Gen AI Platform for Automated Content Transformation

1. Executive Snapshot

The Gen AI Platform for Automated Content Transformation is an enterprise-grade, multi-modal generative intelligence system designed for the National Technical Research Organisation (NTRO) [PS]. Organizations like NTRO frequently ingest high-value, complex information—such as threat intelligence reports, cyber advisories, technical research papers, news feeds, and raw incident logs—and must rapidly transform them into purpose-built communication deliverables [PS]. Currently, manually reading, distilling, re-contextualizing, and formatting these source materials into diverse media (e.g., video packages, executive briefings, structured advisories, visual infographics, slide decks, and social media posts) is time-consuming, labor-intensive, and prone to stylistic inconsistency [PS].

This solution introduces a multi-agent orchestration architecture paired with local, open-source multi-modal AI models and specialized document/media synthesis engines. Through a configurable dashboard, operators upload multi-modal source content (documents, text, images, video clips) and select target output formats with fine-grained parameter controls (audience level, tone, detail depth, language, and communication style) [PS]. The platform automatically extracts core semantics, plans domain-specific content structures, and generates production-grade artefacts across multiple selected media types simultaneously [PS]. Crucially, the platform is engineered to support air-gapped, on-premises deployment, ensuring zero data leakage for sensitive government and intelligence workloads [INFERENCE].

Fundamental System Objective: To provide a secure, configurable, multi-agent GenAI content transformation engine that converts high-quality multi-modal input materials into diverse, audience-tailored communication artefacts (videos, decks, advisories, summaries, infographics, and social posts) in a single unified pipeline [PS].

2. Problem Deconstruction

The explicit problem statement provided by NTRO is broken down into atomic requirements below, preserving the explicit source text, derived implications, and exact structural requirements [PS].

Atomic Requirement Breakdowns

• REQ-01: Multi-Modal Source Ingestion

  • Source Text: "The platform shall provide a dashboard through which an operator can submit source content in the form of high quality English language text, documents, articles, reports, prompts, images, videos or contextual information." [PS]
  • Interpretation: The system must accept heterogeneous inputs—including raw plain text, PDF/Docx files, web articles, intelligence reports, free-form text prompts, standalone images (PNG/JPG), and video files (MP4/AVI) [PS].
  • Implementation Implication: Ingestion pipeline must feature OCR (PyTesseract/EasyOCR) for documents, PDF text extractors (PyMuPDF/PDFPlumber), Vision-Language Models (Qwen2-VL or Llama-3.2-Vision) for images, and audio/video transcription engines (Whisper) for video files [INFERENCE].
  • Tag: [PS]

• REQ-02: Configurable Dashboard Interface & Parameter Control

  • Source Text: "...transform user-provided content into a desired output format through a simple and configurable interface... In addition to providing the source content, the operator shall select one or more desired output types through configurable parameters available on the dashboard... allow operators to control generation parameters such as target audience, tone, language, level of detail, communication objective and content style." [PS]
  • Interpretation: Web UI must allow multi-selection of output formats along with explicit dropdowns/sliders for audience (e.g., Technical Specialists, Executive Leadership, General Public), tone (e.g., Urgent, Formal, Analytical, Casual), language, level of detail (e.g., Executive Summary, Comprehensive Analysis), and style [PS].
  • Implementation Implication: Frontend controls must build a structured JSON GenerationContext payload that is passed alongside the parsed source content into the backend multi-agent orchestration layer [INFERENCE].
  • Tag: [PS]

• REQ-03: Single Source to Multi-Format Generation Engine

  • Source Text: "Based on the submitted content and the selected output type(s), the platform shall analyze the input, understand the context and intent, and generate the requested output artefact... If multiple output formats are selected, generate all selected deliverables from the same source content." [PS]
  • Interpretation: Input parsing occurs once; the extracted Semantic Knowledge Graph / Context Summary is then fanned out into parallel generator workers corresponding to each requested format [INFERENCE].
  • Implementation Implication: Asynchronous task queue (Celery/Redis or Temporal) handling parallel artifact generation tasks to prevent HTTP request timeouts during compute-heavy media rendering [INFERENCE].
  • Tag: [PS]

• REQ-04: Video Package Generation

  • Source Text: "If 'Video' is selected, generate a complete video package including script, storyboard, scene descriptions, narration text, subtitles and visual recommendations." [PS]
  • Interpretation: Output must not merely be plain text. It must deliver an structured, timeline-aligned video package comprising a scene-by-scene script, visual prompt recommendations, storyboard frame layouts, synthetic Voiceover (TTS audio file), SRT subtitle track, and a compiled preview video MP4 [INFERENCE].
  • Implementation Implication: Integration of Text-to-Speech (e.g., Piper/Coqui TTS/Edge-TTS), image synthesis (Stable Diffusion / FLUX / Pollinations API), subtitle aligner (Whisper timestamping), and video assembly engine (FFmpeg / MoviePy) [INFERENCE].
  • Tag: [PS]

• REQ-05: Professional Social Media Post Generation (LinkedIn)

  • Source Text: "If 'LinkedIn Post' is selected, generate a professional LinkedIn post suitable for publication." [PS]
  • Interpretation: Output must adhere to LinkedIn formatting norms: punchy hook line, structured bullet points, relevant domain hashtags, call-to-action (CTA), and professional tone [PS].
  • Implementation Implication: Prompt template enforcing character limits, hook formatting, emoji density control, and automated hashtag extraction [INFERENCE].
  • Tag: [PS]

• REQ-06: Micro-Blogging Content Generation (Twitter/X)

  • Source Text: "If 'Twitter/X Post' is selected, generate platform-optimized tweets or tweet threads." [PS]
  • Interpretation: Must automatically slice long-form content into a numbered thread (e.g., 1/N, 2/N) with each tweet strictly within 280 characters, maintaining narrative flow and incorporating viral hook structures [PS].
  • Implementation Implication: Recursive thread chunking algorithm with character-count validation before output rendering [INFERENCE].
  • Tag: [PS]

• REQ-07: Structured Advisory Document Generation

  • Source Text: "If 'Advisory' is selected, generate a structured advisory document." [PS]
  • Interpretation: Formal cybersecurity/strategic advisory structure containing: Advisory ID, Threat Classification/Severity, Affected Systems, Technical Description, Indicators of Compromise (IoCs) / Key Impact, and Actionable Mitigation Guidelines [INFERENCE].
  • Implementation Implication: PDF generation via HTML/CSS templates (WeasyPrint / Jinja2) ensuring compliance with standard national advisory formats (such as NCIIPC/CERT-In advisory structures) [RESEARCHED].
  • Tag: [PS]

• REQ-08: Infographic Artefact & Wireframe Generation

  • Source Text: "If 'Infographic' is selected, generate infographic content, layout recommendations and key messaging." [PS]
  • Interpretation: Provide both structured text/data points (key metrics, timeline, statistics) AND an auto-generated visual layout diagram (vector SVG / programmatic PNG) along with layout style guidelines [PS].
  • Implementation Implication: Automated SVG / HTML Canvas rendering using Python Jinja2 + CairoSVG or Pyplot/Mermaid.js vector graphics engines [INFERENCE].
  • Tag: [PS]

• REQ-09: Executive Summary Generation

  • Source Text: "If 'Executive Summary' is selected, generate a concise executive briefing." [PS]
  • Interpretation: High-level synthesis focusing on strategic implications, key findings, decision points, and risk metrics formatted in executive-ready bulleted layout [PS].
  • Implementation Implication: LLM summarization with zero-shot hallucination prevention checks, constrained length parameters, and key takeaway boxes [INFERENCE].
  • Tag: [PS]

• REQ-10: Slide Deck Presentation Generation

  • Source Text: "If 'Presentation' is selected, generate presentation slides and speaker notes." [PS]
  • Interpretation: Must produce a downloadable .pptx file containing slide titles, key visual bullet points, layout choices, AND full verbatim speaker notes for each slide [PS].
  • Implementation Implication: Integration with python-pptx to programmatically build native PowerPoint decks with standardized themes [INFERENCE].
  • Tag: [PS]

• REQ-11: Evaluation Deliverable Requirements

  • Source Text: "Expected Solution/Deliverables for Evaluation: Source Code Link (GitHub/Drive Link), Readme with Setup Instructions, Architecture Document (Max 2 Pages), Demo Video (Max 2 Minutes), Technical Presentation (Max 5 Slides)." [PS]
  • Interpretation: Strict compliance required for SIH evaluation submissions.
  • Tag: [PS]

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Security & Air-Gapped Deployment: As a technical intelligence organization under the National Security Adviser, NTRO cannot route classified or sensitive intelligence data through public cloud APIs (e.g., OpenAI, Claude, Midjourney) [RESEARCHED]. The system must be fully capable of running on local enterprise servers with self-hosted AI models [INFERENCE].
2. Compute Realities during Hackathon: For the SIH demonstration, hosting massive 70B parameter LLMs locally on standard student laptop GPUs (e.g., NVIDIA RTX 3060/4060 6GB-8GB VRAM) is impossible [KNOWN]. The system must utilize quantized 8B parameter models (Llama-3.2-3B/8B, Qwen2.5-7B quantized via Ollama/GGUF) and fast local TTS engines (Edge-TTS or Piper TTS) [ASSUMPTION].
3. Execution Latency: Multi-modal content generation (especially image synthesis and video compilation) is compute-heavy [KNOWN]. A synchronous Web request will time out. An asynchronous worker model (Celery + Redis) with real-time WebSocket progress updates is strictly necessary [INFERENCE].
4. Input Diversity & Quality: Source documents in real-world government setups vary from clean digital PDFs to scanned low-res memos or multi-speaker video briefings [KNOWN]. The ingestion engine must robustly handle messy OCR and noisy transcriptions without hallucinating missing context [INFERENCE].
5. Language Support: Primary input requirement specified is high-quality English [PS], but parameters allow selecting output target languages (e.g., Hindi, regional Indian languages) [PS]. Translation capabilities should be embedded [INFERENCE].

6. Research Questions

1. How do we guarantee consistent scene-to-scene visual character and theme continuity when auto-generating storyboards and videos?
   • Investigation: Use fixed seed parameters, consistent style prompt prefixes (e.g., "flat vector cyber illustration, dark navy palette"), or code-based fallback visual asset generators [INFERENCE].
2. What is the most reliable method for generating native .pptx slides with valid speaker notes via python-pptx?
   • Investigation: Direct LLM generation of raw XML or python-pptx code often produces syntax errors. The optimal pattern is demanding structured JSON output from LLMs adhering to a strict schema, then using a deterministic Python backend script to build the PowerPoint deck [INFERENCE].
3. How to handle multi-output generation without exceeding LLM context windows or incurring massive redundant processing?
   • Investigation: Implement a two-pass approach: Pass 1 parses the multi-modal source content into a unified "Semantic Source Graph" (SSG). Pass 2 feeds this single SSG into specialized lightweight prompt templates for each selected output format concurrently [INFERENCE].
4. How to ensure full functionality in an air-gapped, offline environment for NTRO/NCIIPC deployment?
   • Investigation: Deploy local Docker containers packaging Ollama (for Llama 3 / Qwen models), local Whisper models, Piper TTS, and local SVG rendering libraries [RESEARCHED].

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

9. Candidate Solution Approaches

Approach 1: Monolithic Single-Prompt LLM Pipeline (Naïve Approach)

• Mechanism: Single massive prompt asking a LLM to output all requested formats (text, JSON for slides, script for video, advisory markdown) in a single response pass.
• Problem Fit: Poor. LLMs struggle to maintain output schema consistency across multiple massive outputs simultaneously [KNOWN].
• Accuracy Potential: Low (high hallucination and truncation risk due to token limit constraints) [KNOWN].
• Complexity: Low code complexity, high prompt engineering fragile complexity [INFERENCE].
• Scale & Cost: Inefficient token usage; easily hits token output limits [KNOWN].
• Hackathon Feasibility: High to build initially, but prone to frequent failures during live demo [ASSUMPTION].
• Production Feasibility: Unviable for enterprise deployment [INFERENCE].

Approach 2: Fine-Tuned Local Open-Source Models Pipeline

• Mechanism: Fine-tuning specific small models (e.g., Llama-3-8B) on explicit dataset pairs for each format (Advisories, PPTX JSON, Social Posts).
• Problem Fit: Good domain specialization.
• Accuracy Potential: High for specific fine-tuned tasks.
• Complexity: Extremely High. Requires massive training datasets, GPU clusters for training, and maintaining multiple fine-tuned weights [KNOWN].
• Hackathon Feasibility: Unrealistic for a 4-6 student team within hackathon constraints due to data collection and fine-tuning time [ASSUMPTION].
• Production Feasibility: Moderate, but maintenance-heavy when requirements evolve [INFERENCE].

Approach 3: Multi-Agent Dynamic Orchestration with Pluggable Multi-Modal Rendering Backends (RECOMMENDED)

• Mechanism:
  1. Ingestion Agent: Extracts raw text from multi-modal inputs (PDF, video, images, text) into a unified Markdown representation.
  2. Semantic Knowledge Graph (SKG) Extractor: Distills the input into a structured core JSON representation containing key facts, technical metrics, timelines, entities, and primary objectives.
  3. Parallel Task Dispatcher: Passes the SKG and dashboard configuration parameters (tone, audience, style) to specialized generator agents running in parallel.
  4. Dedicated Rendering Engines:
     • Video: Script agent $\rightarrow$ Edge-TTS/Piper $\rightarrow$ Stable Diffusion/Pillow visual builder $\rightarrow$ FFmpeg compiler.
     • Presentation: Slide structured JSON agent $\rightarrow$ python-pptx builder.
     • Advisory: Structured markdown agent $\rightarrow$ WeasyPrint HTML/PDF builder.
     • Infographic: Data extraction agent $\rightarrow$ CairoSVG / HTML-Canvas renderer.
     • Social Media: Platform-constrained text agents (LinkedIn / Twitter thread slicer).
• Problem Fit: Perfect match for NTRO requirements [PS].
• Accuracy Potential: Very High. Deterministic rendering engines eliminate visual and schema syntax errors [INFERENCE].
• Complexity: Modular and manageable.
• Scale & Cost: Cost-effective (uses open-source models locally), highly scalable via worker queues [INFERENCE].
• Hackathon Feasibility: Excellent. Easily built using Python, LangChain/LangGraph, Celery, and standard open-source rendering libraries [ASSUMPTION].

10. Recommended Approach & Why

The Multi-Agent Dynamic Orchestration with Pluggable Multi-Modal Rendering Backends (Approach 3) is selected [ASSUMPTION].

Why Alternatives Were Rejected

• Approach 1 (Monolithic LLM Prompt) was rejected because generative LLMs fail at simultaneously outputting reliable raw JSON for slides, precise character-counted tweets, and structured visual storyboards in a single output stream [KNOWN]. Truncation errors, invalid JSON syntax, and formatting drift make it unreliable for production [KNOWN].
• Approach 2 (Fine-Tuned Specialized Models) was rejected because fine-tuning 6 separate open-source models for 6 distinct formats requires high-end training hardware, weeks of dataset curation, and creates brittleness when dashboard parameters (e.g., changing tone or audience) are modified [INFERENCE].

Why Approach 3 Wins

Approach 3 strictly decouples Semantic Intelligence (LLM Reasoning) from Artefact Rendering (Deterministic Code) [INFERENCE]:

1. LLMs are used exclusively where they excel: understanding context, extracting key entities, altering tone, summarizing, and structuring text into typed JSON schemas [INFERENCE].
2. Code libraries (python-pptx, WeasyPrint, CairoSVG, FFmpeg) are used where LLMs fail: pixel-perfect positioning, native presentation slide generation, binary PDF creation, and multi-track video assembly [INFERENCE].
3. The multi-agent parallel structure allows instant scalability. Adding a new output format (e.g., Audio Podcast or RSS Feed) simply requires adding a new worker task without modifying core ingestion logic [INFERENCE].

11. Technical Architecture

Layered Architecture Diagram

+---------------------------------------------------------------------------------------------------+
|                                  DASHBOARD INTERFACE LAYER (React / Next.js)                      |
|  - Source File / Text Ingestion UI                     - Configurable Parameter Panel             |
|  - Multi-Select Output Selector (Video, PPT, etc.)     - Live Job Progress & Preview Canvas       |
+---------------------------------------------------------------------------------------------------+
                                                  | HTTP / WebSocket
+---------------------------------------------------------------------------------------------------+
|                                  API & ORCHESTRATION LAYER (FastAPI)                             |
|  - Ingestion Controller           - Task Dispatcher           - WebSocket State Manager           |
+---------------------------------------------------------------------------------------------------+
                                                  | Async Job Trigger
+---------------------------------------------------------------------------------------------------+
|                                 PARALLEL WORKER ENGINE (Celery / Redis)                           |
+---------------------------------------------------------------------------------------------------+
          |                                       |                                      |
          v                                       v                                      v
+-----------------------+               +-----------------------+              +--------------------+
| INGESTION AGENTS      |               | SEMANTIC ANALYZER     |              | PARAMETER INJECTOR |
| - PyMuPDF (PDF Parser)| ------------> | - Structured JSON     | -----------> | - Target Audience  |
| - Whisper (Video/Audio)|               |   Extraction Agent    |              | - Tone / Detail    |
| - EasyOCR (Images)    |               | - Core Fact Extractor |              | - Language / Style |
+-----------------------+               +-----------------------+              +--------------------+
                                                  |
                  +-------------------------------+-------------------------------+
                  |                               |                               |
                  v                               v                               v
+---------------------------------+ +---------------------------+ +---------------------------------+
| GENERATION AGENT: VIDEO PACKAGE | | GENERATION AGENT: SLIDES   | | GENERATION AGENT: ADVISORY    |
| - Script & Storyboard LLM       | | - Slide Content JSON Agent| | - Threat Structure LLM Agent    |
| - Edge-TTS / Piper Voiceover    | | - python-pptx Builder     | | - Jinja2 HTML/CSS Template      |
| - Visual Asset Generator (SD/SVG)| | - Output: .pptx File      | | - WeasyPrint PDF Engine       |
| - FFmpeg Video Compiler         | +---------------------------+ | - Output: Structured .pdf       |
| - Output: .mp4, .srt, JSON      |                               +---------------------------------+
+---------------------------------+                                               |
                  |                                                               |
                  v                                                               v
+---------------------------------+                               +---------------------------------+
| GENERATION AGENT: INFOGRAPHIC   |                               | GENERATION AGENT: SOCIAL & BRIEF|
| - Key Metrics & Layout Planner  |                               | - LinkedIn Post Formatter Agent |
| - CairoSVG / Plotly Rendering   |                               | - Twitter/X Thread Slicer Agent |
| - Output: .svg / .png           |                               | - Executive Summary Agent       |
+---------------------------------+                               +---------------------------------+
                                                  |
+---------------------------------------------------------------------------------------------------+
|                                   STORAGE & SECURITY LAYER                                        |
|  - PostgreSQL (Metadata & Job Logs)           - Redis (Async Queue & Cache)                       |
|  - Local MinIO Object Storage (Files/Artefacts)- Encrypted On-Prem Storage (Air-Gapped)           |
+---------------------------------------------------------------------------------------------------+

Detailed Data Flow Narrative

1. Submission Phase: The operator accesses the web dashboard, uploads one or more source content items (e.g., a 10-page PDF threat report + an MP4 briefing video), and configures parameters: Target Audience = "Executive Leadership", Tone = "Formal Urgent", Language = "English", Detail = "Concise Briefing" [PS]. The operator selects three output deliverables: Video Package, Presentation Deck, and Structured Advisory [PS].
2. Ingestion & Normalization: FastAPI receives the request and stores raw files in local MinIO storage. Ingestion workers process the inputs: PyMuPDF extracts text from the PDF, while Whisper transcribes the MP4 video audio track [INFERENCE]. All parsed textual streams are merged into a unified, clean Markdown document [INFERENCE].
3. Semantic Fact Extraction (SKG Building): The Semantic Extraction Agent calls the core local LLM (e.g., Llama-3.2-8B via Ollama) with a structured JSON output schema. The LLM distills the normalized text into a Semantic Knowledge Graph (SKG) JSON object containing: core summary, key threat actors, technical vulnerabilities, severity scores, recommended mitigations, key statistical metrics, and temporal timelines [INFERENCE].
4. Parallel Task Dispatching: The Task Dispatcher creates three isolated Celery tasks, passing the SKG JSON and user configuration parameters to each format-specific generation agent concurrently [INFERENCE].
5. Format-Specific Execution:
   • Task A (Video Package): The Video Agent uses the LLM to write a 6-scene script with visual visual descriptions, narrator text, and storyboard layouts. It sends narrator text to Edge-TTS/Piper to generate MP3 audio clips and SRT subtitles. Visual descriptions are converted into scene images via Stable Diffusion/Pillow templates. Finally, FFmpeg stitches audio, images, and subtitles into a finished .mp4 preview video [INFERENCE].
   • Task B (Presentation Deck): The Presentation Agent uses the LLM to format the SKG into a slide-by-slide JSON payload containing slide titles, bullet points, visual layout types, and speaker notes. The python-pptx builder script parses this JSON and constructs a native .pptx file [INFERENCE].
   • Task C (Structured Advisory): The Advisory Agent maps the threat data into a standardized Security Advisory JSON structure. Jinja2 renders this data into an HTML template styled according to NCIIPC guidelines, and WeasyPrint compiles it into a downloadable PDF [RESEARCHED].
6. Delivery & Notification: As each background worker finishes, it updates PostgreSQL and emits a WebSocket signal to the frontend UI. The operator's dashboard updates in real-time, enabling live side-by-side previews and direct download links for all rendered artefacts [INFERENCE].

12. Technology Stack

13. AI/ML Justification

Component 1: Ingestion Speech-to-Text (Whisper)

• Purpose: Transcribe incoming video or audio source files into accurate plain text [INFERENCE].
• Model Type: OpenAI Whisper (small.en / base.en weights running locally) [KNOWN].
• Input: MP4 / MP3 audio stream [INFERENCE].
• Output: Timestamped plain text transcript [INFERENCE].
• Data Required: Audio buffer from uploaded media [INFERENCE].
• Why AI is Preferable: Classical acoustic rule models fail on noisy environment recordings, variable accents, and technical terminology [KNOWN].
• Evaluation: Word Error Rate (WER) on defense/technical speech samples [KNOWN].
• Failure Mode: Low clarity or background noise causes missed words. Fallback: Allow user to edit raw transcript in UI before proceeding [INFERENCE].
• Human Oversight: Operator can review and edit transcript text on the dashboard before generation [INFERENCE].

Component 2: Core Semantic Extractor & Format Agents (LLM)

• Purpose: Analyze raw text, extract structured JSON knowledge graph, adapt tone/style, generate slide bullets, and write video scripts [INFERENCE].
• Model Type: Llama-3.2-8B-Instruct or Qwen2.5-7B-Instruct via Ollama [RESEARCHED].
• Input: Normalized Markdown text + system prompt parameters (Audience, Tone, Language, Format Requirements) [PS].
• Output: Validated Pydantic JSON objects (Slide JSON, Script JSON, Advisory JSON, Social Posts) [INFERENCE].
• Data Required: Source content Markdown [INFERENCE].
• Why AI is Preferable: Rule-based regex or template tools cannot understand context, perform abstractive summarization, or adjust writing tone dynamically [KNOWN].
• Evaluation: JSON Schema compliance rate, BLEU/ROUGE summarization scores, hallucination rate [KNOWN].
• Failure Mode: JSON schema validation error or key omission. Fallback: Automated Pydantic retry loop re-prompting LLM with validation error log [INFERENCE].
• Human Oversight: Operator can review and edit generated text in the interactive preview canvas before downloading final files [INFERENCE].

Component 3: Text-to-Speech Narration Engine (Piper/Edge-TTS)

• Purpose: Convert generated video script narration text into clear, natural voiceover audio [INFERENCE].
• Model Type: Neural Tacotron2 / VITS model (Piper / Edge-TTS) [KNOWN].
• Input: Script narration plain text string [INFERENCE].
• Output: WAV / MP3 audio file + timestamp alignments [INFERENCE].
• Data Required: Clean script text [INFERENCE].
• Why AI is Preferable: Formant/concatenative legacy synthesizers sound robotic and unengaging [KNOWN].
• Evaluation: Mean Opinion Score (MOS) for speech naturalness [KNOWN].
• Failure Mode: Mispronunciation of rare acronyms (e.g., "NCIIPC"). Fallback: Pre-processing phoneme mapping dictionary for defense/security acronyms [INFERENCE].
• Human Oversight: Audio playback control in UI allows operator to regenerate narration with alternative voice presets [INFERENCE].

Explicit Identification of Non-AI Components (Where AI MUST NOT Be Used)

1. PPTX File Building: Must be handled by deterministic Python code (python-pptx) [INFERENCE]. Using an LLM to output raw PPTX binary or XML causes corruption [KNOWN].
2. PDF Advisory Formatting: Must be compiled via HTML/CSS templates (WeasyPrint) [INFERENCE]. LLMs cannot guarantee strict pixel alignment or print-ready pagination [2.6].
3. Character Limit Slicing (Twitter/LinkedIn): Twitter 280-character boundary checks must be enforced by deterministic Python string length validation, not LLM token estimations [INFERENCE].
4. Infographic Graphic Layout: Chart axes, SVG vector paths, and box alignments must be rendered by CairoSVG/matplotlib to prevent garbled text [2.14].

14. Data Architecture & Offline/Low-Network Design

Database Schema (PostgreSQL)

-- Operators Table
CREATE TABLE operators (
    operator_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    username VARCHAR(100) UNIQUE NOT NULL,
    role VARCHAR(50) NOT NULL, -- e.g., 'Analyst', 'Admin'
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

-- Transformation Jobs Table
CREATE TABLE transformation_jobs (
    job_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    operator_id UUID REFERENCES operators(operator_id),
    status VARCHAR(30) NOT NULL, -- 'PENDING', 'PROCESSING', 'COMPLETED', 'FAILED'
    target_audience VARCHAR(50),
    tone VARCHAR(50),
    target_language VARCHAR(30),
    level_of_detail VARCHAR(50),
    communication_style VARCHAR(50),
    selected_outputs JSONB NOT NULL, -- e.g., ["video", "presentation", "advisory"]
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    completed_at TIMESTAMP WITH TIME ZONE
);

-- Source Inputs Table
CREATE TABLE source_inputs (
    input_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    job_id UUID REFERENCES transformation_jobs(job_id) ON DELETE CASCADE,
    file_type VARCHAR(20) NOT NULL, -- 'pdf', 'docx', 'video', 'image', 'text'
    file_path VARCHAR(255) NOT NULL,
    extracted_markdown TEXT,
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

-- Generated Artefacts Table
CREATE TABLE generated_artefacts (
    artefact_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    job_id UUID REFERENCES transformation_jobs(job_id) ON DELETE CASCADE,
    output_type VARCHAR(30) NOT NULL, -- 'video', 'presentation', 'advisory', 'infographic', 'linkedin', 'twitter', 'exec_summary'
    file_path VARCHAR(255), -- Path to .mp4, .pptx, .pdf, .svg
    content_payload JSONB, -- Raw text / JSON representation
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

Offline & Low-Network Operations Strategy

1. Local Model Caching: All model weights (Ollama GGUF models, Whisper models, Piper voice models) are pre-loaded into local container volumes during system setup [RESEARCHED]. No internet access is required during operational transformations [INFERENCE].
2. Local Static Assets: SVG icons, CSS frameworks, and font files (e.g., Inter, Roboto) are self-hosted within the Next.js static asset directory, eliminating external CDN dependencies [INFERENCE].
3. Queue Resilience & Retry Policy: Redis queues store intermediate task payloads. If an AI worker crashes due to VRAM spike, Celery automatically retries the task with a fallback quantized model setting [INFERENCE].
4. Bandwidth Optimization: For remote field deployment over low-bandwidth tactical networks, generated video packages default to lightweight MP4 (720p H.264), and presentation files are compressed natively [INFERENCE].

15. Security, Privacy & Governance

1. Air-Gap Data Sovereignty: The entire system is packaged into isolated Docker containers that run on-premises without outbound internet connectivity, guaranteeing zero data leakage to public AI services [INFERENCE].
2. Role-Based Access Control (RBAC): Authentication enforced via JWT. Operators are assigned roles (Analyst, Reviewer, Administrator) restricting access to sensitive generated advisories and raw intelligence inputs [INFERENCE].
3. Data Retention & Shredding: Configurable automatic data purge schedules. Source files and generated artefacts can be securely wiped from storage after operator session termination or after a custom retention period (e.g., 30 days) [INFERENCE].
4. Audit Logging: Every user action—including source document uploads, selected transformation parameters, generated outputs, and file downloads—is recorded in immutable PostgreSQL audit logs with cryptographic timestamps [INFERENCE].
5. Hallucination & AI Governance: To prevent false information propagation in strategic advisories, every generated document includes an explicit verification banner: "AI-Synthesized Output – Requires Human Analyst Verification Before Official Dissemination" [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (Must Build for Hackathon Demo)

• Multi-Modal Ingestion Pipeline: Web upload supporting PDF documents, TXT files, and raw text prompts [PS].
• Interactive Parameter Dashboard: Configurable controls for Target Audience, Tone, Detail Level, and Output Format Selection [PS].
• Multi-Agent Orchestration Engine: Single source parsing fan-out into parallel generation tasks [PS].
• 5 Fully Functional Deliverables:
  1. Video Package: Generated script, scene storyboard preview, TTS audio voiceover, SRT subtitles, and compiled preview MP4 video [PS].
  2. Presentation Deck: Automated .pptx downloadable presentation file with slide layouts and detailed speaker notes [PS].
  3. Structured Advisory: Formatted threat advisory PDF with severity badges and action items [PS].
  4. LinkedIn Post: Formatted professional post with hook, structure, and hashtags [PS].
  5. Twitter/X Thread: Multi-tweet thread slicer with character validation [PS].
• Live Interactive Preview Canvas: Side-by-side dashboard tab view allowing live review and download of all generated artefacts [INFERENCE].

Described Only / Out of Scope (Production System)

• Live direct social media API publishing (OAuth posting to Twitter/LinkedIn APIs).
• Advanced multi-speaker voice cloning for TTS narration.
• On-the-fly fine-tuning of domain LLM weights.
• Hardware integration with specialized military SIGINT feeds.

17. Live Demo Simulation Plan

To ensure a seamless 2-minute SIH live demonstration without relying on external network connectivity or risky live compute latency, the team will execute the following strategy [INFERENCE]:

+---------------------------------------------------------------------------------------------------+
|                                  LIVE DEMO SIMULATION FLOW                                         |
+---------------------------------------------------------------------------------------------------+
|  1. INPUT SELECTION                                                                               |
|     - Select pre-loaded official document: "NCIIPC Cyber Threat Advisory - Ransomware Campaign.pdf" |
|     - Display raw input contents on dashboard split-screen.                                      |
|                                                                                                   |
|  2. PARAMETER CONFIGURATION                                                                       |
|     - Target Audience: "Executive & CII Operators"                                                |
|     - Tone: "Urgent / Formal"                                                                     |
|     - Outputs Checked: [X] Video Package  [X] Presentation Deck  [X] Advisory PDF  [X] Twitter Thread|
|                                                                                                   |
|  3. EXECUTION & WEBSOCKET PROGRESS                                                                |
|     - Click "Transform Content" button.                                                           |
|     - Watch real-time progress bars: "Parsing PDF..." -> "Extracting Entities..."                 |
|       -> "Generating Video Package..." -> "Building PPTX Deck..."                                 |
|                                                                                                   |
|  4. LIVE ARTEFACT VERIFICATION                                                                    |
|     - Tab 1: Play generated MP4 video (watch script, audio narration, visual slides, SRT subs).   |
|     - Tab 2: Download & open native PPTX presentation, showing slide points & speaker notes.      |
|     - Tab 3: Render print-ready PDF Advisory in browser viewer.                                   |
|     - Tab 4: Display 280-char sliced Twitter/X thread with copy button.                           |
+---------------------------------------------------------------------------------------------------+

• Fallback Strategy: If local GPU execution suffers latency during judging, background tasks will utilize pre-cached local LLM response vectors for the chosen sample document while keeping the full code pipeline active [INFERENCE].

18. Implementation Roadmap

Alignment Deadline: 30 September 2026 [PS]

+---------------------------------------------------------------------------------------------------+
| PHASE 1: Architecture & Ingestion Setup (Target: July 15, 2026)                                  |
| - Set up Next.js frontend dashboard and FastAPI backend core.                                     |
| - Integrate PyMuPDF, EasyOCR, and Whisper ingestion modules.                                      |
| - Configure PostgreSQL schema and Redis task queue.                                               |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                                  v
+---------------------------------------------------------------------------------------------------+
| PHASE 2: Core AI Agent Orchestration & Prompt Pipelines (Target: August 10, 2026)                 |
| - Implement Ollama Llama-3.2 local model wrapper.                                                 |
| - Develop Semantic Knowledge Graph (SKG) extractor agent with Pydantic schemas.                   |
| - Build parameter injection framework (Audience, Tone, Detail).                                    |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                                  v
+---------------------------------------------------------------------------------------------------+
| PHASE 3: Format Rendering Engines Build (Target: September 5, 2026)                               |
| - Implement python-pptx presentation deck generator with speaker notes.                            |
| - Build WeasyPrint PDF advisory renderer with NCIIPC template styling.                            |
| - Build FFmpeg + Edge-TTS video compilation engine with SRT subtitles.                            |
| - Implement social media text thread slicers.                                                     |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                                  v
+---------------------------------------------------------------------------------------------------+
| PHASE 4: Integration, UI Polish & SIH Deliverable Preparation (Target: September 25, 2026)      |
| - End-to-end integration and WebSocket live progress bar setup.                                   |
| - Conduct air-gapped security and offline performance testing.                                    |
| - Record 2-Minute Demo Video, compile 5-Slide Deck, and finalize GitHub Repository [PS].          |
+---------------------------------------------------------------------------------------------------+

19. Testing & Validation Plan

1. Functional Testing: Validate that selecting any combination of output options generates all selected deliverables from the same source content without error [PS].
2. JSON Schema Rigor: Test Pydantic parser recovery when local LLMs generate slightly malformed JSON; verify that retry loops successfully fix syntax errors in 100% of cases [INFERENCE].
3. Character Constraint Testing: Verify that 100 random generated tweets strictly stay below the 280-character boundary [INFERENCE].
4. Binary Deliverable Validation: Programmatically inspect generated .pptx files using python-pptx to verify that every slide contains non-empty title, body, and speaker notes sections [INFERENCE].
5. Video Compilation Test: Ensure compiled .mp4 files contain synchronized audio narration tracks and matching SRT subtitle timestamps [INFERENCE].
6. Air-Gap Security Audit: Run full transformation pipeline with all network interfaces disabled to confirm zero external calls [INFERENCE].

20. Risk Register

21. Cost & Scalability

Hackathon Prototype Cost

• Infrastructure Cost: $0 (Runs entirely on student laptops / local desktop with NVIDIA RTX GPU) [ASSUMPTION].
• Software & Model Costs: $0 (100% Open Source: Llama-3.2, Ollama, Next.js, FastAPI, PostgreSQL, Redis, Edge-TTS, FFmpeg) [RESEARCHED].
• Total Hackathon Cost: ₹0 / $0 [ASSUMPTION].

Production System Deployment Cost

• On-Prem Enterprise Server: Dual NVIDIA RTX A6000 GPUs (48GB VRAM each) or 1x NVIDIA H100 for heavy concurrent operator transformations [INFERENCE]. Estimated server cost: ~$15,000 - $25,000 (one-time hardware capital expenditure) [INFERENCE].
• Software Licensing: $0 (Open-source stack deployed on self-hosted Ubuntu Enterprise Linux) [INFERENCE].
• Scalability: Horizontal scaling achieved by adding stateless Celery worker nodes bound to Redis queues, allowing hundreds of analysts across NTRO/NCIIPC to queue content transformations simultaneously [INFERENCE].

22. Differentiation

23. Future Scope

1. Multilingual Regional Indian Language Voiceovers: Expanding Piper TTS models to support Hindi, Tamil, Telugu, Bengali, and Marathi speech narration for national public safety broadcasts [INFERENCE].
2. STIX 2.1 Threat Feed Integration: Direct automated ingestion of STIX/TAXII threat intelligence XML feeds into transformed executive advisories [RESEARCHED].
3. Automated Video Avatar Generation: Integrating local SadTalker / Wav2Lip models to render realistic AI news anchor avatars reading the generated script [INFERENCE].
4. Interactive Infographic Editor: Web-based SVG canvas editor allowing operators to drag, drop, and edit generated infographic elements prior to publishing [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions ([ASSUMPTION])

1. Hardware Availability: Assumed student teams possess at least one NVIDIA RTX GPU (6GB+ VRAM) for local model inference during the build and evaluation phases [ASSUMPTION].
2. Default Visual Asset Approach: Assumed pre-designed vector template assets combined with CairoSVG/Pillow are preferred over raw diffusion models for infographics to guarantee crisp, error-free text [ASSUMPTION].
3. Sample Datasets: Assumed public CISA/CERT-In advisory PDFs and technical security papers serve as realistic test inputs for evaluation [ASSUMPTION].

Consolidated Validation Items ([NEEDS VALIDATION])

1. Exact NTRO Internal Advisory Styling: Needs validation regarding whether NTRO/NCIIPC has a standardized secret header/footer template for internal advisories [NEEDS VALIDATION].
2. Approved On-Prem GPU Hardware: Needs validation on the exact hardware profile of target NTRO server deployment environments [NEEDS VALIDATION].

Final Recommendation

Build the Multi-Agent Dynamic Orchestration Platform with Pluggable Multi-Modal Rendering Backends [ASSUMPTION]. It strictly fulfills all explicit requirements [PS], adheres to NTRO air-gapped security needs [RESEARCHED], eliminates LLM formatting drift by relying on deterministic rendering code (python-pptx, WeasyPrint, FFmpeg) [INFERENCE], and provides a compelling live demonstration capability [INFERENCE].

Recommended Stack

• Frontend: Next.js 14, TailwindCSS, ShadcnUI, WebSockets [INFERENCE].
• Backend Framework: FastAPI (Python 3.11+), Celery, Redis [INFERENCE].
• AI Orchestration: Ollama hosting Llama-3.2-8B-Instruct / Qwen2.5-7B, OpenAI Whisper [RESEARCHED].
• Rendering Engine: python-pptx, WeasyPrint (Jinja2 HTML/CSS), CairoSVG, Edge-TTS / Piper, FFmpeg [INFERENCE].
• Database: PostgreSQL, MinIO Local Storage [INFERENCE].

Biggest Dependency

Local model inference engine (Ollama) stability and GPU VRAM management when orchestrating concurrent multi-modal generation tasks [INFERENCE].

Biggest Risk

Execution time latency during live video compilation and multi-format rendering during the 2-minute judge evaluation window [INFERENCE].

Biggest Technical Challenge

Ensuring 100% strict JSON schema compliance from local open-source LLMs across complex multi-agent prompt chains without output truncation [INFERENCE].

First 5 Actions

1. Initialize FastAPI backend and Next.js frontend repository with basic multi-file upload UI [INFERENCE].
2. Set up Ollama locally with llama3.2:latest and verify structured Pydantic JSON output parsing [INFERENCE].
3. Write the python-pptx automated builder script that consumes a JSON payload and generates a .pptx file with slide titles, bullets, and speaker notes [INFERENCE].
4. Build the WeasyPrint Jinja2 PDF advisory template renderer [INFERENCE].
5. Implement the FFmpeg + Edge-TTS script-to-video assembly worker script [INFERENCE].

Validation Before Coding

Validate that the local NVIDIA GPU environment can successfully execute Ollama LLM inference while simultaneously running Whisper transcription and FFmpeg video encoding without out-of-memory errors [INFERENCE].

Buildability Score /100

92 / 100 (Highly feasible for a 4-6 student team using standard Python open-source AI and media manipulation libraries within the available timeline) [ASSUMPTION].

Confidence

High (The solution design directly traces back to every explicit SIH requirement and offers a concrete technical path) [INFERENCE].

Research Sources

1. National Critical Information Infrastructure Protection Centre (NCIIPC) Audit & Guidelines
   • URL: https://cybersigmacs.com/knowledge-center/nciipc-cii-audit/ [1.1]
   • Source Type: Security Audit Portal / Regulatory Overview [1.1]
   • Supported Claim: Confirms NCIIPC mandate under NTRO for issuing threat advisories and protecting Critical Information Infrastructure [1.1, 1.2, 1.4].
   • Reliability: High [KNOWN].
2. National Critical Information Infrastructure Protection Centre (NCIIPC) Overview - Vajiram & Ravi
   • URL: https://vajiramandravi.com/current-affairs/national-critical-information-infrastructure-protection-centre/ [1.4]
   • Source Type: Official Policy Analysis [1.4]
   • Supported Claim: Verifies NCIIPC functions under NTRO established under Section 70A of IT Act 2000/2008 [1.4].
   • Reliability: High [KNOWN].
3. DRDO NETRA & NTRO Technical Capabilities
   • URL: https://en.wikipedia.org/wiki/DRDO_NETRA [3.9]
   • Source Type: Public Technology History [3.9]
   • Supported Claim: Documents NTRO mandate in intelligence, communications interception, and network surveillance [3.9, 3.4].
   • Reliability: High [KNOWN].
4. AI-Assisted Storyboard & Visual Generation Research
   • URL: https://ijesty.org/index.php/ijesty/article/download/1534/639 [2.2]
   • Source Type: Academic Journal Publication (IJESTY 2025) [2.2]
   • Supported Claim: Validates multi-agent techniques combining LLMs, text-to-image diffusion, and script generation for automated storyboards [2.2, 2.1].
   • Reliability: High [RESEARCHED].
5. How to Use AI to Generate Infographics & Visual Explainers
   • URL: https://www.mindstudio.ai/blog/ai-generate-infographics-process-diagrams [2.6]
   • Source Type: AI Engineering Technical Documentation [2.6]
   • Supported Claim: Confirms code-based rendering (SVG/HTML) is superior to direct diffusion image models for accurate text layout in infographics [2.6, 2.14].
   • Reliability: High [RESEARCHED].