Case study · Founder project
Variorum
The memory your codebase never had.
An AI-powered engineering memory layer for GitHub repositories. It preserves the decisions, the history, and the reasoning around your code, so context does not disappear when engineers move on.
- My role
- Founder and builder.
- Built for
- Teams and developers maintaining codebases that outlive their authors.
The problem
Code is version-controlled. The reasoning behind it is not.
Why a decision was made, why a workaround exists, which docs are stale, and what is risky to touch all live in people’s heads, scattered commits, and old pull requests. That knowledge erodes as teams grow and people leave. New engineers lose days rebuilding context that used to exist.
Why I built it
I built Variorum because I kept hitting this problem myself. Joining an unfamiliar codebase, the code and the git history were always there, but the why was gone. Variorum is my attempt to make that reasoning durable and queryable.
It is built for teams and individual developers who maintain codebases that outlive the people who wrote them. Onboarding, code review, and safe refactoring are the moments it targets.
What it does
Install the GitHub App on a repository and Variorum indexes the code, docs, and history. From there you can ask how something works and get a cited answer, plan a change and see what it will affect, and get automatic insight on every pull request, with a weekly health digest in Slack.
- Cited answers from your real codebase
- Ask in plain English and get an answer grounded in the actual code, docs, and pull requests, with citations that link to the exact lines and PRs.
- Change-risk analysis before you code
- Plan a change and see the files it touches, how risky each one is, who owns it, and which tests are missing.
- Docs that fix themselves
- When documentation drifts from the code, Variorum detects it and opens a fix pull request with the evidence attached.
- Insight where you already work
- Risk-scored briefings on every pull request, untested scenarios surfaced, and a weekly health digest in Slack.
Portfolio and expertise views round it out: knowledge health across repositories, and who actually knows each area.
The rule it follows
Variorum proposes, humans decide.
Coding assistants help you write the next line. Variorum understands the code that already exists and keeps that knowledge alive. It does not generate your features; it builds the cited memory around them and surfaces it when you need it. It never auto-merges and never force-pushes.
How it is built
Frontend
Next.js (App Router), React, TypeScript strict, Tailwind CSS
For a fast, typed, product-grade dashboard.
Backend
FastAPI, Python, Pydantic, SQLAlchemy, Alembic
For async IO-heavy AI and GitHub workloads with typed models and real migrations.
Data
PostgreSQL, full-text search, optional pgvector
For hybrid retrieval today, semantic search at scale when it is needed.
Code understanding
tree-sitter
For a real structural map of files, functions, and classes instead of fragile regex.
AI architecture
Provider-agnostic layer over Gemini, DeepSeek, and Perplexity
For automatic fallback, so one outage or quota limit never breaks the product.
Integration
GitHub App
For least-privilege, per-installation tokens and verified webhooks.
Retrieval is a hybrid of keyword and semantic search over embeddings, blending code, docs, and engineering history into one cited answer. The emphasis is on grounding and reducing hallucination, not just calling an LLM.
The hard part was the AI engineering: turning a whole repository into something a model can reason about reliably, and citing every claim back to a source.