Sameer Himati

I build AI products for real problems.

Right now I'm building Bean — an assistant that drafts your customer-support replies and flags the ones it isn't sure about instead of guessing. Join the beta →

I run Itamih, forward-deployed AI for small businesses since 2023 — most recently a from-scratch practice-management system that runs daily at a dental hospital. Before that I built Fend (an invite-only network for the startup ecosystem) solo, ground up.

GitHubLinkedInXEmail

Projects

Bean

Drafts customer-support replies in the owner's voice, grounded only in facts it can source — and refuses to draft when it isn't sure, instead of guessing. Deployed, receiving live mail.

PythonClaude APIAgentsEvals

Clinic

From-scratch practice-management system for a dental hospital. In daily use by the doctors and staff. Advisory-lock money serialization, invoice-sequence race fixes, a corporate-claim engine backtested against real bills.

Next.jsPrismaBunProduction

sourcery

A retrieval-eval harness for web-search APIs — same query, same judge, swap only the provider. Finding: Firecrawl extracts 3× more sources than Bright Data and answers worse. More context, worse output.

TypeScriptEvalsRetrieval

LibStack

Offline-capable reading viewer for a knowledge vault. IndexedDB write queue with retry classification, and a Cloudflare Worker that syncs highlights back to GitHub with optimistic concurrency.

TypeScriptCloudflare WorkersOffline-first

Other

claude-code-kitMy Claude Code setup — skills, agents, hooks, rulesnanogpt-mlxBuilding a GPT from scratch by hand in MLX. In progressAdMuteChrome extension that mutes non-blockable ads on streaming platforms

Writing

View all

Outside the Agentic Loop

Everyone in the valley reduces an AI agent to a while loop. The loop is the easy five percent. This is a plain map of the other ninety five, the part that actually makes an agent good.

All (Machine) Learning is just a measure of surprise

Hamming's back-of-envelope math made me realize my own math fundamentals weren't strong enough, so I rebuilt it from the floor (exponentials, logs, surprise) until I got to cross-entropy loss, finally explaining itself from the bottom up.

A Month Off Code

Reflections from a month in San Francisco where I took a deliberate break from coding agents, two hackathons, and what stepping back taught me about what I want to build.