Manas Dalvi

All projects

01AI · Infra

PRism

Dependency-aware AI code review that posts context-rich feedback directly on GitHub pull requests.

View source on GitHub

Stack

Tree-sitter · Gemini · Qdrant · Redis · Celery · AWS ECS · Docker

In one line

AST-scoped PR review

Overview

PRism is an event-driven developer tool that reviews GitHub pull requests with real codebase context instead of just the raw diff. When a PR event fires, it walks the repo's dependency graph to find every file and symbol impacted by the change, then assembles a tightly scoped prompt so the model reviews the change the way a senior engineer familiar with the codebase would.

Architecture

Demo

Demo coming soon

Highlights5

01

AST parsing with Tree-sitter builds repo-wide dependency graphs, identifying impacted files and call sites per PR change.

02

Context-scoped prompting assembles diffs, impacted call sites, and internal engineering rules into tightly scoped Gemini prompts.

03

Scoped RAG via Qdrant enforces internal style guides, with per-commit-SHA caching of ASTs and symbol maps in Redis.

04

Deployed on AWS ECS with Celery and isolated Docker workers, webhook signature verification, and short-lived GitHub App tokens.

05

Precision/recall evaluation pipeline runs against known-bad PR test sets for continuous quality measurement.

Outcomes

Repo-wide
AST dependency graph
Per-SHA
cached ASTs + symbol maps
Scoped RAG
style-guide enforcement

Related notes

NextSASD