Projects

AI products, each one tested

Six tools I built with Claude, Cursor, ChatGPT and other AI tools. Each card leads with the number its test measured. The scoreboard below puts them side by side.

The model extracts and suggests; tested code does the counting and makes the call. Where a result is still weak, the repo says so in its EVALS.md.

All projects

01

ClientLens

Sales AI copilot (Chrome extension)

Google Meet copilot: drafts pitches from your knowledge base, coaches live, fact-checks every claim before it reaches a doc.

Tab audio via Deepgram STT into a client-side multi-agent pipeline: next-best sentences, objection handles, sentiment.

  • AI-Built
  • Chrome Extension
  • Multi-Agent RAG
  • Deepgram STT
  • Sales
Live SiteGitHub
TESTED67/72 answers rightfrom its knowledge base with Gemini search, on new questions it was never tuned on (24 test questions x 3 runs, made-up KB, test harness; keyword alone 60/72). Separate test: 13/13 hidden false claims caught, 0 reached a doc.

02

FollowThrough

AI-native meeting intelligence

Every client promise captured and tracked to completion.

Multi-pass LLM extraction, citation-verified.

  • AI-Built
  • TypeScript
  • LLM Extraction
  • Open Source
Live DemoGitHub
TESTED100% rightevery promise it flagged on new calls was real. It found 76% of them.

IN USESales team pilot, October 2026.

03

EOD

End-of-day work receipts

Your day's shipped work as a clean receipt, so updates stop being a memory exercise.

  • AI-Built
  • Python
  • Productivity
Live DemoGitHub
TESTED5/10 → 0/10days where the update claimed more than was done.

IN USE5 people in product and marketing, daily since mid-2026.

04

PM-OS

AI product management operating system

82 PM modules across 10 disciplines, 81 with AI drafting on your own key: PRDs, roadmaps, execution.

TESTED28 → 2made-up figures, before and after the fix.
  • AI-Built
  • TypeScript
  • PM Workflows
  • Live
Live DemoGitHub

05

Kokum

Pantry-aware cooking

Code decides what you can cook; Claude only ranks and explains.

TESTED3/80 → 0/80lines naming an ingredient the dish does not use. Code now puts dishes you can cook first.
  • AI-Built
  • Claude
  • Cooking
Live DemoGitHub

06

Bidline

Procurement on your own Claude, over MCP

Your Claude reads messy supplier quotes through 58 MCP tools; Bidline's code converts, compares, records every edit and gets the award approved. No AI inside the app.

TESTED150/150 pricesmatch a checked reference to the rupee, and an end-to-end test drives the whole flow through MCP.
  • AI-Built
  • Procurement
  • MCP
Live DemoGitHub