Skip to content
Abhinav Anand
work

Agentic AI / Hackathon

FairSprint AI

Agent that checks what team members say they did against the evidence that exists, with a mentor approving every consequential change.

Live Code
FairSprint AI admin portal overview: sprint metrics and standup participation for eight team members, running on synthetic demo data.
Admin overview in the demo build, running on synthetic data.

[01] problem

Mentors must judge team progress they mostly cannot see, and commit counts under-count reviewers, testers, writers and designers.

[02] approach

  1. 01A LangGraph agent loads each member's standup claim and collects evidence from Supabase and read-only GitHub activity.
  2. 02Deterministic code assigns one of four verification statuses (verified, partially verified, unverified, insufficient evidence), classifies risk and proposes reassignments with capacity and dependency checks.
  3. 03The run pauses for mentor approval; approving a reassignment applies it once, rejecting changes nothing, and the last decision resumes the run to draft the weekly report.

[03] what's interesting

  • Verification is deterministic code, not the language model, and the model's output is validated against schemas.
  • A missing API response is never treated as proof of no activity, and no output is worded as a finding of misconduct.
  • Gemini is the primary model with Groq as an independent fallback.

[04] stack

  • Python
  • FastAPI
  • LangGraph
  • React 19
  • TypeScript
  • Supabase
  • Gemini
  • Groq