indieNik · Bengaluru, India
Nikhil Patil builds AI systems by day and runs cafés with them.
I build AI products, autonomous systems and the software behind my own café business.
- Projects live
- 13
- Cafés
- 2 ventures
- Primary tools
- Claude Code, Angular, GCP
- Also
- Beatboxer
Selected work
Things that are live and doing a job.
Most of these touch real money, real customers or a real counter, so the interesting parts are the guardrails.
IgniteAI
One product photo in, one finished UGC video ad out.
Alphie
An intraday trading bot for NSE stocks, with a Claude sleeve inside hard risk rails.
DhanChakra
An affiliate income loop that scouts, creates, publishes, measures and heals itself.
Baristalk
Scan a table QR, speak your order in any language, it lands on the counter priced and paid.
PanelPrep
A live mock interview with two AI panelists who have two real voices and interrupt each other.
Hoist
A JavaScript and TypeScript learning PWA built for recall under pressure, not just comprehension.
Oasis POS
The offline-first Android point of sale my café chain runs on every day.
The cafés
Started brewing chai, now building systems.
Every system in the café column exists because something at the counter was slow, wrong or on paper. The café is the test bench and the customer.
How I work
Distilled from the CLAUDE.md files, not written for this page.
Verify live before claiming anything.
A feature is done when it has been exercised in production and the receipt is written down: the revision id, the HTTP status, the object id. Anything not verified is labelled not verified. Status files keep three buckets: claimed and verified, not verified, deliberately incomplete.
“Run every non-obvious code sample before writing the output down.”
Push complexity into deterministic code.
Language models are probabilistic and most business logic is not. Directives in Markdown describe intent, an agent orchestrates, and Python or TypeScript does everything that has to be repeatable. The model writes prose and classifies. Code decides.
“90% accuracy per step is 59% success over five steps. Push complexity into deterministic code.”
Money rules never live in agent judgment.
Budget caps, kill switches, dry-run defaults and paused-first launches are enforced on the server and pinned by tests that fail if loosened. Automated de-risking is allowed. Automated re-risking is not. Turning real risk on is always a human call.
“Gemini writes analyst notes only; decisions stay reproducible.”
Docs ship in the same commit as the code.
Every repo has a CLAUDE.md that an agent reads first, a STATUS file that is read first and updated last, numbered architecture decision records, a changelog, and per-PR notes. A doc update is part of the change, not a follow-up.
“If you make a decision and there is no ADR for it, you are not done.”
Off duty
Not everything is software.
- Beatboxing. Drums, bass and snares with a microphone. Tracks on SoundCloud since 2015, a beatbox Instagram, and the occasional cover. Listen
- Chai. Jaggery chai, specifically. Enough to open a café about it.
- Espresso. Dialing in shots at home. I built an app for it in a day.
- Writing. Frontend performance, automation and analogies on Medium since 2017.
Contact