swift-fully

byhrmn

Build a fully working web app (not a mockup) for Smart India Hackathon 2026, PS 26189 — "AI-Powered Criminal Network Analysis System" by TEAM UPLIFT. PURPOSE Fuse fragmented police data (FIRs, call records, financial transactions, surveillance logs, social media, criminal databases) into one knowledge graph, extract entities and relationships, surface key persons and suspicious patterns, and present it all in a visual investigator console. STACK React + TypeScript + Vite + Tailwind, Cytoscape.js (graph), Recharts (charts), Leaflet + OpenStreetMap (maps). FastAPI backend with NetworkX, PyTorch Geometric, scikit-learn and spaCy. Neo4j + PostgreSQL with an in-memory fallback so it boots with zero setup. Docker compose, no API keys. DESIGN Dark intelligence console. Background #0E1A2B, panels #132840, borders #22496F, text #E6EDF5, accent orange #E8833A, green #3FBF7F, amber #E0A93B, blue #4C9BE8. Condensed uppercase headers, dense 12px text, thin borders, compact tables. Top bar: shield emblem, title "INTELLIGENCE ANALYTICS & CASE MANAGEMENT SYSTEM", notification bell, avatar. Sidebar: Dashboard, Cases, Subjects (X/Y), Evidence Locker, Link Analysis, Timeline, Geospatial Map, Reports, Settings. Responsive. FEATURES — every one must actually function INGEST — Upload CSV/JSON/PDF/TXT/images. Live pipeline: Ingest → OCR/NLP → Entity Extract → Relation Extract → Entity Resolution → Graph. Review table where the officer can Accept, Edit or Reject each extraction. Fuzzy duplicate merging ("Ravi Kumar" = "R. Kumar"). GRAPH — Interactive canvas with Person, Phone, Location, Vehicle, Org, Account and Event nodes. Pan, zoom, click-to-expand neighbours, right-click menu, four switchable layouts, shortest-path finder, filters by type/date/confidence, export PNG and JSON. KEY PERSONS — Degree, betweenness, closeness, eigenvector and PageRank in a sortable table. Node size scales by the selected metric; a "Key Person" badge applies above a threshold. PATTERNS — Louvain community detection with coloured clusters, plus anomaly rules (call bursts, money structuring, circular transfers, unusual co-location, bridge nodes) that raise real alerts with explanations. PREDICT — GNN link prediction with an Adamic-Adar fallback so the demo never fails. Predicted links render as dashed edges with probability labels, queued for an officer to Confirm (becomes a real edge) or Dismiss. COMPARE (hero screen) — Subject X vs Subject Y. Two profile cards (name, DOB, aliases, last known address, reported suspicion, behavioural indicators, key evidence) beside a Comparison View with working tabs: Similarities & Differences, Timelines, Relationships, Evidence Strength, Confidence Levels. Live charts: Evidence bar chart, Shared Locations area chart, Evidence Strength bars, Confidence Level horizontal bars, and a mini link graph. Include a computed table — Verified connections 3 vs 5 (CONFIRMED FACT), AI-predicted links 1 vs 2 (AI INFERENCE), Network centrality 0.42 vs 0.71 (AI INDICATOR) — plus common contacts, shared locations, a similarity score and an auto-written summary. Export to PDF. EVOLVE — Event timeline with a date-range slider and Play button that animates the graph rebuilding over time; metrics recompute per window. LOCATE — Map with clustered markers for incidents, sightings, towers and jails. Heatmap toggle, draw-a-radius "who was here between these dates" query, time-synced to the timeline. Clicking a jail opens a facility panel with location, current population and a searchable inmate table (Name, Detail, Date In, Jail ID) linking back to that person's graph node. COLLABORATE — JWT login with five roles (Admin, Senior Investigator, Investigator, Analyst, Auditor) with genuinely different permissions. Comments and @mentions on entities, notifications, live multi-user updates over WebSockets, and a tamper-evident audit log (SHA-256 hash chain with a Verify Integrity button). EXPLAIN — Every AI output has a "Why this?" panel showing contributing evidence, feature weights and a confidence percentage. REPORTS — Generate a downloadable PDF case dossier: graph snapshot, key persons, detected patterns, X-vs-Y comparison, timeline, map and predicted links. EVIDENCE LABELS Tag every piece of evidence everywhere with one of four chips: CONFIRMED FACT (green), ALLEGATION/REPORT (amber), AI INFERENCE (blue), UNVERIFIED (grey). Separating verified fact from AI guess is the core value of this product. SEED DATA Generate a synthetic trafficking and hawala network: 120 persons, 150 phones, 60 Indian locations, 40 vehicles, 25 organisations, 8 jails with inmate rosters, 5,000 call records, 1,500 transactions and 60 FIR documents across 24 months. Deliberately plant three communities, two broker nodes, one money-structuring pattern, one call-burst anomaly and four discoverable missing links. Pre-wire Subject X and Subject Y as a strong comparison pair. Footer on every page: "Illustrative synthetic data — not real persons or cases." RULES No dead buttons, no placeholder images, no "coming soon" — every control must change real state from real data. Include a README with three-command setup, a seed script, and a Demo Mode that replays a three-minute judging walkthrough. Tagline: "PS asks: connect the dots. TEAM UPLIFT adds: Compare · Predict · Evolve · Locate · Collaborate · Explain."