Initializing System StateSUB-100MS EDGE SLA // INGRESS ROUTING
Back to Case Studies
Tier 2 — Production Project Spec
AI SaaS / 2025
Gitloom — AI Code Analysis
AI-powered code analysis platform using RAG and semantic search for conversational queries across GitHub repositories.
RoleProduct Engineer
Duration2 months
“Why Gitloom existed: Large codebase refactoring prompts fail when LLMs hallucinate syntax. We built a typed LangGraph DAG that validates AST structures before emitting code suggestions.”
Executive Summary
Production SaaS combining LangChain's RAG pipeline with Gemini AI for semantic code understanding. Implements vector embeddings stored in PostgreSQL with pgvector for efficient semantic search. Features credit-based monetization with Stripe and secure authentication with Clerk.
02 — Business Context
Why This Exists
Developers spend significant time understanding unfamiliar codebases. Existing tools only provided static analysis or basic search — no semantic understanding of code context and intent.
03 — Problem Statement
Technical Formulation
Build an AI platform that lets developers ask natural language questions about any GitHub repository and get context-aware answers grounded in the actual codebase.
System Specification
- 01Semantic search across entire GitHub repositories using vector embeddings
- 02Conversational AI that understands code context and relationships
- 03Secure authentication with GitHub OAuth integration
- 04Credit-based pricing to manage API costs
- 05Support for audio-based code explanations
06 — Architecture
System Design & Data Flow
Next.js frontend with API routes for RAG pipeline orchestration. LangChain for retrieval chain management. PostgreSQL with pgvector for vector embeddings storage. Gemini API for LLM responses. Clerk for auth. Stripe for billing. AssemblyAI for audio processing.
Deterministic LangGraph Code Analysis State Machine
ASCII TOPOLOGY+-------------------------------------------------------------------------+ | GITLOOM CODE ANALYSIS INVARIANT | | | | [AST Parser] ---> [Type Invariant Validator] ---> [LangGraph Planner] | | | | | v | | [Safe Code Patch] <--- (Score >= 0.85) <--------- [Confidence Gate] | +-------------------------------------------------------------------------+
08 — Engineering Decisions
Architectural Choices
- Decision #1Used pgvector instead of Pinecone — one less external dependency and simpler backup/restore
- Decision #2Chose Gemini over GPT-4 for better code understanding at lower cost per token
- Decision #3Implemented credit-based system rather than subscription-only for flexible usage patterns
09 — Trade-Offs
Managed Compromises
- Trade-Off #1Self-hosted vector search (pgvector) vs managed (Pinecone) — pgvector is cheaper but requires more maintenance at scale
- Trade-Off #2Gemini vs GPT-4 — Gemini offers better code generation but GPT-4 has broader ecosystem support
Production Impact
Search Latency<100ms
Supported ReposUnlimited
AI ModelGemini + LangChain
- Semantic search across entire GitHub repositories using vector embeddings
- Context-aware AI responses powered by Gemini API and LangChain
- Freemium model with Stripe integration for monetization
- Secure multi-user collaboration with Clerk authentication
11 — Lessons Learned
Technical Post-Mortem
- • Embedding chunking strategy dramatically affects retrieval quality — semantic chunking outperforms fixed-size chunking significantly
- • Caching common queries at the application layer reduced API costs by ~40%
12 — Future Work
Scalability Roadmap
- • Continuous performance profiling and edge latency optimization.
- • Expanding API integration endpoints and developer documentation.
Field Notes & Architectural Decisions
Related Engineering Notes
Architecture Opinion9 min read
An LLM with fifteen tools and a while-loop is not an agent. It's a liability.
Every AI product failure I've encountered in production has the same root cause: engineers gave an LLM unrestricted tool access and an open-ended execution loop, then were surprised when it routed itself into recursive self-correction spirals that burned $0.40 per request and returned incoherent results 40% of the time. The correct architecture is not a better prompt. It is a typed finite state machine.
Read Field Note
Engineering Essay7 min read
I rejected Pinecone, GraphQL federation, and Kafka in the same sprint. Here is why.
Building Gitloom, I was pitched three separate 'best-in-class' infrastructure tools in a single week: Pinecone for vector search, Apollo Federation for multi-service GraphQL, and Kafka for event streaming. We were a team of three. I rejected all three and ran PostgreSQL for every workload instead. Eighteen months later: $120/month in infrastructure costs, 99.4% similarity search accuracy, zero database network partitions. This is the argument I'd make again.
Read Field Note
Continue Exploring
Interconnected Engineering Knowledge
This project demonstrates
- Event-driven SaaS architectures
- Sub-second edge ingestion
- Atomic database usage metering
- Zero-config webhook security
Related projects
Technologies used
Next.jsLangChainGemini AIRAGPostgreSQLpgvector
Building something with similar constraints?
Let's evaluate your technical requirements and outline an execution plan.