A multi-agent system that evaluates developer GitHub profiles the way a senior engineer would — in 60 seconds.
Every recruiter, hiring manager, and engineer looking at a GitHub profile asks the same question:
Is this real work, or tutorial follows wearing a portfolio costume?
Surface metrics lie. Stars, streaks, and pinned repos can be gamed. A senior engineer takes two hours to give a real verdict on a profile — reading commits, scanning READMEs, checking depth versus breadth, sniffing out copy-paste tutorials, cross-checking claims against code.
tier-zero does that review in 60 seconds.
Point it at any GitHub username. Five specialized agents analyze the profile across independent dimensions, then debate the verdict.
| Agent | What it does |
|---|---|
| Forensics | Originality detection — fork ratio, commit pattern analysis, signs of streak farming, tutorial clusters |
| Depth | Technical depth signals — language breadth vs depth, repo quality, presence of tests/CI/docs/deploys |
| Claims | Cross-checks the profile bio against actual code. If bio says "RAG / Agents / Fine-tuning" — does the code show it, or just LangChain hello-worlds? |
| Senior Reviewer | Synthesizes a verdict the way a Staff Engineer would |
| Critic | Pushes back on the verdict. Forces sharper takes. Prevents shallow praise. |
The agents run in a graph, not a chain. They debate until verdict converges.
A structured report containing:
- Verdict — overall hire-ability signal (0-100) with one-line summary
- Three strengths — what this profile is doing right
- Three concerns — what would worry a senior engineer
- Originality forensics — fork ratio, tutorial-cluster detection, streak-farming indicators
- Depth signals — language portfolio, repo quality distribution, production-readiness markers
- Claims vs evidence — bio claims mapped against actual repository evidence
- What to fix — actionable improvements ranked by impact
┌───────────────────────────┐
│ Orchestrator (LangGraph) │
└────────────┬──────────────┘
│
┌────────────────────┼────────────────────┐
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│Forensics│ │ Depth │ │ Claims │
│ Agent │ │ Agent │ │ Agent │
└────┬────┘ └────┬────┘ └────┬────┘
│ │ │
└────────────────────┼────────────────────┘
▼
┌──────────────────┐
│ Senior Reviewer │
└────────┬─────────┘
│
▼
┌──────────────────┐
│ Critic │
└────────┬─────────┘
│
▼
Final Report
See architecture.md for the full design.
- Agent orchestration — LangGraph
- LLMs — Llama 3.3 70B via Groq (primary), OpenRouter (fallback)
- GitHub data — PyGithub (REST API)
- Backend — FastAPI (async background tasks, in-memory store)
- Frontend — Next.js + Tailwind + shadcn/ui
- Observability — Langfuse
- Eval — custom ground-truth set of 4 profiles, scored against senior engineer labels
# Clone
git clone git@github.com:KrishanKVerma/tier-zero.git
cd tier-zero
# Configure
cp .env.example .env
# Add GROQ_API_KEY (and optionally OPENROUTER_API_KEY for fallback)
# Install
pip install -e .
# Run the API server
uvicorn apps.api.main:app --reload --port 8000
# Or run the pipeline directly on a username
python -c "from apps.api.graph import run; print(run('some-github-user'))"- Architecture defined
- Repo scaffold
- GitHub data fetching
- Forensics Agent
- Depth Agent
- Claims Agent
- Senior Reviewer + Critic
- LangGraph orchestration
- Eval set + benchmarks
- FastAPI backend
- Next.js frontend
- Safety + legal hardening
- Deploy to internet
- Public launch
Most AI tools never measure if they work. tier-zero ships with a 4-profile ground-truth set hand-labeled against senior engineer judgment. Every PR runs against the eval. Performance is tracked in evals/results/.
Issues and PRs welcome. See docs/ for design rationale before opening a PR.
MIT — see LICENSE.
Built by Krishan Kumar Verma — open to remote roles and freelance.