D3 Admissions Briefing Agent

An AI system that autonomously generates research briefs for startup applications at District 3, a publicly funded incubator at Concordia University in Montreal.

Built as part of my application for the Wealthsimple AI Builder role.

System design whiteboard (tldraw)

What it does

When a new startup application is submitted, an agent autonomously produces a 9-section research brief: founder profiles, competitive analysis, SDG alignment, stream classification, scored rubric, risk flags, and interview questions — all cited to real URLs or application fields.

~15-20 minutes and ~$2.50 CAD per brief, replacing 3-4 hours of manual research.

How it works

The agent operates in a sandbox rather than following a scripted pipeline:

• Knowledge — a /knowledge folder with D3's mandate, evaluation rubric, stream definitions, and SDG framework
• Tools — 13 MCP tools for web research, self-assessment, section writing, human flagging, and working memory
• Goal — produce 9 brief sections, all cited, with a recommendation
• Guidelines — citation requirements, confidence thresholds, a mandate to revise earlier work when later findings contradict it

Within this sandbox, the agent decides what to research, in what order, and how deep to go.

Key design decisions

• Humans decide, AI prepares. The agent has no tool to accept, reject, email, or route. The responsibility boundary is architectural.
• Self-correcting. The agent self-assesses after each research phase and revises earlier sections when new evidence contradicts them.
• Human-in-the-loop. Two touchpoints: mid-run questions (with 5-min timeout) and post-brief review where a mini-agent rewrites sections based on reviewer input.
• Research resilience. Fetch cascade: GitHub API, direct HTTP, agent-driven sub-page exploration, Jina Reader (JS rendering), Wayback Machine. If all fail, the agent flags the gap honestly.

Setup

# Clone and install
git clone https://github.com/shaynelarocque/briefbot.git
cd briefbot
pip install -r requirements.txt

# Add your API key
cp .env.example .env
# Edit .env with your Anthropic API key

# Run
uvicorn app.main:app --reload

Open http://localhost:8000 to submit an application and watch the agent work.

Deploy to Render

Or manually:

1. Create a new Web Service on Render and connect this repo
2. Set environment variables:
   • ANTHROPIC_API_KEY — your Anthropic API key
   • AUTH_USERNAME — reviewer username (optional, enables Basic Auth)
   • AUTH_PASSWORD — reviewer password (optional, enables Basic Auth)
3. Render will auto-detect the Dockerfile and deploy

When AUTH_USERNAME and AUTH_PASSWORD are both set, every page is gated behind HTTP Basic Auth — share the credentials with reviewers for private access.

Stack

• Agent: Claude Agent SDK (Claude Sonnet 4.6)
• Backend: FastAPI, SSE streaming
• Frontend: Vanilla JS, HTMX
• Storage: File-based (no database)