AI/NLP Engine to Detect Serious Injury & Fatality (SIF) Precursors in OIL's Unsafe-Act/Unsafe-Condition and Near-Miss Reports

1. Executive Snapshot

Oil India Limited (OIL) captures significant volumes of Unsafe Act / Unsafe Condition (UA/UC) observations, near-misses, and incident reports across onshore and offshore installations via its Health, Safety, Security, and Environment (HSSE) platform [PS]. However, these safety logs are currently triaged manually at periodic intervals (monthly or quarterly) [PS]. This batch-processing latency creates a critical safety vulnerability: high-energy hazards with fatal potential remain unaddressed until after manual review [INFERENCE]. Modern safety science (DEKRA Martin & Black 2015; EEI SIF Precursor Model; VelocityEHS 2024 PSIF Classifier) proves that non-fatal minor accidents and fatal incidents stem from fundamentally different root causes [PS]. While non-fatal US industrial accidents decreased by 51% over 15 years, fatalities fell by only 25.5% [PS]. Leading oil and gas operators separately isolate and prioritize the ~20–25% of safety reports that contain genuine Serious Injury and Fatality (SIF) potential [PS]. This proposal presents "SIF-Guard AI", an intelligent, offline-capable NLP engine and spatial-temporal safety dashboard designed for OIL [ASSUMPTION]. SIF-Guard AI automatically ingests free-text safety reports, classifies SIF-potential vs. non-SIF-potential incidents, maps records to the official 9 IOGP Life-Saving Rules [RESEARCHED], pinpoints barrier failures, and ranks operational assets by precursor density to enable real-time, proactive HSE interventions [PS].

2. Problem Deconstruction

The problem statement provides specific requirements for automated SIF detection [PS]. Below is the atomic breakdown, mapping each explicit requirement to its technical and operational implications:

1. Inbound Safety Report Ingestion [PS]

   • Source Text: "Build a prototype that ingests OIL's free-text safety reports..." [PS]
   • Interpretation: The system must accept unstructured, multi-lingual (English, Hindi, Assamese) free-text data submitted by field personnel via OIL's HSSE portal, mobile apps, or batch CSV/JSON uploads [INFERENCE].
   • Implementation Implication: Ingestion pipeline requiring textual preprocessing, multi-lingual normalization, OCR for scanned forms, and handling of field jargon/abbreviations (e.g., "LOTO", "SWL", "Derrick", "DG Set", "BOP") [INFERENCE].
   • Confidence Tag: [PS]

2. Requirement (a): Binary SIF-Potential Classification [PS]

   • Source Text: "...and automatically a) Classifies each as SIF-potential vs non-SIF-potential..." [PS]
   • Interpretation: A binary machine learning / NLP classification model must evaluate every report description, assessing exposure to high-energy sources and control effectiveness to isolate the ~20–25% of reports carrying SIF potential [PS].
   • Implementation Implication: Supervised Fine-Tuned NLP Transformer (e.g., DeBERTa-v3 / RoBERTa) combined with high-energy hazard rule heuristics (EEI/DEKRA SIF decision logic) [INFERENCE].
   • Confidence Tag: [PS]

3. Requirement (b): Tagging to IOGP Life-Saving Rules [PS]

   • Source Text: "...b) Tags it to the relevant IOGP Life-Saving Rule (e.g., Energy Isolation, Hot Work, Confined Space, Line of Fire)..." [PS]
   • Interpretation: Multi-label classification to map each report to one or more of the 9 standardized IOGP (International Association of Oil & Gas Producers) Life-Saving Rules [RESEARCHED].
   • Implementation Implication: Multi-label NLP classifier trained on standard IOGP definitions (Bypassing Safety Controls, Confined Space, Driving, Energy Isolation, Hot Work, Line of Fire, Safe Mechanical Lifting, Work Authorisation, Working at Height) [RESEARCHED].
   • Confidence Tag: [PS]

4. Requirement (c): Recurring Precursor Pattern & Barrier Failure Extraction [PS]

   • Source Text: "...c) Surfaces recurring precursor patterns (activity, location, barrier failure) via a dashboard." [PS]
   • Interpretation: Information extraction (NER + Intent Mining + Topic Modeling) to extract structured entities: Activity (e.g., pipe racking, welding), Location (e.g., Rig #4, Duliajan Wellhead 12), and Barrier Failure (e.g., uncalibrated gas detector, missing safety latch, bypassed interlock) [INFERENCE].
   • Implementation Implication: Entity extraction pipeline using Named Entity Recognition (NER), dependency parsing, and semantic clustering (HDBSCAN / Sentence-BERT) [INFERENCE].
   • Confidence Tag: [PS]

5. Expected Outcome: Interactive Safety Ranking & Risk Heatmap Dashboard [PS]

   • Source Text: "A working AI/NLP with an interactive dashboard that ranks sites/activities by SIF-precursor density and auto-maps to Life-Saving Rules, enabling HSE to focus interventions where fatal potential is highest." [PS]
   • Interpretation: A Web/GIS dashboard displaying real-time precursor density metrics, geographical heatmaps across OIL fields, drill-down analytics, and automated HSE alert triggers [INFERENCE].
   • Implementation Implication: React.js/Deck.gl/Leaflet interactive dashboard backed by PostgreSQL/PostGIS and FastAPI analytical endpoints [INFERENCE].
   • Confidence Tag: [PS]

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Geographic & Field Conditions: OIL's primary operations are concentrated in Assam (Duliajan, Moran, Digboi), Arunachal Pradesh, Tripura, KG Basin, and offshore assets [RESEARCHED]. Operating environments include dense jungles, heavy rainfall regions, riverine areas, and active drill sites with classified hazardous zones (Zone 0, Zone 1, Zone 2 explosion risks) [KNOWN].
2. Connectivity Constraints: Remote production wells, seismic survey sites, and exploratory drilling rigs frequently suffer from non-existent or patchy 2G/cellular connectivity [KNOWN]. Centralized cloud-only inference will fail in field conditions. Edge caching and offline-first capabilities are strictly required [INFERENCE].
3. Language & Literacy Realities: Field technicians, contractor labor, and roustabouts report incidents in mixed languages—Hinglish, Assamese, Benglish, or non-standard English with local oilfield terminology (e.g., "derrick man", "BOP pressure", "mud pump spill", "LOTO not applied") [INFERENCE].
4. Hardware Constraints: Field personnel use low-to-mid-tier Android smartphones (2GB–4GB RAM) or shared ruggedized industrial tablets [KNOWN]. AI models cannot run raw multi-billion parameter LLMs locally on field devices; lightweight web-based sync and efficient cloud server-side inference (or lightweight edge models) are required [INFERENCE].
5. Regulatory & Statutory Constraints: Oil Industry Safety Directorate (OISD) guidelines (OISD-STD-105) and Petroleum & Natural Gas (PNG) Safety Rules 2008 mandate strict records for "High Potential (Hi-Po) Near Misses" [RESEARCHED]. All data must reside within sovereign Indian cloud infrastructure or OIL on-premise servers [KNOWN].

6. Research Questions

1. What specific text schema and field structures exist in OIL's current HSSE software platform? [NEEDS VALIDATION]
2. What is the baseline distribution of unstructured free-text lengths (word count) in historical UA/UC reports at OIL? [NEEDS VALIDATION]
3. How accurately do fine-tuned small language models (e.g., DeBERTa-v3-base) perform on industrial safety text compared to large commercial LLMs (GPT-4o / Claude 3.5)? [RESEARCHED]
4. What set of high-energy hazard indicators (e.g., electrical > 50V, fall > 1.8m, suspended load, pressurized fluids > 100 PSI) defines SIF potential under the EEI Safety Classification & Learning (SCL) model? [RESEARCHED]
5. What are the specific network security protocols (firewalls, VPNs, API gateways) required to integrate with OIL's corporate SAP/Oracle enterprise systems? [NEEDS VALIDATION]

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

                      +-------------------------------------------------------------+
                      |                 EXISTING SAFETY SYSTEMS                     |
                      +------------------------------+------------------------------+
                                                     |
                   +---------------------------------+---------------------------------+
                   |                                                                   |
                   v                                                                   v
+------------------------------------+                               +------------------------------------+
| Traditional Enterprise EHS Portals |                               | Standard Generic LLM API Wrappers  |
| (SAP EHS, Enablon, Gensuite)       |                               | (Generic ChatGPT / Claude Prompts) |
+------------------------------------+                               +------------------------------------+
| Strengths:                         |                               | Strengths:                         |
| • Robust workflow management       |                               | • Fast natural language parsing    |
| • Scalable database infrastructure |                               | • Zero initial training required   |
| Weaknesses:                        |                               | Weaknesses:                        |
| • Manual reporting categorization  |                               | • Hallucination of safety rules    |
| • Batch review latency             |                               | • Data privacy compliance risks    |
| • Ignores SIF precursor science    |                               | • High latency & API costs         |
| • High false positive clutter      |                               | • Fails offline in remote fields   |
+------------------------------------+                               +------------------------------------+
                   |                                                                   |
                   +---------------------------------+---------------------------------+
                                                     |
                                                     v
                      +-------------------------------------------------------------+
                      |                   PROPOSED SYSTEM GAP                       |
                      | No offline-capable, domain-tuned, SIF-precursor specialized |
                      | NLP platform tailored for Indian O&G operational realities. |
                      +-------------------------------------------------------------+

1. Traditional Enterprise EHS Platforms (SAP EHS, Enablon, Intelex)

   • Description: Comprehensive corporate EHS management tools used for regulatory compliance and incident tracking [KNOWN].
   • Strengths: Reliable database tracking, integrated CAPA (Corrective and Preventive Action) workflows [KNOWN].
   • Weaknesses: Relies heavily on manual dropdown inputs by field workers; lacks NLP-driven automatic SIF classification; processes incident reports in batch intervals [PS].
   • Gap: Fails to isolate the ~20% high-energy SIF precursor reports from thousands of minor observations [PS].
   • Source: Industry EHS platform standards [KNOWN]. Confidence: [KNOWN]

2. Commercial AI PSIF Classifiers (VelocityEHS AI PSIF Insights, EHS Insight SIF Detection)

   • Description: Proprietary SaaS solutions incorporating machine learning models to detect Potential SIF (PSIF) in Western industrial plants [RESEARCHED].
   • Strengths: Validated ML pipelines; multi-factor risk scoring; integration with Western OSHA standards [RESEARCHED].
   • Weaknesses: Expensive SaaS subscription models; cloud-dependent (fails in offline oilfield sites); non-compliant with local Indian sovereign cloud mandates; lacks support for Assamese/Hindi code-switched free-text [INFERENCE].
   • Gap: Cannot operate in remote, low-bandwidth Indian oilfield environments (e.g., Duliajan fields) with localized language nuances [INFERENCE].
   • Source: VelocityEHS Press Release 2025 / EHS Insight [RESEARCHED]. Confidence: [RESEARCHED]

3. Generic LLM-Based Safety Summarizers (Wrapper Applications)

   • Description: Basic OpenAI/Claude API prompt-based wrappers that summarize safety logs [INFERENCE].
   • Strengths: Quick to implement during hackathons [INFERENCE].
   • Weaknesses: High hallucination risk; non-deterministic SIF tagging; privacy violation (sending OIL proprietary safety data to external public APIs); expensive token costs; slow inference speed [INFERENCE].
   • Gap: Unsuitable for safety-critical industrial decision support where deterministic, auditable decisions are required [INFERENCE].
   • Source: Empirical AI research [KNOWN]. Confidence: [INFERENCE]

9. Candidate Solution Approaches

+---------------------------------------------------------------------------------------------------+
|                                candidate solution approaches                                     |
+----------------------------------+----------------------------------+-----------------------------+
| Approach 1: Deterministic Rules  | Approach 2: Pure Cloud LLM       | Approach 3: Hybrid Fine-    |
| & Regex Classifier               | API Pipeline                     | Tuned SLM + Rule Engine     |
+----------------------------------+----------------------------------+-----------------------------+
| • High explainability            | • High natural language understanding| • Domain-tuned accuracy    |
| • Zero ML infrastructure         | • Expensive API tokens           | • Fully offline capable     |
| • Fails on complex free-text     | • High latency & cloud dependency| • Zero API vendor lock-in   |
| • High false negative rate       | • Data privacy compliance risks  | • Real-time deterministic   |
|                                  | • Non-deterministic outputs      |   safety decision fallback  |
+----------------------------------+----------------------------------+-----------------------------+

Approach 1: Deterministic Keyword & Rule-Based Regex Engine

• Mechanism: String matching, dictionary lookups (e.g., "height", "gas release", "crane"), and IF-THEN heuristic decision trees based on EEI rules.
• Problem Fit: Low. Field safety reports use non-standard phrasing, regional jargon, and spelling errors ("scafold unsecure", "gas leak near DG set").
• Accuracy Potential: Poor (<60% F1-score) [INFERENCE]. High false negatives for subtle SIF precursors.
• Data Requirements: Minimal (dictionary lists).
• Complexity: Very Low.
• Explainability: 100% transparent.
• Cost: Very Low ($0).
• Scalability: High speed, but brittle maintainability.
• Hackathon Feasibility: High.
• Production Feasibility: Unacceptable due to poor accuracy.

Approach 2: Pure Cloud LLM API Pipeline (e.g., GPT-4o / Claude 3.5 Prompting)

• Mechanism: Direct REST API calls passing safety reports to a commercial cloud LLM with structured JSON output prompts.
• Problem Fit: Medium. High language understanding, but introduces data privacy and connectivity bottlenecks.
• Accuracy Potential: High (~85–90% zero-shot accuracy) [INFERENCE].
• Data Requirements: Zero training data required; requires prompt engineering.
• Complexity: Low implementation effort, high operational complexity (network dependency).
• Explainability: Low (black-box API).
• Cost: High (recurring per-token API fees across thousands of daily OIL reports) [INFERENCE].
• Scalability: Limited by API rate limits and network bandwidth at remote rig sites.
• Hackathon Feasibility: High.
• Production Feasibility: Unacceptable due to OIL data sovereignty policies, recurring API costs, and lack of offline connectivity in field operations [INFERENCE].

Approach 3 (RECOMMENDED): Hybrid Architecture — Fine-Tuned Transformer (DeBERTa-v3 / RoBERTa) + Deterministic High-Energy Heuristic Rule Engine

• Mechanism: A dual-stage hybrid engine. Stage 1 uses an explicit rule engine implementing the EEI Safety Classification & Learning (SCL) decision tree to flag obvious high-energy precursors (e.g., falls > 1.8m, voltage > 50V, unisolated pressure > 100 PSI). Stage 2 uses a fine-tuned, quantized lightweight NLP Transformer (DeBERTa-v3-base) trained on domain-specific safety text to perform multi-task SIF-potential classification and IOGP LSR multi-label tagging. Stage 3 utilizes HDBSCAN & Sentence-Transformers for clustering recurring precursor patterns.
• Problem Fit: Perfect fit for OIL requirements.
• Accuracy Potential: Excellent (>90% Precision/Recall on SIF potential) [INFERENCE].
• Data Requirements: Requires ~1,000–2,000 labeled safety report training samples (synthesized/augmented for hackathon).
• Complexity: Medium-High.
• Explainability: High (Rule stage is 100% transparent; ML stage outputs confidence scores & attention highlight maps).
• Cost: Low (Open-source model deployed on local GPU/edge server; zero recurring API token fees).
• Scalability: High throughput (>100 reports/second per GPU node) [INFERENCE].
• Hackathon Feasibility: High (Can be trained using Google Colab/Kaggle GPUs and deployed via FastAPI + ONNX Runtime).
• Production Feasibility: Excellent. Fully compatible with OIL on-premise servers and offline edge deployments.

10. Recommended Approach & Why

The Hybrid Architecture (Approach 3) is selected as the winning approach for the SIH solution.

Justification:

1. Safety-Critical Reliability: Pure machine learning models can occasionally hallucinate or miss critical edge cases. Integrating deterministic EEI High-Energy Rule filters ensures that explicit fatal hazards (e.g., "worker inside vessel without gas test") are flagged 100% of the time, regardless of NLP confidence scores [INFERENCE].
2. Data Sovereignty & On-Premise Compliance: OIL is a Navratna Public Sector Undertaking (PSU) under the Ministry of Petroleum & Natural Gas [RESEARCHED]. Transferring operational hazard reports containing exact rig locations and personnel names to external commercial APIs violates national data sovereignty guidelines [INFERENCE]. Approach 3 runs completely on-premise or on sovereign Indian cloud instances.
3. Offline Field Capability: Fine-tuned Transformer models compressed via ONNX Runtime / INT8 quantization can run locally on field laptops or edge gateways situated at remote drilling sites without active internet access [INFERENCE].
4. Superior Cost-Efficiency: Eliminates recurring per-token cloud API bills, allowing OIL to process millions of historical reports cost-effectively.

Rejection of Alternatives:

• Approach 1 (Rules Only) was rejected because regex/keywords fail to capture complex semantic context, syntax errors, and informal field descriptions, leading to dangerously high false-negative rates [INFERENCE].
• Approach 2 (Cloud LLM API) was rejected due to data security policy violations, high recurring operational costs, latency, and complete failure in remote, zero-bandwidth oilfield locations [INFERENCE].

11. Technical Architecture

Layered System Architecture Diagram

+---------------------------------------------------------------------------------------------------+
|                                  SIF-GUARD AI SYSTEM ARCHITECTURE                                 |
+---------------------------------------------------------------------------------------------------+

[ PRESENTATION LAYER - WEB & MOBILE UI ]
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+
  |   Field Officer PWA / Mobile App  |  |  Interactive HSE Command Center   |  |   Executive GIS Precursor Map     |
  |  (React / PWA / Offline IndexedDB)|  |   (React.js / Tailwind / Recharts)|  |     (Deck.gl / Leaflet / PostGIS) |
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+
                                                        |
                                                        v (REST / WebSockets / HTTPS)
[ API GATEWAY & SECURITY LAYER ]
  +-------------------------------------------------------------------------------------------------+
  |  FastAPI Gateway | OAuth2 / JWT Auth | Rate Limiter | Data Scrubbing & PII Anonymizer           |
  +-------------------------------------------------------------------------------------------------+
                                                        |
                                                        v
[ INGESTION & DATA PROCESSING LAYER ]
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+
  |   Real-Time Field Stream Ingestion|  |    Batch CSV / SAP EHS Ingest     |  | Multi-Lingual Text Preprocessing  |
  |      (RabbitMQ / Kafka / Celery)  |  |      (Pandas / Polars Pipeline)   |  | (Spacy /IndicNLP / FastText)      |
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+
                                                        |
                                                        v
[ HYBRID AI / NLP HYBRID INFERENCE ENGINE ]
  +-------------------------------------------------------------------------------------------------+
  |  STAGE 1: Deterministic EEI High-Energy Hazard Rule Filter (IF-THEN Precursor Logic)            |
  +-------------------------------------------------------------------------------------------------+
                                                        |
  +-----------------------------------------------------+-------------------------------------------+
  |                                                                                                 |
  v                                                                                                 v
  |  STAGE 2A: SIF-Potential Classifier                 |  STAGE 2B: IOGP Life-Saving Rule Classifier |
  |  (Fine-Tuned DeBERTa-v3 ONNX Runtime)               |  (Multi-Label Transformer Model)          |
  +-----------------------------------------------------+-------------------------------------------+
                                                        |
  +-----------------------------------------------------+
  |  STAGE 3: Entity & Pattern Extraction Engine (Spacy NER + HDBSCAN Semantic Clustering)          |
  +-------------------------------------------------------------------------------------------------+
                                                        |
                                                        v
[ DATA ARCHITECTURE & STORAGE LAYER ]
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+
  | PostgreSQL + PostGIS (Metadata)   |  | Vector DB (Qdrant / pgvector)     |  | Redis Cache & Task Queue           |
  | (Reports, Assets, Audit Logs)     |  | (Semantic Search & Clustering)    |  | (Real-time Alert Distribution)    |
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+
                                                        |
                                                        v
[ INTEGRATION & NOTIFICATION LAYER ]
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+
  |  SMS / WhatsApp / Email Gateway   |  |   SAP EHS / Oracle Sync Adapter   |  |  OISD Statutory Audit Export      |
  |   (Twilio / Local SMS Gateway)    |  |       (REST / Enterprise Bus)     |  |   (PDF / Excel Report Generator)  |
  +-----------------------------------+  +-----------------------------------+  +-----------------------------------+

Detailed Step-by-Step Data Flow Narrative:

1. Ingestion: A field worker submits an Unsafe Act/Unsafe Condition (UA/UC) report narrative (e.g., "Worker seen operating welding torch on gas pipeline without hot work permit and gas test near Derrick Rig #3") via the offline-first Progressive Web App (PWA) or batch CSV upload [INFERENCE].
2. Preprocessing & Anonymization: The FastAPI backend receives the report. The PII Anonymizer strips personnel names, phone numbers, and employee IDs using Spacy NER. Text is standardized, removing typos and translating regional Assamese/Hindi terms to standardized English [INFERENCE].
3. Stage 1 — Rule-Based Filtering: The text is passed through the EEI High-Energy Heuristic Rule Filter. It checks for hardcoded critical keywords and energy thresholds (e.g., "hot work", "gas pipeline", "no permit"). If an explicit life-threatening rule violation is triggered, it is instantly tagged as SIF-Potential = High (Rule-Triggered) [INFERENCE].
4. Stage 2 — Deep Transformer Inference: Concurrently, the preprocessed text is tokenized and processed by the fine-tuned DeBERTa-v3 ONNX model.
   • Head A: Predicts probability of SIF Potential ($P_{SIF} \in [0,1]$). If $P_{SIF} \ge 0.5$, classified as SIF-Potential [INFERENCE].
   • Head B: Multi-label classification predicts applicable IOGP Life-Saving Rules (e.g., Hot Work, Work Authorisation) [INFERENCE].
5. Stage 3 — Information & Barrier Extraction: The NER engine extracts structured entities:
   • Activity: Welding / Hot Work
   • Location: Derrick Rig #3
   • Barrier Failure: Missing Work Permit & Omission of Gas Test [INFERENCE].
6. Storage & Analytics Ingestion: Results are committed to PostgreSQL (relational metadata + PostGIS location tags) and vector embeddings are stored in pgvector/Qdrant for similarity clustering [INFERENCE].
7. Precursor Density Calculation: The analytics engine recalculates the SIF Precursor Density Metric for "Derrick Rig #3" and the "Welding/Hot Work" activity category: $$\text{SIF Precursor Density} = \frac{\sum \text{SIF-Potential Reports in Last 30 Days}}{\text{Total Reports in Asset Cluster}} \times 100$$
8. Real-time Alerting & Visualization: If the asset's precursor density exceeds the threshold (>25%), an automated alert is pushed to the Installation HSE Manager via SMS/WhatsApp, and the location glows red on the Executive GIS Dashboard [INFERENCE].

12. Technology Stack

13. AI/ML Justification

1. SIF-Potential Binary Classifier

• Purpose: Classify whether an unstructured safety report description carries genuine fatal potential (~20-25%) or is a low-severity event [PS].
• Model Type: Fine-Tuned Transformer (microsoft/deberta-v3-base) [RESEARCHED].
• Input: Unstructured text string (e.g., "Scaffolding plank slipped from 15ft height while worker was passing tools without safety harness").
• Output: SIF-Potential Flag (True/False) + Probability Score (e.g., 0.92).
• Data Required: ~2,000 safety report descriptions annotated with ground-truth SIF labels based on EEI SCL criteria [ASSUMPTION].
• Why AI is Preferable: Simple rules fail to comprehend linguistic variations, implied context, and complex interactions between hazards and missing controls [INFERENCE].
• Evaluation Metric: Precision, Recall, F1-score, ROC-AUC (Target: Recall $\ge 90%$ to minimize false negatives on fatal risks) [INFERENCE].
• Failure Mode & Mitigation: If model yields a false negative on an ambiguous report, the Stage 1 Rule Filter acts as a safety net [INFERENCE].
• Human Oversight: All reports flagged as SIF-Potential require mandatory sign-off/review by the Installation HSE Manager [INFERENCE].

2. Multi-Label IOGP Life-Saving Rules Classifier

• Purpose: Automatically map safety reports to one or more of the 9 official IOGP Life-Saving Rules [PS].
• Model Type: Multi-Label Classification Transformer Head attached to DeBERTa-v3 [RESEARCHED].
• Input: Unstructured text narrative.
• Output: Binary vector representing presence of 9 IOGP LSRs (e.g., [Energy Isolation: 0, Hot Work: 1, Confined Space: 0, Line of Fire: 1, ... ]).
• Data Required: Multi-label dataset categorized by certified HSE domain experts [ASSUMPTION].
• Why AI is Preferable: Safety reports frequently involve multiple overlapping rules (e.g., Hot Work in a Confined Space). Standard keyword search cannot accurately distinguish primary context [INFERENCE].
• Evaluation Metric: Micro/Macro F1-score, Hamming Loss [INFERENCE].
• Failure Mode & Mitigation: Low confidence predictions (<0.4) trigger a prompt for manual tagging by the HSE reviewer [INFERENCE].
• Human Oversight: HSE Officers can manually modify auto-tagged LSR labels on the dashboard UI [INFERENCE].

3. Precursor Entity & Pattern Extraction (NER & Clustering)

• Purpose: Extract specific Entities: Activity, Location, Barrier Failure, and cluster recurring patterns [PS].
• Model Type: Custom Spacy Named Entity Recognition (NER) + Sentence-BERT (all-MiniLM-L6-v2) + HDBSCAN Density-Based Clustering [RESEARCHED].
• Input: Text narrative.
• Output: Extracted Entity JSON + Cluster ID (e.g., Cluster #12: Recurring Gas Detector Calibration Failures in Moran Rig #2).
• Data Required: Entity-annotated safety text + pre-computed semantic embeddings.
• Why AI is Preferable: Automatically discovers emerging systemic failures without predefined categories [INFERENCE].
• Evaluation Metric: Silhouette Coefficient, Cluster Density Score [INFERENCE].
• Failure Mode & Mitigation: Outlier reports assigned to Noise Cluster (-1) for manual inspection [INFERENCE].
• Human Oversight: HSE Directors review auto-generated cluster summaries [INFERENCE].

Explicit Non-AI Component Identification:

• Precursor Density Calculation: Pure mathematical calculation ($\frac{\text{SIF Reports}}{\text{Total Reports}}$) executed directly in PostgreSQL SQL queries. AI MUST NOT be used for statistical aggregation [INFERENCE].
• User Permission & Access Control: Role-Based Access Control (RBAC) managed strictly via standard JWT authentication logic. AI MUST NOT participate in security decisions [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

Database Schema (PostgreSQL DDL Concept)

-- Core Table: Safety Reports
CREATE TABLE safety_reports (
    report_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    report_code VARCHAR(50) UNIQUE NOT NULL,
    field_asset_id UUID REFERENCES field_assets(asset_id),
    report_timestamp TIMESTAMPTZ NOT NULL,
    raw_narrative TEXT NOT NULL,
    cleaned_narrative TEXT,
    reporter_role VARCHAR(50),
    is_sif_potential BOOLEAN NOT NULL,
    sif_confidence_score FLOAT NOT NULL,
    classification_source VARCHAR(20) CHECK (classification_source IN ('RULE_ENGINE', 'ML_MODEL', 'HYBRID')),
    extracted_activity VARCHAR(100),
    extracted_location VARCHAR(100),
    barrier_failure_type VARCHAR(150),
    status VARCHAR(20) DEFAULT 'PENDING_REVIEW',
    created_at TIMESTAMPTZ DEFAULT NOW()
);

-- Junction Table: Multi-Label IOGP LSR Tags
CREATE TABLE report_iogp_tags (
    report_id UUID REFERENCES safety_reports(report_id) ON DELETE CASCADE,
    lsr_id INT CHECK (lsr_id BETWEEN 1 AND 9),
    confidence_score FLOAT NOT NULL,
    PRIMARY KEY (report_id, lsr_id)
);

-- Spatial Table: Field Assets (Rigs, Stations, Wells)
CREATE TABLE field_assets (
    asset_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    asset_name VARCHAR(100) NOT NULL,
    asset_type VARCHAR(50) CHECK (asset_type IN ('DRILLING_RIG', 'PRODUCTION_STATION', 'PIPELINE', 'REFINERY')),
    region VARCHAR(50) NOT NULL, -- e.g., 'Duliajan', 'Moran', 'KG Basin'
    location GEOMETRY(Point, 4326) NOT NULL,
    is_active BOOLEAN DEFAULT TRUE
);

Offline & Low-Network Design Strategy

+---------------------------------------------------------------------------------------------------+
|                               OFFLINE-FIRST SYNCHRONIZATION FLOW                                  |
+---------------------------------------------------------------------------------------------------+

 [ FIELD USER DEVICE ]                                           [ OIL CENTRAL SERVER ]
  +-----------------------+                                       +-----------------------+
  | Submit UA/UC Report   |                                       |                       |
  +-----------+-----------+                                       |                       |
              |                                                   |                       |
              v                                                   |                       |
  { Network Available? } --(YES)--> [ Direct FastAPI REST POST ] -------> [ Ingest & Process ]
              |                                                   |                       |
            (NO)                                                  |                       |
              v                                                   |                       |
  [ Save to IndexedDB ]                                           |                       |
  [ Queue for Sync    ]                                           |                       |
  [ Display Offline UI]                                           |                       |
              |                                                   |                       |
  (Background Listener Checks Network)                            |                       |
              |                                                   |                       |
   { Connectivity Restored }                                       |                       |
              |                                                   |                       |
              +-------------------> [ ServiceWorker Sync Task ] -> [ Batch Sync Endpoint  |
                                    (Exponential Backoff Retry)   |  with Conflict Log ]  |
                                                                  +-----------------------+

1. Local Storage Mechanism: The PWA UI uses Dexie.js (IndexedDB wrapper) to store unsubmitted reports locally on field smartphones/tablets [INFERENCE].
2. Bandwidth Reduction: Inbound text reports are light (~1–2 KB per record). Attachments/photos are compressed client-side (WebP format, max 800x600 resolution) before queueing [INFERENCE].
3. Synchronization & Conflict Resolution:
   • Uses a background ServiceWorker (BackgroundSync API) that detects network restoration [KNOWN].
   • Reports are sent in compressed batch JSON payloads (gzip).
   • Conflict Handling: Every offline report generates a client-side UUID and timestamp. Server accepts all unique client UUIDs. Duplicate timestamps from the same user are deduplicated automatically [INFERENCE].
4. Retry Strategy: Exponential backoff with jitter (initial retry 5s, doubling up to max 5 minutes) to avoid overloading gateway upon network reconnection [INFERENCE].

15. Security, Privacy & Governance

1. PII Anonymization & Data Protection: Safety reports frequently contain employee names, contractor agency details, and supervisor identities. An automatic Named Entity Recognition (NER) scrubber replaces individuals' names with tokens ([PERSONNEL_ANONYMIZED]) prior to model ingestion and vector embedding storage [INFERENCE].
2. Authentication & Role-Based Access Control (RBAC):
   • OAuth2 with JSON Web Tokens (JWT) and SHA-256 signatures [KNOWN].
   • Roles: Field_Reporter (Create reports), Rig_HSE_Manager (Review/verify local SIF flags), Corporate_HSE_Director (View system-wide analytics/heatmaps), System_Admin (Manage user roles) [INFERENCE].
3. Data Encryption Standard:
   • Data at Rest: AES-256 bit encryption for PostgreSQL tables and disk volumes [KNOWN].
   • Data in Transit: TLS 1.3 encryption across all REST/WebSocket API endpoints [KNOWN].
4. Data Ownership & Regulatory Compliance:
   • All safety data, trained model weights, and analytics remain the 100% exclusive property of Oil India Limited [PS].
   • System complies with Oil Industry Safety Directorate (OISD) Incident Reporting Guidelines (OISD-STD-105) and Indian Digital Personal Data Protection (DPDP) Act 2023 [RESEARCHED].
5. AI Failure & Human-in-the-Loop Governance:
   • The system operates explicitly as a Decision Support System (DSS), NOT an autonomous safety enforcement tool [INFERENCE].
   • AI-generated SIF flags and IOGP LSR tags serve as recommendations. No site shutdown or disciplinary action can occur without human HSE verification [INFERENCE].
   • Full Audit Logging: Every modification to an AI prediction by a human manager is logged with user ID, timestamp, and justification reason for auditing [INFERENCE].

16. MVP Scope (Hackathon Build)

+---------------------------------------------------------------------------------------------------+
|                                      HACKATHON MVP BUILD SCOPE                                    |
+---------------------------------------------------------------------------------------------------+

                         MUST BUILD (Core Value Proposition)
                         ------------------------------------
                         • Free-text report ingestion (Single/Batch CSV)
                         • Dual SIF Classifier (Fine-Tuned DeBERTa-v3 + EEI Rules)
                         • Multi-label IOGP Life-Saving Rules Tagging Engine (9 Rules)
                         • Interactive GIS Risk Heatmap (Deck.gl / PostGIS)
                         • Precursor Density Ranking (Site / Activity)
                         • Human-in-the-Loop Feedback & Audit UI

                         SHOULD BUILD (High Impact Differentiators)
                         ------------------------------------------
                         • Offline-first PWA submission form (IndexedDB sync)
                         • Entity Extraction Engine (Activity, Location, Barrier)
                         • Automated PDF/Excel OISD Report Generator

                         NICE TO HAVE (Time Permitting)
                         ------------------------------
                         • Assamese/Hindi Voice-to-Text reporting
                         • Real-time WhatsApp/SMS alert simulation

                         DO NOT BUILD DURING HACKATHON
                         -----------------------------
                         • Production SAP/Oracle ERP Deep Connectors
                         • Biometric Login / Enterprise Active Directory Sync
                         • Autonomous Drone/CCTV Video Analytics Integration
+---------------------------------------------------------------------------------------------------+

In Scope (Hackathon Build)

• Core NLP Engine: Fine-tuned Transformer model + EEI heuristic rule filter for binary SIF detection [PS].
• IOGP LSR Mapper: Automated tagging across all 9 IOGP Life-Saving Rules [PS].
• Entity Extraction: Basic NER model extracting Activity, Location, and Barrier Failure [PS].
• Interactive Dashboard: Web UI displaying spatial asset heatmaps, precursor density rankings, and drill-down audit views [PS].
• Offline Demo Mode: Functional PWA demonstrating offline report logging and re-sync upon toggling offline/online mode [INFERENCE].
• Dataset Engine: Realistic synthetic dataset of 1,500+ Indian O&G safety reports generated and annotated for training and live demo testing [ASSUMPTION].

Described Only / Out of Scope (Production System)

• Live direct database connector to OIL's production SAP EHS / Oracle enterprise deployment [INFERENCE].
• Full integration with OIL's internal Single Sign-On (SSO) / Active Directory infrastructure [INFERENCE].
• Hardware-level integration with intrinsically safe explosion-proof handheld mobile terminals [INFERENCE].

17. Live Demo Simulation Plan

To present a convincing demonstration to SIH judges without exposing proprietary OIL data or requiring live oilfield connectivity, the team will implement the following simulation protocol:

+---------------------------------------------------------------------------------------------------+
|                                     LIVE DEMO SIMULATION FLOW                                     |
+---------------------------------------------------------------------------------------------------+

 [ STEP 1: REAL-TIME NARRATIVE INPUT ]
 Inject complex free-text report (e.g., "Rig #4 Duliajan: Gas monitor alarm beeped during grinding.
 Operator bypassed detector and continued work without permit.")

                                   |
                                   v
 [ STEP 2: LIVE HYBRID INFERENCE VISUALIZATION ]
 Display side-by-side processing execution:
   • Rule Engine: Triggers "Bypassing Safety Control" & "Hot Work" Rules.
   • DeBERTa-v3 Model: Outputs SIF Probability = 96.4% (SIF-Potential).
   • Multi-Label Head: Highlights IOGP LSR Icons (#1 Bypassing Controls, #5 Hot Work).

                                   |
                                   v
 [ STEP 3: DYNAMIC MAP & DENSITY RE-RANKING ]
 Dashboard map updates instantly: Duliajan Rig #4 glows RED.
 Precursor Density Metric jumps from 18% to 28% (Exceeds Threshold).

                                   |
                                   v
 [ STEP 4: OFFLINE PWA SYNC SIMULATION ]
 Disconnect laptop Wi-Fi -> Submit new report on mobile PWA UI -> Reconnect Wi-Fi.
 PWA automatically pushes queued payload; dashboard updates without page refresh via WebSockets.
+---------------------------------------------------------------------------------------------------+

1. Simulated Ingestion Feeds: A real-time "Stream Simulator" tool built into the UI that injects synthetic field reports at a controlled rate (1 report every 3 seconds) to demonstrate live background streaming, queueing, and processing [INFERENCE].
2. Realistic Ground-Truth Test Cases: Prepare 5 distinct test scenarios representing diverse oilfield operational contexts:
   • Case A (High-Energy SIF Precursor): Gas detector bypass during hot work near gas separator [INFERENCE].
   • Case B (Low-Severity Non-SIF): Trip hazard over tripping hazard tape on walkway (Minor UA/UC) [INFERENCE].
   • Case C (Complex Multi-LSR Event): Unisolated valve opening during confined space vessel entry [INFERENCE].
   • Case D (Jargon/Regional Phrasing): Hinglish description ("Derrick man missed safety latch on pipe elevator") [INFERENCE].
   • Case E (False Positive Stress Test): Worker dropped a plastic coffee cup from 2-meter platform (Height present, but low mass/energy) -> Model correctly classifies as Non-SIF [INFERENCE].
3. Data Transparency Statement: Explicitly state to judges: "This live demo utilizes a realistic synthetic dataset designed according to OISD/IOGP standards to preserve operational confidentiality." [INFERENCE]

18. Implementation Roadmap

Target Deadline: 30 September 2026 [PS].

+---------------------------------------------------------------------------------------------------+
|                                      IMPLEMENTATION ROADMAP                                       |
+---------------------------------------------------------------------------------------------------+

 [ PHASE 1: DATA PIPELINE & RULE ENGINE ] (Duration: 2 Weeks | Target: Sept 10, 2026)
   • Finalize taxonomy for IOGP 9 Life-Saving Rules and EEI high-energy criteria.
   • Synthesize and curate 2,000 annotated O&G safety reports (SIF vs Non-SIF, LSR tags).
   • Build Stage 1 EEI High-Energy Rule Engine in Python.

                                   |
                                   v
 [ PHASE 2: MODEL TRAINING & NLP OPTIMIZATION ] (Duration: 2 Weeks | Target: Sept 18, 2026)
   • Fine-tune DeBERTa-v3-base multi-task model on synthetic dataset.
   • Train Spacy NER model for Activity, Location, Barrier Failure extraction.
   • Quantize models to INT8 ONNX format for accelerated inference.

                                   |
                                   v
 [ PHASE 3: BACKEND API & DATABASE ARCHITECTURE ] (Duration: 1.5 Weeks | Target: Sept 23, 2026)
   • Set up PostgreSQL + PostGIS database schema and `pgvector` indexing.
   • Build FastAPI REST endpoints, Celery async background processing, and WebSocket stream.
   • Implement IndexedDB offline sync logic in ServiceWorker.

                                   |
                                   v
 [ PHASE 4: FRONTEND DASHBOARD & GIS VISUALIZATION ] (Duration: 1 Week | Target: Sept 28, 2026)
   • Develop React.js UI with Deck.gl spatial mapping, risk heatmaps, and ranking tables.
   • Integrate Human-in-the-Loop review panel and export capabilities.

                                   |
                                   v
 [ PHASE 5: SYSTEM TESTING, SIMULATION & FREEZE ] (Duration: 2 Days | Target: Sept 30, 2026)
   • Perform end-to-end stress testing, offline sync verification, and false-positive evaluation.
   • Code Freeze & Final Presentation rehearsed for SIH judging panel.
+---------------------------------------------------------------------------------------------------+

19. Testing & Validation Plan

20. Risk Register

21. Cost & Scalability

Hackathon Prototype Cost

• Compute / GPUs: $0 (Utilizing free Google Colab T4 / Kaggle GPUs for model training) [ASSUMPTION].
• Database & Hosting: $0 (Local Docker deployment on developer laptops) [ASSUMPTION].
• Mapping API: $0 (Open-source Leaflet / Deck.gl + OpenStreetMap tiles) [ASSUMPTION].
• Total Hackathon Cost: ₹0 / $0 [ASSUMPTION].

Production Deployment Cost (Oil India Limited On-Premise / Sovereign Cloud)

• On-Premise Compute Hardware: 2x GPU Server Nodes (NVIDIA L4 24GB or A10G) for real-time inference & batch retraining ~ ₹8,000,000 (One-time capital expenditure) [INFERENCE].
• Database & Application Server Infrastructure: Existing OIL Private Cloud VMs (PostgreSQL + Redis + FastAPI Docker nodes) ~ ₹0 incremental hardware cost [INFERENCE].
• Software Licensing: 100% Open-Source Stack (Linux, PostgreSQL, DeBERTa, FastAPI, React) ~ ₹0 recurring vendor software license fees [INFERENCE].
• Scalability Capability:
  • Single GPU node handles >150 reports/second [INFERENCE].
  • Horizontal API scaling via Docker Swarm / Kubernetes to support thousands of concurrent field workers across all OIL operational assets [INFERENCE].

22. Differentiation

23. Future Scope

1. Multimodal Incident Reporting: Extend model to accept field photographs and audio voice notes, utilizing lightweight Vision Transformers (ViT) to detect missing PPE / safety barrier breaches in images [INFERENCE].
2. Predictive Root Cause & CAPA Recommendation: Integrate Retrieval-Augmented Generation (RAG) over OIL's historical Corrective and Preventive Action (CAPA) database to automatically suggest past successful mitigation strategies for newly detected SIF precursors [INFERENCE].
3. Wearable IoT & Geofencing Integration: Connect real-time SIF precursor alerts to field workers' intrinsically safe smartwatches/beacons to trigger proximity warnings when entering high-density SIF hazard zones (e.g., under active suspended loads) [INFERENCE].

24. Assumptions & Open Items Log

Consolidated [ASSUMPTION] Items:

1. Synthetic Dataset Baseline: Assumed a synthetic dataset of 2,000 O&G safety reports generated from OSHA/IOGP public domain logs is sufficient for prototype fine-tuning [ASSUMPTION].
2. Spatial Asset Coordinates: Assumed anonymized geospatial coordinates representing Assam oilfield clusters for GIS heatmap demonstration [ASSUMPTION].
3. Zero Software Budget: Assumed open-source software stack and free tier development tools are used for hackathon build [ASSUMPTION].
4. User Persona & Hardware: Assumed field workers possess basic Android smartphones capable of running modern WebPWAs [ASSUMPTION].

Consolidated [NEEDS VALIDATION] Items:

1. OIL HSSE Database Schema: Exact schema and API interfaces of OIL's existing HSSE reporting software platform [NEEDS VALIDATION].
2. Historical Report Length: Mean word count and exact language distribution (English vs Assamese vs Hindi) in OIL historical safety logs [NEEDS VALIDATION].
3. Corporate Firewall Protocols: Network integration and VPN clearance requirements for deploying containerized services inside OIL's intranet [NEEDS VALIDATION].

Final Recommendation

Build Approach 3 (Hybrid Fine-Tuned SLM + EEI Rule Engine). It provides high precision on SIF detection, strictly adheres to IOGP standards, operates offline in remote oilfields, ensures 100% data privacy for OIL, and directly answers all explicit requirements of SIH26165 [PS].

Recommended Stack

• Frontend: React.js + Vite + TailwindCSS + Deck.gl + Dexie.js (PWA IndexedDB) [KNOWN].
• Backend: FastAPI (Python 3.12) + Celery + Redis + Spacy NER [KNOWN].
• AI/ML: Fine-Tuned DeBERTa-v3-base + ONNX Runtime + EEI Rule Engine [RESEARCHED].
• Database: PostgreSQL 16 + PostGIS + pgvector [KNOWN].

Biggest Dependency

Acquisition or high-fidelity synthesis of realistic Indian oil & gas free-text safety reports annotated with ground-truth SIF and IOGP Life-Saving Rule labels [INFERENCE].

Biggest Risk

Alarm fatigue caused by high false-positive rates, leading HSE officers to ignore dashboard warnings [INFERENCE].

Biggest Technical Challenge

Fine-tuning a lightweight, quantized NLP Transformer model that achieves $>90%$ SIF recall while correctly handling mixed English/Hinglish field jargon and spelling anomalies [INFERENCE].

First 5 Actions

1. Action 1: Construct and validate the standard 9 IOGP Life-Saving Rules JSON taxonomy and EEI High-Energy decision tree rules [RESEARCHED].
2. Action 2: Generate and annotate a high-quality synthetic dataset of 2,000 O&G UA/UC reports [ASSUMPTION].
3. Action 3: Build and validate the Stage 1 EEI High-Energy Python Rule Engine [INFERENCE].
4. Action 4: Fine-tune microsoft/deberta-v3-base on the annotated dataset for SIF binary classification and multi-label LSR tagging [RESEARCHED].
5. Action 5: Develop the React.js + Deck.gl interactive spatial dashboard and integrate the FastAPI ONNX inference engine [INFERENCE].

Validation Before Coding

• Validate that the synthetic training dataset accurately reflects the 20–25% SIF distribution ratio cited in research literature (DEKRA / VelocityEHS) [PS].
• Verify that ONNX Runtime inference latency on standard x86 CPU is under 150 ms per report [INFERENCE].

Buildability Score /100

92 / 100 Reasoning: Excellent alignment with open-source stack and standard NLP classification techniques. The primary effort lies in dataset annotation and UI dashboard development, which are completely achievable within the hackathon timeline by a 4-6 student team [INFERENCE].

Confidence

HIGH [INFERENCE]

Research Sources

1. Title: Safety Classification and Learning (SCL) Model

   • URL: https://www.eei.org/-/media/Project/EEI/Documents/Issues-and-Policy/Power-to-Prevent-SIF/eeiSCLmodel.pdf
   • Source Type: Official Industry Research Report (Edison Electric Institute) [RESEARCHED]
   • Supported Claim: Defines scientific SIF precursor conditions, high-energy decision logic, and rules for distinguishing SIF-potential events (Martin & Black 2015 criteria) [PS].
   • Reliability: High (Peer-reviewed industry standard) [RESEARCHED]

2. Title: IOGP Life-Saving Rules (IOGP Report 459)

   • URL: https://www.iogp.org/workstreams/safety/safety/life-savingrules/
   • Source Type: International Industry Standard Organization [RESEARCHED]
   • Supported Claim: Standardized definitions of the 9 IOGP Life-Saving Rules (Bypassing Controls, Confined Space, Driving, Energy Isolation, Hot Work, Line of Fire, Safe Lifting, Work Authorisation, Height) [RESEARCHED].
   • Reliability: High (Global oil & gas industry consensus standard) [RESEARCHED]

3. Title: Automatic Identification of Incidents Involving Potential Serious Injuries and Fatalities (PSIF)

   • URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC10998882/
   • Source Type: Academic Journal Publication / PubMed Central [RESEARCHED]
   • Supported Claim: Proves machine learning (Transformers + XGBoost) combined with NLP effectively automates SIF/PSIF precursor classification from unstructured incident logs [RESEARCHED].
   • Reliability: High (Peer-reviewed scientific journal) [RESEARCHED]

4. Title: VelocityEHS Launches New "PSIF" AI to Identify Next Serious Injury or Fatality

   • URL: https://www.globenewswire.com/news-release/2025/07/01/3108237/0/en/velocityehs-launches-new-psif-ai-to-identify-the-next-serious-injury-or-fatality-before-it-happens.html
   • Source Type: Official Enterprise Press Release / Product Announcement [RESEARCHED]
   • Supported Claim: Demonstrates commercial viability of applying machine learning to automatically flag PSIFs in safety reports without altering user workflows [PS].
   • Reliability: High (Official corporate release) [RESEARCHED]

5. Title: Incident Reporting Guidelines — Oil Industry Safety Directorate (OISD)

   • URL: https://oisd.gov.in/incident-reporting-guidelines
   • Source Type: Official Indian Government Regulatory Directorate [RESEARCHED]
   • Supported Claim: Outlines statutory definitions for High Potential (Hi-Po) Near Misses, Form 2 / Form 3 quarterly reporting, and OISD safety compliance rules for Indian petroleum operators [RESEARCHED].
   • Reliability: High (Official Government of India Regulatory Body) [RESEARCHED]