PRISM is a human-in-the-loop AI system that helps engineering teams understand pull requests, track repository health over time, and make safer decisions during fast development.
It does not auto-merge code, spam PR comments, or blindly enforce rules. PRISM analyzes, explains, and advises — humans stay in control.
Modern teams move fast. Code review tools haven’t kept up.
Most existing tools: • Dump lint errors • Focus on a single PR in isolation • Enforce rules without context • Answer only: “Is this PR okay?”
They fail to answer: • How is this repo evolving over time? • Is our development getting riskier? • Why does this PR matter in the bigger picture?
PRISM was built to fill that gap.
PRISM provides repo-level supervision, not just PR checks.
🔍 Pull Request Analysis
For every PR, PRISM analyzes: • Change size and surface area • Files and directories touched • Risk-sensitive domains (auth, infra, payments, etc.) • Semantic intent using LLMs
It produces: • A plain-English summary • Key risks (if any) • Actionable suggestions • A quantified health delta
📊 Repository Health Scoring
Each repository maintains a rolling health score based on: • Baseline risk heuristics • Semantic risk from PR intent • Directional changes over time
Health states: • Healthy • At Risk • Critical
📈 Visual Health Trends
The dashboard shows: • Current health score • Repo-specific risk reasons • Recent PR activity • Health trends over time (demo visualization)
🧠 Human-in-the-Loop Design
PRISM: • Never auto-merges • Never blocks developers • Never enforces rules blindly
It advises. Humans decide.
Backend • Python + FastAPI – API layer • Heuristic engine – Baseline risk scoring • Gemini API – Semantic PR understanding • MongoDB – PR history & repo health storage
Backend responsibilities: • Analyze PR payloads • Generate risk signals • Compute health deltas • Store and retrieve historical context
Frontend • Next.js (App Router) • TypeScript • Tailwind CSS
Frontend responsibilities: • Repo health dashboard • Repo detail views • PR summaries and insights • Demo-friendly visualizations
Architecture Philosophy • Clear backend / frontend contract • Deterministic mocks for demos • Human-readable outputs • Scalable to GitHub Actions integration
• Designing a health score that is intuitive, directional, and explainable
• Avoiding noisy or repetitive AI output
• Mapping semantic risk into something engineers trust
• Frontend–backend integration under hackathon time pressure
• Git branch chaos (character-building experience)
• Built a repo-level supervision model, not just a PR checker
• Combined heuristics + LLM reasoning coherently
• Created an opinionated but non-blocking developer experience
• Delivered a clean, demo-ready UI
• Kept humans in control at every step
• AI is most effective when it augments judgment, not replaces it
• Context over time matters more than single-PR correctness
• Explainability builds trust faster than automation
• Clean contracts between systems save lives (and hackathons)
Planned extensions: • GitHub App + GitHub Actions integration • Long-term health trend analytics • Team-level risk dashboards • Configurable risk sensitivity per repo • PRISM comments as suggestions, not commands
PRISM aims to become a copilot for code review decisions, not a gatekeeper.
• Python
• FastAPI
• MongoDB
• Google Gemini API
• Next.js
• TypeScript
• Tailwind CSS
⸻
PRISM is an experiment in responsible, human-centered AI for software engineering.