LARPbot

Always-on AI agent that verifies developer claims against their actual GitHub code.

A recruiter forwards an applicant's email (with a github.com/<username> link) to tungtungrecruiting@gmail.com. Within ~60 seconds, a LARP Report lands in the same thread — scored 0–100, with receipts pulled from real commits and source files.

How it works

applicant emails the recruiting inbox
        │
        ▼
Gmail publishes a Pub/Sub event
        │
        ▼
Vercel webhook (/api/gmail-webhook) pings Tensorlake
        │
        ▼
Tensorlake agent (poll_recruiting_inbox)
   ├─ fetches the full email thread (preserves context across replies)
   ├─ extracts GitHub URL + claims (GPT-4o)
   ├─ runs an agentic verification loop over the candidate's repos
   │     using tools: github_commits, github_tree, github_file, nia_search
   ├─ synthesizes a 0–100 LARP score + subscores + receipts
   └─ replies in-thread with the report
        │
        ▼
recruiter sees the report under the original email

The whole pipeline is event-driven — no human triggers anything, no polling. Cron is a fallback only.

What's in a report

• LARP Score (0–100) and one-sentence verdict
• Subscores: skill inflation, project substance, role authenticity, code depth
• Receipts: per-claim links to specific commits or files in their actual repos
• Recent Activity: 5 most recent commits per analyzed repo
• What to Ask Next: targeted interview questions for unverified claims

If the email doesn't contain a GitHub URL (or points to a non-existent profile), the applicant gets a friendly plain-text reply asking for one. The recruiter inbox stays clean.

Tech stack

Layer	What it does	Where
Next.js (Vercel)	On-demand UI + Gmail Pub/Sub webhook	app/, lib/
Tensorlake	Always-on agent runtime — runs the verification pipeline	agent/larpbot_agent.py
OpenAI GPT-4o	Claim extraction + agentic verification loop	tool-use API
Gmail API	Inbox polling, push notifications, sending replies	inlined in agent
Cloud Pub/Sub	Event delivery from Gmail to webhook	GCP
GitHub REST	Source-of-truth for repos, commits, file contents	lib/github.ts, agent
SQLite	Durable per-candidate memory inside Tensorlake MicroVM	/tmp/larpbot_memory.db

Getting started

Prerequisites

• A Gmail account dedicated to the recruiting inbox
• A Google Cloud project with Gmail API + Cloud Pub/Sub enabled
• Tensorlake account, OpenAI API key, GitHub PAT
• Vercel project for the Next.js app

Quick deploy

1. Set Tensorlake secrets:

   GITHUB_PAT, OPENAI_API_KEY, NIA_API_KEY,
   GMAIL_CLIENT_ID, GMAIL_CLIENT_SECRET, GMAIL_REFRESH_TOKEN
   

2. Deploy the agent — see CLAUDE.md for the full uv command.
3. Deploy the Next.js app to Vercel; set TENSORLAKE_API_KEY in env.
4. Wire up Gmail push: create a Pub/Sub topic, grant gmail-api-push@system.gserviceaccount.com Publisher role, create a push subscription pointed at https://<your-vercel-domain>/api/gmail-webhook.
5. Register the watch weekly (Gmail watches expire after 7 days):

   GMAIL_WATCH_TOPIC=projects/<gcp-project>/topics/gmail-watch \
     ... python agent/setup_gmail_watch.py

Send a test email with a GitHub URL — the report should arrive in under a minute.

Project layout

agent/
  larpbot_agent.py        Single-file Tensorlake deployment.
                          Gmail / GitHub / OpenAI / SQLite logic all inlined.
  setup_gmail_watch.py    Registers Gmail push subscription. Run weekly.
  setup_cron.py           (Legacy) Tensorlake cron — broken upstream, unused.
  gmail_oauth_setup.py    One-time refresh-token bootstrapper.

app/
  api/analyze/            On-demand UI: kicks off a job
  api/status/[jobId]/     Poll endpoint for the UI
  api/gmail-webhook/      Pub/Sub push receiver (the real trigger)
  api/candidates/         Lists past candidates from Tensorlake memory

lib/
  openai.ts               GPT-4o agentic loop (mirrors the agent's loop)
  github.ts               GitHub REST client
  nia.ts                  Nia client (currently DNS-fails; non-fatal)
  types.ts                Canonical AnalysisResult shape
  jobStore.ts             In-memory UI job state

.github/workflows/poll.yml   Backup 5-min cron trigger.
public/demo.json             Offline fallback for the UI demo button.

Known limitations

• Tensorlake's built-in cron fails with internal_error in milliseconds before any code runs — confirmed bug for this app. The Gmail Pub/Sub trigger is the supported path.
• Nia API (api.nia.ai) is unreachable from the Tensorlake runtime due to DNS failures. The agent falls back to GitHub-only verification tools and produces reports without it.
• GitHub Actions schedules are best-effort; the 5-minute backup cron may drift to 10–15 minutes during runner contention. It's a fallback, not the primary.
• Gmail watch expires every 7 days. Production should add a weekly cron to re-run setup_gmail_watch.py.

License

Hackathon project. Use at your own risk.