Skip to content

TASKS.md

Latest commit

 

History

History
266 lines (216 loc) · 11.8 KB

TASKS.md

File metadata and controls

266 lines (216 loc) · 11.8 KB

JARVIS — Build Tasks

Ordered by critical path. Check off as you go.

PHASE 1: Foundation (Hours 0-4)

  • T01: Convex project setup

    • npx create-convex in frontend/
    • Copy schema from TECH_DOC.md → convex/schema.ts
    • Create mutations: persons.create, persons.update, captures.create
    • Create queries: persons.listAll, persons.getById, connections.listAll, intel.recentActivity
    • Verify with npx convex dev — should deploy without errors
    • Acceptance: Can create + query a person from Convex dashboard

  • T02: FastAPI skeleton

    • pip install fastapi uvicorn python-dotenv pydantic-settings
    • Create backend/main.py with /api/health, /api/capture stubs
    • Create backend/config.py with Settings class (see SYSTEM_DESIGN.md §8)
    • Create .env from .env.example
    • Acceptance: uvicorn main:app --reload starts, /api/health returns OK

  • T03: Next.js project setup

    • npx create-next-app frontend --typescript --tailwind --app
    • Install: framer-motion, convex, @use-gesture/react, lucide-react
    • Add Google Fonts (Bebas Neue, Courier Prime, Inter) to layout.tsx
    • Add CSS variables from DESIGN_HANDOFF.md §1 to globals.css
    • Wire up Convex provider in layout.tsx
    • Acceptance: npm run dev shows page with correct fonts

  • T04: Convex ↔ Frontend proof of life

    • Create a test button that inserts a dummy person via Convex mutation
    • Display the person's name on the page via useQuery
    • Verify real-time: change data in Convex dashboard → page updates instantly
    • Acceptance: Button click → name appears in <1 second

PHASE 2: Identification Pipeline (Hours 4-8)

  • T05: PimEyes Browser Use agent

    • Create backend/identification/pimeyes.py (see SYSTEM_DESIGN.md §2.3)
    • Implement PimEyesSearcher class with Browser Use SDK
    • Navigate to pimeyes.com, upload image, extract results page screenshot
    • Use Gemini 2.0 Flash to parse screenshot → names + URLs
    • Acceptance: Given a face photo, returns [{name, confidence, urls}]

  • T06: Account pool rotation

    • Create backend/identification/pool.py (see SYSTEM_DESIGN.md §2.4)
    • Implement AccountPool with 3+ PimEyes accounts
    • 30-second cooldown per account, round-robin selection
    • Acceptance: 5 sequential searches use different accounts, no rate limits

  • T07: Capture service + Telegram bot

    • Create backend/capture/service.py (see SYSTEM_DESIGN.md §1.3)
    • Create backend/capture/telegram_bot.py (see SYSTEM_DESIGN.md §1.4)
    • ffmpeg frame extraction from video
    • mediapipe face detection on extracted frames
    • POST detected faces to /api/capture
    • Acceptance: Send video to Telegram bot → faces detected → Convex shows new capture

  • T08: End-to-end identification pipeline

    • Wire together: capture → PimEyes → Gemini parse → create Person in Convex
    • Test with a real photo of a known person
    • Acceptance: Upload photo → Person card appears in Convex with name + confidence

PHASE 3: Agent Swarm (Hours 8-14)

  • T09: Exa fast-pass enrichment

    • Create backend/enrichment/exa_client.py (see SYSTEM_DESIGN.md §3.4)
    • Search person name + context → extract LinkedIn URL, company, title
    • Write results as intel fragments to Convex
    • Acceptance: Name search returns structured person data in <500ms

  • T10: LinkedIn Browser Use agent

    • Create backend/agents/linkedin_agent.py (see SYSTEM_DESIGN.md §3.2)
    • Navigate to LinkedIn profile URL (from Exa or PimEyes)
    • Extract: headline, about, experience, education, skills, recent posts
    • Use Voyager API interception where possible for speed
    • Acceptance: Given LinkedIn URL → returns structured profile data

  • T11: Twitter/X agent (twscrape)

    • Create backend/agents/twitter_agent.py (see SYSTEM_DESIGN.md §3.3)
    • Use twscrape for reverse GraphQL extraction
    • Fallback to Browser Use if twscrape fails
    • Extract: bio, recent tweets, followers, engagement
    • Acceptance: Given Twitter handle → returns structured profile + recent tweets

  • T12: Google search agent

    • Create backend/agents/google_agent.py
    • Browser Use agent: search "{name} {company}" on Google
    • Extract top 5 result titles + snippets
    • Feed into Exa for deeper page-level extraction if relevant
    • Acceptance: Returns relevant search results for a person

  • T13: Swarm orchestrator

    • Create backend/agents/orchestrator.py (see SYSTEM_DESIGN.md §3.1)
    • Two-tier execution: Exa first (fast), then Browser Use agents in parallel
    • Stream results to Convex as they arrive (don't wait for all agents)
    • 180-second timeout per agent
    • Acceptance: Trigger research → Convex shows data arriving incrementally from multiple sources

PHASE 4: Frontend Magic (Hours 14-18)

  • T14: Corkboard canvas

    • Create components/Corkboard.tsx (see DESIGN_HANDOFF.md §6)
    • Cork texture background (repeating image)
    • Mouse drag to pan, scroll to zoom
    • CSS transform on container div
    • Acceptance: Cork texture visible, can pan and zoom smoothly

  • T15: PersonCard component

    • Create components/PersonCard.tsx (see DESIGN_HANDOFF.md §4.1)
    • Paper texture, pushpin, slight rotation
    • Photo thumbnail, name, title, status badge
    • Spawn animation with Framer Motion (see §5.1)
    • Acceptance: Cards appear on board with spring animation

  • T16: Real-time card updates

    • Connect PersonCard to Convex useQuery
    • Cards update live as backend writes new data
    • Typewriter text animation for new data arriving
    • Status badge transitions (pending → researching → complete)
    • Acceptance: Backend mutation → card visibly updates on screen within 1 second

  • T17: ConnectionLine component

    • Create components/ConnectionLine.tsx (see DESIGN_HANDOFF.md §4.2)
    • SVG overlay on corkboard
    • Dashed red line with glow
    • Draw animation (stroke-dashoffset)
    • Midpoint label
    • Acceptance: String draws itself between two connected cards

  • T18: DossierView panel

    • Create components/DossierView.tsx (see DESIGN_HANDOFF.md §4.3)
    • Slide-in from right on card click
    • All dossier fields: work history, education, social, conversation hooks, risk flags
    • Scrollable content area
    • Acceptance: Click card → panel slides in with full dossier data

  • T19: LiveFeed sidebar

    • Create components/LiveFeed.tsx (see DESIGN_HANDOFF.md §4.4)
    • Real-time event stream from Convex
    • New events slide in from top
    • Click event → scroll to related card on board
    • Acceptance: Backend activity shows up as live feed items

PHASE 5: Synthesis + Polish (Hours 18-22)

  • T20: Report synthesizer

    • Create backend/synthesis/synthesizer.py (see SYSTEM_DESIGN.md §4)
    • Aggregate all intel fragments for a person
    • Gemini 2.0 Flash generates: summary, conversation hooks, risk flags
    • Connection detection between people
    • Write final dossier to Convex
    • Acceptance: Person with multiple intel fragments → coherent dossier with hooks

  • T21: TopBar + StatusBar

    • Create components/TopBar.tsx and components/StatusBar.tsx
    • JARVIS logo, agent status indicators, military clock
    • Bottom bar with system status counts
    • Acceptance: UI chrome looks polished and shows live status

  • T22: Laminar observability

    • Install laminar SDK: pip install lmnr
    • Add @observe() decorators to all agent functions
    • Verify traces appear in Laminar dashboard
    • Acceptance: Full agent trace visible in Laminar for one end-to-end run

  • T23: "CLASSIFIED" stamp + polish

    • ClassifiedStamp component with spring animation
    • Appears when dossier status → complete
    • Sound effects (optional): shutter, paper, typewriter
    • Acceptance: Completed dossier gets a dramatic stamp animation

PHASE 5B: Ambient Intelligence + OSINT Depth (Hours 20-22)

  • T28: WiFi probe sniffer

    • Create backend/ambient/wifi_sniffer.py (see SYSTEM_DESIGN.md §9.2)
    • Raspberry Pi 4 + Alfa AWUS036ACH in monitor mode
    • tshark captures probe requests → extract MAC + SSID
    • WiGLE API enriches SSID → GPS location (workplace/school inference)
    • Acceptance: Running at venue, capturing SSIDs from nearby devices

  • T29: BLE device scanner

    • Create backend/ambient/ble_scanner.py (see SYSTEM_DESIGN.md §9.3)
    • Uses bleak library for async BLE scanning
    • Extract device names ("John's AirPods") → owner name parsing
    • Acceptance: Returns list of nearby BLE devices with inferred owner names

  • T30: FEC donation search agent

    • Create backend/agents/public_records/fec_agent.py (see SYSTEM_DESIGN.md §10.4)
    • FEC API (free key) → search by name → political donations since 1977
    • Acceptance: Name search returns donation history with amounts and committees

  • T31: Public records Browser Use agents

    • Create backend/agents/public_records/people_search.py
    • Browser Use agent for FastPeopleSearch / TruePeopleSearch
    • Extract: address, phone, relatives, associates
    • Fire in parallel during Tier 2 enrichment
    • Acceptance: Given a name → returns address + phone + associates

  • T32: OSINT username sweep (Sherlock + Maigret)

    • Create backend/agents/osint/username_sweep.py
    • Run Sherlock (400+ sites, ~30s) and Maigret (2500+ sites, ~60s) in parallel
    • Parse results → list of confirmed accounts with URLs
    • Acceptance: Username → returns 20+ confirmed accounts across platforms

  • T33: Email intelligence (Holehe + HIBP)

    • Create backend/agents/osint/email_intel.py
    • Holehe checks 120+ sites for email registration
    • HIBP checks breach count (ethical: count only, no data)
    • Acceptance: Email → list of registered services + breach count

  • T34: Live camera feed widget

    • Create backend/ambient/cameras/caltrans.py + nyc.py (see SYSTEM_DESIGN.md §11)
    • Pull nearby camera feeds (Caltrans for SF, NYCTMC for NYC)
    • Frontend mini-map component showing live feeds on corkboard
    • Acceptance: Corkboard shows 2-3 live camera thumbnails refreshing every 15s

PHASE 6: Demo Prep (Hours 22-24)

  • T24: Pre-seed demo data

    • Create scripts/seed_demo.py
    • Pre-cache 2-3 complete dossiers with all fields populated
    • Ensure they appear on corkboard with connections on app load
    • Acceptance: App loads with 2-3 fully populated cards + connections

  • T25: Record backup video

    • Screen record a full demo flow (capture → identify → research → dossier)
    • Edit with cuts for slow parts (PimEyes, Browser Use agent wait times)
    • Keep under 3 minutes
    • Acceptance: Video exists and plays smoothly

  • T26: Practice demo script

    • Follow the 3-minute demo script from TECH_DOC.md §6
    • Time each section
    • Identify and fix any dead spots
    • Acceptance: Can deliver full demo in under 3 minutes without dead air

  • T27: Apply to manual prize tracks

    • Submit for: Most Hardcore Infra, Best Design, Best Use of Real-Time Data
    • Check hackathon platform for submission deadlines
    • Acceptance: All 3 manual track applications submitted

Quick Reference: Doc Locations

Doc What's in it
PRD.md Product requirements, judging alignment, features, risks
ARCHITECTURE.md System architecture, data flow, component deep dives, API contracts
TECH_DOC.md Project structure, build plan phases, Convex schema, agent patterns, env vars, demo script
SYSTEM_DESIGN.md Full implementation code for backend/agents/pipelines — copy-paste into files
DESIGN_HANDOFF.md Full visual spec for frontend — colors, typography, layouts, animations, components
CREDITS_CHECKLIST.md Sponsor credit claim checklist