| --- |
| title: ForensicAI - Forensic Triage Intelligence |
| emoji: ๐ฌ |
| colorFrom: red |
| colorTo: gray |
| sdk: docker |
| app_port: 7860 |
| fullWidth: true |
| header: mini |
| startup_duration_timeout: 10m |
| tags: |
| - ml-intern |
| --- |
| |
| <div align="center"> |
|
|
| # ๐ฌ ForensiX AI |
|
|
| ### AI-Powered Forensic Triage & Postmortem Intelligence System |
|
|
| []() |
| []() |
| []() |
| []() |
| []() |
|
|
| *An intelligent investigative support system integrating GraphRAG, NLP, Multi-Agent AI, and custom fine-tuned models to assist forensic investigations with real-time analysis, evidence correlation, and predictive risk scoring.* |
|
|
| </div> |
|
|
| --- |
|
|
| ## ๐ Table of Contents |
|
|
| - [Problem Statement](#-problem-statement) |
| - [System Architecture](#-system-architecture) |
| - [Multi-Agent System](#-multi-agent-system-feat) |
| - [GraphRAG Knowledge Engine](#-graphrag-knowledge-engine) |
| - [Custom AI Models](#-custom-ai-models) |
| - [Forensic NLP Engine](#-forensic-nlp-engine) |
| - [Digital Evidence Correlation](#-digital-evidence-correlation) |
| - [Cross-Case Intelligence](#-cross-case-intelligence) |
| - [Time-of-Death Estimation](#-time-of-death-estimation) |
| - [Risk Scoring & Explainability](#-risk-scoring--explainability) |
| - [Universal LLM Provider](#-universal-llm-provider) |
| - [Frontend Architecture](#-frontend-architecture) |
| - [API Reference](#-api-reference) |
| - [Deployment](#-deployment) |
| - [Novelties & Innovation](#-novelties--innovation) |
| - [Ethical Considerations](#-ethical-considerations) |
| - [Future Scope](#-future-scope) |
|
|
| --- |
|
|
| ## ๐ฏ Problem Statement |
|
|
| Investigative agencies and forensic departments face critical challenges: |
|
|
| - **Volume overload**: Processing large amounts of forensic/digital evidence within limited time |
| - **Manual bottlenecks**: Autopsy reports, CCTV logs, mobile metadata analyzed separately by different teams |
| - **Missed connections**: Cross-evidence correlations invisible to siloed manual analysis |
| - **Delayed triage**: Critical leads buried under routine cases due to lack of intelligent prioritization |
| - **No standardization**: Each investigator approaches evidence differently, creating inconsistency |
|
|
| **ForensiX AI solves this** by providing an AI-powered unified platform that: |
| 1. Automatically extracts structured data from unstructured forensic reports |
| 2. Correlates physical evidence with digital traces in real-time |
| 3. Identifies anomalies and generates investigative leads |
| 4. Provides explainable risk scoring for case prioritization |
| 5. Operates with full chain-of-custody integrity (SHA-256 ledger) |
|
|
| --- |
|
|
| ## ๐๏ธ System Architecture |
|
|
| ``` |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| โ ForensiX AI v4.0 Architecture โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค |
| โ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ PRESENTATION LAYER โ โ |
| โ โ React 18 + Vite + Tailwind + Zustand + Framer Motion + Recharts โ โ |
| โ โ โ โ |
| โ โ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โ โ |
| โ โ โDashboard โ โ Timeline โ โ Evidence โ โ Agents โ โ AI Chat โ โ โ |
| โ โ โ โ โ โ โ Graph โ โ โ โ (GraphRAG)โ โ โ |
| โ โ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โ โ |
| โ โ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โ โ |
| โ โ โ Risk โ โ Autopsy โ โ Cases โ โ Custody โ โ Settings โ โ โ |
| โ โ โ Scoring โ โWorkspace โ โ CRUD โ โ Chain โ โ LLM Cfg โ โ โ |
| โ โ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โโโโโโโโโโโโ โ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ REST + WebSocket + SSE โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ APPLICATION LAYER (FastAPI) โ โ |
| โ โ โ โ |
| โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ |
| โ โ โ MULTI-AGENT ORCHESTRATOR (7 Agents) โ โ โ |
| โ โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ โ |
| โ โ โ โ Autopsy โโTimeline โโ Digital โโToxicol. โโ Risk โ โ โ โ |
| โ โ โ โ Agent โโ Agent โโ Agent โโ Agent โโ Agent โ โ โ โ |
| โ โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ โ |
| โ โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ โ |
| โ โ โ โ Correlation Agent โโ Explainability Agent (SHAP) โ โ โ โ |
| โ โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ โ |
| โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ |
| โ โ โ โ |
| โ โ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโ โ โ |
| โ โ โ GraphRAG โ โ Forensic โ โ Cross-Case โ โ Universal โ โ โ |
| โ โ โ Knowledge โ โ NLP Engine โ โ Intelligence โ โ LLM Prov. โ โ โ |
| โ โ โ (25 nodes) โ โ (Henssge) โ โ (Serial Det.)โ โ (12+ APIs)โ โ โ |
| โ โ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโ โ โ |
| โ โ โ โ |
| โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ |
| โ โ โ CUSTOM AI MODELS (Offline, No API Keys) โ โ โ |
| โ โ โ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโโ โ โ โ |
| โ โ โ โ GLiNER-biomedโ โ DeBERTa-v3 โ โ Qwen2.5-0.5B+LoRA โ โ โ โ |
| โ โ โ โ Zero-shot NERโ โ Classifier โ โ Extraction Model โ โ โ โ |
| โ โ โ โ (280MB) โ โ (180MB) โ โ (1.2GB) โ โ โ โ |
| โ โ โ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโโ โ โ โ |
| โ โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ SQLAlchemy + REST โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ DATA LAYER โ โ |
| โ โ โโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ |
| โ โ โ SQLite โ โ Supabase โ โ Chain of Custody โ โ โ |
| โ โ โ (Local) โ โ PostgreSQL + โ โ SHA-256 Immutable โ โ โ |
| โ โ โ โ โ pgvector (Cloud) โ โ Ledger โ โ โ |
| โ โ โโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| ``` |
|
|
| --- |
|
|
| ## ๐ค Multi-Agent System (FEAT) |
|
|
| **Forensic Evidence Analysis & Triage** โ 7 specialized AI agents that independently analyze evidence and collaborate through a master orchestrator. |
|
|
| ``` |
| โโโโโโโโโโโโโโโโโโโโโโโโ |
| โ FEAT ORCHESTRATOR โ |
| โ Sequential+Parallel โ |
| โโโโโโโโโโโโฌโโโโโโโโโโโโ |
| โ |
| โโโโโโโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโ |
| โ โ โ |
| โโโโโโดโโโโโ โโโโโโดโโโโโ โโโโโโดโโโโโ |
| โ AUTOPSY โ โTIMELINE โ โ DIGITAL โ |
| โ AGENT โ โ AGENT โ โ AGENT โ |
| โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโโ |
| โโข NER โ โโข Gap โ โโข CCTV โ |
| โโข COD โ โ detect โ โโข Mobile โ |
| โโข Injuriesโ โโข Clusterโ โโข GPS โ |
| โโข Toxicol.โ โโข Sequenceโ โโข Patternโ |
| โโโโโโฌโโโโโ โโโโโโฌโโโโโ โโโโโโฌโโโโโ |
| โ โ โ |
| โโโโโโโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโ |
| โ |
| โโโโโโโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโ |
| โ โ โ |
| โโโโโโดโโโโโ โโโโโโดโโโโโ โโโโโโดโโโโโ |
| โTOXICOL. โ โ RISK โ โEXPLAIN- โ |
| โ AGENT โ โ AGENT โ โ ABILITY โ |
| โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโโ |
| โโข Drug IDโ โโข 6-factorโ โโข SHAP โ |
| โโข Levels โ โ weightedโ โโข LIME โ |
| โโข Lethalityโ โโข Scoring โ โโข Attrib.โ |
| โโโโโโฌโโโโโ โโโโโโฌโโโโโ โโโโโโฌโโโโโ |
| โ โ โ |
| โโโโโโโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโ |
| โ |
| โโโโโโโโโโโโดโโโโโโโโโโโโ |
| โ CORRELATION AGENT โ |
| โ โข Temporal links โ |
| โ โข Spatial patterns โ |
| โ โข Causal inference โ |
| โ โข Behavioral flags โ |
| โโโโโโโโโโโโโโโโโโโโโโโโ |
| ``` |
|
|
| ### Agent Details |
|
|
| | Agent | Purpose | Confidence | Key Findings | |
| |-------|---------|-----------|--------------| |
| | **Autopsy Agent** | Extract COD, manner, injuries, toxicology from reports | 87-95% | Injuries, COD, toxicology entities | |
| | **Timeline Agent** | Reconstruct chronological event sequence | 84% | Gaps (>30min), clusters (<5min) | |
| | **Digital Agent** | Analyze CCTV, mobile, GPS evidence | 86% | Person discrepancy, rapid departure | |
| | **Toxicology Agent** | Interpret drug/poison findings | 90% | Sedation indicators, lethality | |
| | **Risk Agent** | 6-factor weighted risk scoring | 85% | Score 0-100, severity level | |
| | **Explainability Agent** | SHAP-style factor attribution | 82% | Top contributing factors | |
| | **Correlation Agent** | Cross-evidence relationship discovery | 88-94% | Temporal/causal/behavioral links | |
|
|
| --- |
|
|
| ## ๐ง GraphRAG Knowledge Engine |
|
|
| A Retrieval-Augmented Generation system with **25 verified forensic knowledge nodes** covering: |
|
|
| ``` |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| โ GRAPHRAG ARCHITECTURE โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค |
| โ โ |
| โ โโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโ โ |
| โ โ QUERY โโโโโโถโ EMBEDDING โโโโโโถโ RETRIEVAL โ โ |
| โ โ (User/LLM) โ โ (384-dim) โ โ (Top-K) โ โ |
| โ โโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโฌโโโโโโโ โ |
| โ โ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โผ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ FORENSIC KNOWLEDGE BASE โ โ |
| โ โ โ โ |
| โ โ POSTMORTEM (5): โ โ |
| โ โ livor mortis, rigor mortis, algor mortis, โ โ |
| โ โ PMI estimation, decomposition stages โ โ |
| โ โ โ โ |
| โ โ INJURY PATTERNS (8): โ โ |
| โ โ petechiae, defensive wounds, blunt force, โ โ |
| โ โ ligature, sharp force, gunshot, hyoid fracture, โ โ |
| โ โ wound tracks โ โ |
| โ โ โ โ |
| โ โ MANNER DETECTION (4): โ โ |
| โ โ classification, suicide criteria, staging, โ โ |
| โ โ drowning, fire deaths โ โ |
| โ โ โ โ |
| โ โ TOXICOLOGY (5): โ โ |
| โ โ CO poisoning, benzodiazepines, opioids, โ โ |
| โ โ alcohol, specimen collection โ โ |
| โ โ โ โ |
| โ โ EVIDENCE (3): โ โ |
| โ โ CCTV analysis, mobile evidence, Locard exchange โ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ โ |
| โ โผ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ CONTEXT INJECTION โ LLM System Prompt โ โ |
| โ โ "VERIFIED FORENSIC KNOWLEDGE โ do not contradict" โ โ |
| โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ |
| โ โ |
| โ RETRIEVAL STRATEGY (in order): โ |
| โ 1. Supabase pgvector (semantic, production) โ |
| โ 2. HuggingFace embeddings (sentence-transformers) โ |
| โ 3. Keyword fallback (TF-IDF, always works) โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| ``` |
|
|
| --- |
|
|
| ## ๐งช Custom AI Models |
|
|
| ### Model Stack (runs offline, no API keys needed) |
|
|
| ``` |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| โ LOCAL ML PIPELINE โ ForensiX Custom Models โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค |
| โ โ |
| โ Layer 1: ZERO-SHOT NER (GLiNER-biomed, 280MB) โ |
| โ โโโ Define ANY entity type at runtime โ |
| โ โโโ "cause of death", "weapon type", "time indicator" โ |
| โ โโโ No retraining needed โ labels via natural language โ |
| โ โ |
| โ Layer 2: BIOMEDICAL NER (OpenMed, ~109MB each) โ |
| โ โโโ Toxicology: drugs, chemicals, levels (F1=0.96) โ |
| โ โโโ Anatomy: body parts, organs (F1=0.91) โ |
| โ โโโ Pathology: diseases, conditions (F1=0.91) โ |
| โ โ |
| โ Layer 3: MANNER CLASSIFIER (DeBERTa-v3-small, 180MB) โ |
| โ โโโ Fine-tuned on 750 synthetic forensic samples โ |
| โ โโโ 5 classes: HOMICIDE/SUICIDE/ACCIDENTAL/NATURAL/UNDET. โ |
| โ โโโ <100ms inference, F1>0.90 โ |
| โ โ |
| โ Layer 4: EXTRACTION MODEL (Qwen2.5-0.5B+LoRA, 1.2GB) โ |
| โ โโโ Fine-tuned on 300 forensic Q&A pairs โ |
| โ โโโ Outputs structured JSON โ |
| โ โโโ CPU inference ~3-5s per report โ |
| โ โ |
| โ Total RAM: ~3-5GB (NER stack) or ~5-7GB (with LLM) โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| ``` |
|
|
| ### Training Pipeline |
|
|
| ```bash |
| # 1. Generate synthetic training data |
| python custom-models/scripts/generate_dataset.py |
| # โ 750 classification samples + 300 extraction Q&A pairs |
| |
| # 2. Train manner-of-death classifier |
| python custom-models/scripts/train_classifier.py |
| # โ DeBERTa-v3-small, 5 epochs, ~180MB output |
| |
| # 3. Train forensic extraction model |
| python custom-models/scripts/train_extraction_model.py |
| # โ Qwen2.5-0.5B + LoRA, 3 epochs, ~50MB adapter |
| ``` |
|
|
| --- |
|
|
| ## ๐ Forensic NLP Engine |
|
|
| ### Entity Extraction Pipeline |
|
|
| ```python |
| # Input: "Blunt force trauma to the right temporal region with subdural hematoma. |
| # Defensive wounds on forearms. Benzodiazepines detected." |
| |
| # Output: |
| { |
| "entities": [ |
| {"text": "Blunt force trauma to the right temporal region", "label": "INJURY", "confidence": 0.92}, |
| {"text": "subdural hematoma", "label": "INJURY", "confidence": 0.88}, |
| {"text": "Defensive wounds on forearms", "label": "INJURY", "confidence": 0.90}, |
| {"text": "Benzodiazepines detected", "label": "TOXICOLOGY", "confidence": 0.85} |
| ], |
| "riskScore": 100, |
| "riskLevel": "CRITICAL", |
| "anomalies": [{"type": "sedation_indicator", "severity": "HIGH"}] |
| } |
| ``` |
|
|
| ### Pattern Recognition |
|
|
| | Pattern | Detection | Confidence | |
| |---------|-----------|-----------| |
| | CAUSE_OF_DEATH | Regex + LLM extraction | 85-95% | |
| | MANNER_OF_DEATH | Classification + keywords | 90-95% | |
| | INJURY | 8 regex patterns + NER | 88% | |
| | TOXICOLOGY | Drug/substance detection | 90% | |
| | TIME_INDICATOR | Postmortem signs parsing | 85% | |
| | EVIDENCE | Physical trace detection | 80% | |
| | ANOMALIES | Cross-reference contradiction detection | 92% | |
| |
| --- |
| |
| ## ๐ก Digital Evidence Correlation |
| |
| ``` |
| EVIDENCE TIMELINE |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| 01:45 CCTV: Two individuals enter warehouse โโโโโโ |
| 01:52 CCTV: Vehicle parks near entrance โ CLUSTER |
| 02:00 Mobile: Victim last call โโโโโโโโโโโโโโโโโโโ (<15min) |
| |
| 02:00 โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ |
| โโโ 45-MIN GAP โโโ โ ANOMALY |
| 02:45 โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ โ |
| |
| 02:47 CCTV: Single person exits at high speed โโโโ |
| 02:48 Mobile: Phone signal lost (powered off) โ PATTERN |
| 02:50 CCTV: Vehicle departs rapidly โโโโโโโโโโโโโโ |
| |
| CORRELATIONS FOUND: |
| โ ๏ธ PERSON DISCREPANCY: 2 entered, 1 departed (CRITICAL) |
| โ ๏ธ COMMUNICATION CUTOFF: Phone disconnected (HIGH) |
| โ ๏ธ RAPID DEPARTURE: Vehicle at unusual speed (HIGH) |
| ``` |
| |
| --- |
| |
| ## ๐ต๏ธ Cross-Case Intelligence |
| |
| Serial pattern detection against historical case database: |
| |
| ``` |
| CURRENT CASE HISTORICAL MATCH (93% similarity) |
| โโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| Manner: Homicide โโโโถ CASE-2024-0103 (93%) |
| Weapons: ligature+blunt โโโโถ Weapons: blunt+ligature โ |
| Injuries: strangulation โโโโถ Injuries: head_trauma+strangulation โ |
| Toxicology: diazepam โโโโถ Toxicology: diazepam โ |
| Location: industrial โโโโถ Location: commercial โ |
| MO: sedation+violence โโโโถ MO: sedation+violence+abandoned โ |
| Time: night โโโโถ Time: night โ |
| Victim: adult male โโโโถ Victim: adult male โ |
|
|
| โ ๏ธ SERIAL PATTERN DETECTED โ Cross-reference with cold case unit |
| ``` |
| |
| --- |
| |
| ## โฑ๏ธ Time-of-Death Estimation |
| |
| ### Henssge Nomogram (Double-Exponential Model) |
| |
| ``` |
| Formula: Q = 1.25ยทe^(-Bยทt) - 0.25ยทe^(-5Bยทt) |
|
|
| Where: |
| Q = (T_rectal - T_ambient) / (37.2 - T_ambient) |
| B = 1.2815 ยท (W_effective)^(-0.625) + 0.0284 |
| W_effective = Corrective_Factor ร Body_Weight |
| |
| Multi-Method Approach: |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| โ Method โ Weight โ Conf. โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค |
| โ Henssge Nomogram โ 70% โ HIGH โ |
| โ Rigor Mortis Stage โ 15% โ MOD โ |
| โ Livor Mortis State โ 10% โ MOD โ |
| โ Vitreous Potassium โ 5% โ HIGH โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| |
| Output: PMI = 6.9h ยฑ 2.8h (95% CI) |
| ``` |
| |
| --- |
| |
| ## ๐ Risk Scoring & Explainability |
| |
| ### 6-Factor Weighted Algorithm |
| |
| ``` |
| Risk Score = ฮฃ (Factor_Score ร Weight) |
|
|
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| โ Factor โ Weight โ Score โ Contribution โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค |
| โ Violence Level โ 25% โ 85/100โ 21.3 pts โ |
| โ Digital Patterns โ 20% โ 90/100โ 18.0 pts โ |
| โ Manner Match โ 15% โ 95/100โ 14.3 pts โ |
| โ Evidence Gaps โ 15% โ 75/100โ 11.3 pts โ |
| โ Temporal Cluster โ 15% โ 70/100โ 10.5 pts โ |
| โ Toxicology โ 10% โ 60/100โ 6.0 pts โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค |
| โ TOTAL โ 100% โ โ 81.4/100 โ |
| โ Level: CRITICAL โ โ โ โ |
| โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| ``` |
| |
| ### SHAP Attribution (Explainable AI) |
| |
| ``` |
| Feature Importance (case #FTI-2024-0847): |
|
|
| Violence Level โโโโโโโโโโโโโโโโโโโโโโ 85% (+21.3) |
| Digital Patterns โโโโโโโโโโโโโโโโโโโโโโ 90% (+18.0) |
| Manner Classification โโโโโโโโโโโโโโโโโโโโโโ 95% (+14.3) |
| Evidence Gaps โโโโโโโโโโโโโโโโโโโโโโ 75% (+11.3) |
| Temporal Clustering โโโโโโโโโโโโโโโโโโโโโโ 70% (+10.5) |
| Toxicology โโโโโโโโโโโโโโโโโโโโโโ 60% (+ 6.0) |
| ``` |
| |
| --- |
| |
| ## ๐ Universal LLM Provider |
| |
| Supports **12+ LLM providers** via OpenAI-compatible API: |
| |
| | Provider | Speed | Quality | Cost | |
| |----------|-------|---------|------| |
| | Featherless.ai | Medium | High (Llama-3.1-70B) | Free tier | |
| | Groq | Ultra-fast | High (llama-3.3-70b) | Free tier | |
| | Together.ai | Fast | High (Llama-3.1-70B-Turbo) | Paid | |
| | OpenAI | Medium | Highest (GPT-4o) | Paid | |
| | HuggingFace | Slow | High (Qwen2.5-72B) | Free | |
| | TokenRouter | Auto | Auto (cheapest route) | Variable | |
| | DeepSeek | Medium | High (deepseek-chat) | Paid | |
| | Ollama | Local | Variable | Free | |
| | OpenRouter | Variable | Any model | Variable | |
| |
| **Fallback chain**: If primary fails โ tries next provider โ graceful degradation to local analysis. |
| |
| --- |
| |
| ## ๐ฅ๏ธ Frontend Architecture |
| |
| ``` |
| src/ |
| โโโ App.jsx # Router + route definitions |
| โโโ main.jsx # Entry point |
| โโโ pages/ |
| โ โโโ Landing.jsx # Hero + feature grid |
| โ โโโ Login.jsx # JWT authentication |
| โ โโโ Dashboard.jsx # KPI cards + activity feed |
| โ โโโ ForensicCommandDashboard.jsx # Full command center |
| โ โโโ InvestigationWorkspace.jsx # Main workspace (panels) |
| โ โโโ CaseAnalysis.jsx # Deep-dive case view |
| โ โโโ Timeline.jsx # Forensic timeline |
| โ โโโ EvidenceGraph.jsx # Force-directed graph |
| โ โโโ RiskScoring.jsx # SHAP + radar charts |
| โ โโโ Agents.jsx # Multi-agent dashboard |
| โ โโโ InvestigationQuery.jsx # NL query + AI Chat |
| โ โโโ ChainOfCustody.jsx # SHA-256 ledger |
| โโโ components/ |
| โ โโโ layout/ # App shell (header, nav, panels) |
| โ โโโ views/ # Embedded views (autopsy, evidence, etc.) |
| โ โโโ forensic/ # RiskGauge, SHAP chart, suspect card |
| โ โโโ modals/ # Upload, export, review, search |
| โ โโโ panels/ # Event details, evidence preview |
| โ โโโ timeline/ # Forensic timeline component |
| โโโ lib/ |
| โโโ api.js # 50+ API endpoints connected |
| โโโ store.js # Zustand global state |
| โโโ websocket.js # Real-time connection |
| ``` |
| |
| --- |
|
|
| ## ๐ก API Reference |
|
|
| ### Core Endpoints (19 routers, 55+ routes) |
|
|
| | Method | Endpoint | Description | |
| |--------|----------|-------------| |
| | `GET` | `/health` | System health + version | |
| | `GET` | `/api/status` | Full system status (12 features) | |
| | `POST` | `/api/intelligence/` | Unified intelligence engine | |
| | `POST` | `/api/chat/` | AI Chat with GraphRAG | |
| | `POST` | `/api/analyze/` | Full NLP + LLM + GraphRAG analysis | |
| | `POST` | `/api/ml/` | ML Pipeline (NER, classify, embed) | |
| | `POST` | `/api/graphrag/retrieve` | Knowledge retrieval | |
| | `GET` | `/api/graphrag/stats` | Knowledge base stats | |
| | `GET` | `/api/triage/stream` | SSE progress events | |
| | `GET` | `/cases/` | List all cases | |
| | `GET` | `/risk/{id}` | Risk assessment | |
| | `GET` | `/timeline/{id}` | Timeline events | |
| | `GET` | `/graph/{id}` | Evidence graph | |
| | `GET` | `/agents/status` | Agent status (7 agents) | |
| | `GET` | `/agents/analysis/{id}` | Full multi-agent analysis | |
| | `POST` | `/query/` | Natural language query | |
| | `GET` | `/custody/{id}` | Chain of custody | |
| | `WS` | `/ws/{id}` | Real-time WebSocket | |
|
|
| --- |
|
|
| ## ๐ Deployment |
|
|
| ### Docker (Production) |
|
|
| ```bash |
| docker build -t forensix-ai . |
| docker run -p 7860:7860 \ |
| -e FEATHERLESS_API_KEY=your_key \ |
| -e HF_TOKEN=your_token \ |
| forensix-ai |
| ``` |
|
|
| ### HuggingFace Spaces |
|
|
| Push to any HF Space with Docker SDK โ auto-deploys. |
|
|
| ### Manual |
|
|
| ```bash |
| pip install -r backend/requirements.txt |
| cd frontend && npm install && npm run build && cd .. |
| uvicorn backend.main:app --host 0.0.0.0 --port 7860 |
| ``` |
|
|
| --- |
|
|
| ## ๐ก Novelties & Innovation |
|
|
| ### 1. Digital Stratigraphy |
| Unlike existing forensic systems that separately analyze pathology reports OR digital evidence, ForensiX introduces a **unified multimodal architecture** that semantically correlates ALL evidence types into synchronized investigative layers. |
|
|
| ### 2. Zero-Shot Forensic NER |
| Using GLiNER-biomed, investigators can define **ANY custom entity type** at runtime using natural language โ no model retraining needed. Want to detect "ligature material type"? Just add it to the label list. |
|
|
| ### 3. GraphRAG-Enhanced AI Chat |
| Every AI response is grounded in **verified forensic knowledge** retrieved via semantic search. The LLM cannot hallucinate forensic facts because the system prompt is injected with authoritative knowledge nodes. |
|
|
| ### 4. Cross-Case Serial Pattern Detection |
| Automated matching against historical case database identifies potential serial offenders by comparing MO patterns, weapon types, victim profiles, and behavioral signatures. |
|
|
| ### 5. Explainable Risk Scoring |
| Every risk score is decomposable into SHAP-style attributions. Investigators can see EXACTLY which evidence drives the score โ legally defensible transparency. |
|
|
| ### 6. Universal LLM Provider |
| Works with ANY provider (12+ supported). Investigators aren't locked into one AI vendor. Seamless fallback chain ensures the system never goes dark. |
|
|
| ### 7. Offline-Capable Custom Models |
| Fine-tuned models run entirely on CPU without internet. Critical for sensitive investigations where data cannot leave the network. |
|
|
| --- |
|
|
| ## โ๏ธ Ethical Considerations |
|
|
| - **Advisory only**: All outputs support human decision-making, never replace forensic experts |
| - **No legal conclusions**: System uses "suggests", "indicates", "consistent with" โ never "proves" |
| - **Explainable**: Every AI output traceable to specific evidence (SHAP attribution) |
| - **Chain of custody**: SHA-256 immutable ledger ensures evidence integrity |
| - **Data privacy**: Local-first architecture, cloud optional, no mandatory data sharing |
| - **Bias awareness**: Models trained on balanced synthetic data to avoid demographic bias |
| - **Access control**: JWT authentication, role-based permissions |
|
|
| --- |
|
|
| ## ๐ฎ Future Scope |
|
|
| | Enhancement | Status | Description | |
| |-------------|--------|-------------| |
| | Real-time CCTV analysis | Planned | Live video feed processing with person detection | |
| | Multilingual report analysis | Planned | Hindi, Tamil, Spanish autopsy report parsing | |
| | Federated learning | Research | Secure cross-agency model training without data sharing | |
| | IoT sensor integration | Planned | Real-time environmental monitoring (temp, humidity) | |
| | Voice-to-text dictation | Planned | Hands-free field notes during investigation | |
| | Mobile companion app | Planned | On-scene evidence collection with GPS tagging | |
| | Behavioral profiling | Research | LSTM-based suspect behavior prediction | |
| | 3D wound reconstruction | Research | AR-based injury pattern visualization | |
|
|
| --- |
|
|
| ## ๐ Project Structure |
|
|
| ``` |
| forensic-ai-triage/ |
| โโโ backend/ |
| โ โโโ main.py # FastAPI app (55+ routes) |
| โ โโโ requirements.txt # Python dependencies |
| โ โโโ db/ # SQLAlchemy models + init |
| โ โโโ routers/ # 19 API routers |
| โ โ โโโ agents.py # Multi-agent endpoints |
| โ โ โโโ analyze.py # Full analysis (NLP+LLM+RAG) |
| โ โ โโโ chat.py # AI Chat with GraphRAG |
| โ โ โโโ intelligence.py # Unified intelligence engine |
| โ โ โโโ ml_pipeline.py # HuggingFace ML pipeline |
| โ โ โโโ graphrag.py # Knowledge base management |
| โ โ โโโ triage_stream.py # SSE progress events |
| โ โ โโโ ... (12 more routers) |
| โ โโโ services/ # Business logic |
| โ โโโ llm_provider.py # Universal LLM (12+ providers) |
| โ โโโ graphrag_service.py # 25-node knowledge base |
| โ โโโ multi_agent_orchestrator.py # 7-agent system |
| โ โโโ forensic_engine.py # Henssge + NLP + correlation |
| โ โโโ supabase_client.py # Cloud persistence |
| โ โโโ ... (8 more services) |
| โโโ frontend/ |
| โ โโโ src/ |
| โ โ โโโ pages/ (12 pages) |
| โ โ โโโ components/ (28 components) |
| โ โ โโโ lib/ (api.js, store.js, websocket.js) |
| โ โโโ package.json |
| โโโ custom-models/ # ๐ Custom AI Models |
| โ โโโ datasets/ # Training data (750+300 samples) |
| โ โโโ scripts/ # Training + inference code |
| โ โ โโโ generate_dataset.py # Synthetic data generator |
| โ โ โโโ train_classifier.py # DeBERTa-v3-small fine-tuning |
| โ โ โโโ train_extraction_model.py # Qwen2.5-0.5B + LoRA |
| โ โ โโโ inference.py # Unified local inference |
| โ โโโ models/ # Trained weights (after training) |
| โโโ lib/ # TypeScript engines (reference) |
| โโโ multi-agent-system/ # Full TS agent implementation |
| โโโ langgraph-agents/ # LangGraph Python server |
| โโโ prisma/ # PostgreSQL schema |
| โโโ Dockerfile # Multi-stage production build |
| โโโ nginx.conf # Reverse proxy config |
| โโโ .env.example # All provider configs |
| โโโ README.md # This file |
| ``` |
|
|
| --- |
|
|
| ## ๐ Performance Metrics |
|
|
| | Metric | Value | |
| |--------|-------| |
| | API Endpoints | 55+ routes, all verified 200 OK | |
| | Multi-Agent Analysis | 7 agents, ~16s total (with LLM) | |
| | NLP Entity Extraction | <100ms per report (regex) | |
| | Risk Scoring | <50ms computation | |
| | Graph Building | <200ms for full evidence graph | |
| | WebSocket Latency | <50ms real-time updates | |
| | Local ML Inference | <500ms (NER), <100ms (classifier) | |
| | Knowledge Retrieval | <10ms (keyword), <200ms (semantic) | |
|
|
| --- |
|
|
| ## ๐ก๏ธ Security |
|
|
| - **Authentication**: JWT (HS256) with configurable secret |
| - **Evidence Integrity**: SHA-256 hash chain (blockchain-style) |
| - **Data Isolation**: SQLite per-instance, no shared state |
| - **API Security**: CORS configured, rate limiting available |
| - **Sensitive Data**: All LLM calls use local context โ no case data sent to external APIs unless investigator explicitly uses AI Chat |
|
|
| --- |
|
|
| <div align="center"> |
|
|
| **Built for Justice. Powered by AI. Guided by Ethics.** |
|
|
| *ForensiX AI โ Investigative Assistance Only. Not for legal determination.* |
|
|
| </div> |
|
|