Skip to content

PRD.md

Latest commit

 

History

History
121 lines (96 loc) · 6.92 KB

PRD.md

File metadata and controls

121 lines (96 loc) · 6.92 KB

Product Requirements Document: JARVIS

Real-Time Person Intelligence Platform

Version: 0.1 | Date: 2026-02-27 | Hackathon: Web Agents Hackathon (Browser Use + YC) Team: Affaan + Edmund | Track: Research Agent + Identify Leads & Send Targeted Outreach

1. One-Liner

A wearable-powered intelligence agent that identifies people in real-time via Meta glasses, deploys a swarm of browser agents to deep-research them across the internet, and streams results to a cinematic corkboard UI.

2. Problem

At events, conferences, and networking situations, you meet dozens of people but have zero context on who's worth your time. Current solutions (Google someone's name, check LinkedIn manually) are slow, shallow, and socially awkward. You want to know — in real-time — who someone is, what they've done, who they know, and whether they're worth a conversation.

3. Solution

JARVIS uses Meta Ray-Ban glasses to capture photos of people. A facial recognition pipeline identifies them, then a swarm of autonomous browser agents (powered by Browser Use) deep-researches each person across LinkedIn, X/Twitter, Instagram, Facebook, public records, and more. Results stream into a COD-style mission briefing corkboard where each person appears as a dossier with connections drawn between individuals.

4. Target User

  • Founders at networking events who need to identify potential investors, partners, or hires
  • Sales professionals at conferences who need to identify and qualify leads in real-time
  • Anyone at a social/professional event who wants ambient intelligence about the people around them

5. Hackathon Judging Alignment

Criteria Weight Our Play
Impact Potential 40% Real-time person intelligence is a category-defining product. The "wow factor" of walking into a room and knowing everything about everyone is undeniable. Delivers extreme value to founders, salespeople, recruiters.
Creativity 20% COD-style corkboard UI, Meta glasses integration, swarm agent architecture, FBI-board aesthetic with live-streaming intel. Nothing like this exists.
Technical Difficulty 20% Multi-agent swarm orchestration, anti-bot detection bypass for PimEyes, parallel browser agents, real-time streaming pipeline, facial recognition integration, walled-garden data extraction.
Demo & Presentation 20% Live demo: put on glasses, look at a person, watch their dossier materialize on the corkboard in real-time. Backup video with cuts for slow parts.

6. Core Features (MVP — 24 hours)

P0 — Must Have

  1. Photo Capture Pipeline: Meta glasses → photo → send to backend
  2. Facial Recognition: PimEyes integration with parallel accounts for anti-bot
  3. Initial Report Generation: LLM-powered summary from PimEyes results
  4. Browser Agent Swarm: Deploy multiple Browser Use agents to deep-research each person
    • LinkedIn profile scraping
    • X/Twitter activity
    • Instagram (public profiles)
    • Google search aggregation
    • Exa API for structured research
  5. Corkboard UI: COD/FBI-style mission board
    • Papers spawn in as data arrives
    • Click to zoom into individual dossiers
    • Draw-string connections between people
    • Real-time streaming updates
  6. Dossier View: Per-person detail page with all gathered intelligence

P1 — Should Have

  1. Relationship Mapping: Identify connections between people (shared companies, schools, followers)
  2. Notification to Glasses: Stream summary to Telegram/iMessage/web overlay viewable on Meta glasses
  3. Laminar Observability: Tracing and evals to verify accuracy of gathered intel

P2 — Nice to Have

  1. Conversation Audio Capture: Record ambient conversations for additional context
  2. License Plate Recognition: Vehicle identification for additional data points
  3. Walled-Garden Extraction: Reverse-engineer internal APIs of gated platforms

7. Out of Scope (for hackathon)

  • Mobile app
  • User authentication / multi-user
  • Data persistence beyond session
  • GDPR/privacy compliance (it's a hackathon demo)

8. Success Metrics

  • Identify and generate full dossier for 5+ people during live demo
  • End-to-end latency: initial report < 30 seconds, deep research < 3 minutes
  • Accuracy: zero obviously wrong information during demo (use Laminar for verification)
  • Audience reaction: audible "wow" during demo

9. Technical Constraints

  • Time: 20+ hours of build time (Feb 28 1PM → Mar 1 10AM)
  • Must use: Browser Use (primary sponsor)
  • Should use: Sponsor tools for credits/prizes (Exa, Laminar, Vercel, Convex, MongoDB)
  • Starting code: Edmund's existing PimEyes → LLM report pipeline
  • Hardware: Meta Ray-Ban glasses (need to source/buy)

10. Risks & Mitigations

Risk Likelihood Impact Mitigation
PimEyes bot detection blocks us High Critical Multiple accounts, rotating agents, rate limiting
Browser agents too slow for live demo High High Pre-cache some results, video backup with cuts
Wrong information in demo Medium Critical Laminar tracing for verification, manual review
Meta glasses integration issues Medium Medium Fallback to phone camera + manual upload
Sponsor API rate limits Low Medium Graceful degradation, queue management

11. Sponsor Stack Usage Plan

Sponsor Usage Credits
Browser Use Core agent orchestration — all browser automation $100
Exa Structured person/company research API Available via MCP
Laminar Agent tracing, evals, accuracy verification $150
Vercel Deploy corkboard frontend $50
Convex Real-time database for streaming dossier updates Free tier
MongoDB Persistent storage for person data/reports Free cluster
Google DeepMind Gemini for LLM calls (report generation) $20
OpenAI GPT-4o for vision (photo analysis) + research Apply for credits
HUD Agent observability/debugging $200
Daytona Sandboxed environments for browser agents $100
Supermemory Agent memory across research sessions $100
VibeFlow Rapid UI prototyping for corkboard $200

12. Prize Track Strategy

Track Auto-entered? Strategy
Top 3 Overall Yes Go all-in on wow factor + technical depth
Founders Prize Yes Personal story: "We built what we wished we had at YC events"
Most Viral Yes Post demo video on X with #browser-use — the visual is inherently viral
Most Hardcore Infra Apply manually Multi-agent swarm, anti-bot, parallel processing — YES APPLY
Best Design Apply manually COD corkboard is visually stunning — YES APPLY
Best Use of Real-Time Data Apply manually Entire product is real-time streaming — YES APPLY
Best Devtool Apply manually Not really a devtool — SKIP