← Back to the story

Projects

Three things I made.

Two are open source and you can use them today. The third runs inside Guidepoint.

v1.0.2MITJS · Python · .NET

Brevit

Cut LLM costs by 60%. Without losing meaning.

Brevit compresses structured data before it reaches your LLM. It flattens JSON to dot-notation, turns uniform arrays into tables and abbreviates repeated keys, so the model reads the same data in 40 to 60% fewer tokens.

JavaScript
npm install brevitnpm ↗
Python
pip install brevitPyPI ↗
C# / .NET
dotnet add package BrevitNuGet ↗
Brevit output
JSON, approx. tokens114
Brevit, approx. tokens0

Runs the brevit npm package in your browser. Token counts are estimated at four characters per token.

VS Code extensionFreeMIT

Git Tree

Every button shows the exact Git command it runs.

A SourceTree-class Git client inside VS Code: a commit graph that stays fast on 50,000-commit repositories, line-level staging, multi-repo workspaces, worktrees, stashes, merge conflicts and pull requests for GitHub, Azure DevOps, GitLab and Bitbucket. No telemetry.

Press a button
git status
  • Initial commit

Illustrative commands. The extension shows its own in the Command Sheet.

Built at GuidepointInternal

Brainpoint

An engineering assistant that cites its sources.

Answers engineering questions over source code, Confluence and an AI-maintained team wiki. A FastAPI backend runs a tool-calling agent over Claude and self-hosted open models, with hybrid retrieval on PostgreSQL and pgvector, Microsoft Entra ID sign-in, and deployment on Azure Kubernetes Service.

How a question is answered
Code, Confluence, wikiRedact and chunkKeyword + vector searchRank fusionAgent reads and checksCited answer
Recall@5, keywords only0.80
Recall@5, hybrid0.87

Measured on 55 labelled questions. Hybrid search fuses keyword ranking with HNSW vector search on pgvector.