I build AI systems that run unattended in production — and that show their own failure rates.
Bengaluru, India · self-taught · B.Com (Data Analytics, 2025)
Most AI demos work once, on stage. I'm more interested in the boring part: what happens on day 90, at 3am, when nobody is watching and the input is malformed. Every project below has some version of the same idea baked in — a system that tells you when it is wrong.
🔎 glassbox-pro — grounded crime intelligence
Built for the Karnataka State Police Datathon 2026 (Challenge 1), entirely on Zoho Catalyst.
Ask a question in English or Kannada, typed or spoken, and get an answer computed only from case records — never generated. Every claim cites the FIR numbers it came from, and each citation is re-verified against the datastore before it renders. If the data isn't there, it says "no records found" instead of inventing one.
- 6 police roles with jurisdiction scoping enforced server-side, not hidden in the browser
- Explainable 0–100 offender risk scores — every point itemised with its supporting case
- Money-trail engine tracing mule → collector → cash-out using real AML typologies
- A forecast that grades itself: hides its own most recent 30 days, re-runs on older data, and publishes precision — hits and misses
- Aligned to the official CCTNS FIR schema (18-digit CrimeNo, BNS 2023 sections, chargesheet A/B/C)
Node.js Zoho Catalyst Serverless RAG Kannada NLP
📉 trading-algoo — a backtester built to stop me lying to myself
The engine's #1 job is not finding winning strategies. It's refusing to show me a fake one.
- Next-bar execution — a decision on bar N can only fill at bar N+1's open, which makes lookahead bias structurally impossible rather than merely discouraged
- An integrity pass before any strategy runs — gaps, duplicate bars, timezone mistakes, impossible candles (high below low), suspected un-adjusted stock splits, missing values. Bad data in, the backtest never starts
- Real Indian costs charged on every trade: brokerage, STT, exchange fees, SEBI fee, stamp duty, GST, slippage
- Walk-forward testing, Monte Carlo resampling, parameter-sensitivity heatmaps
- A multiple-testing penalty that rises with every variant tried — and a permanent count of every variant ever tested, because forgetting your failures inflates everything after them
- Scorecards attach warnings when the evidence is too thin to believe: too few trades, suspiciously high Sharpe, a drawdown you'd have panicked out of
- 29 test files, in the repo from commit 1
Python pandas NumPy pytest Angel One SmartAPI
🎬 bloom-cafe — a cinematic site that degrades honestly
A scroll-animated café site: the page opens as a solid paper field with the name cut out of it, and scrolling flies the camera through the letter into the room. Plain HTML/CSS/JS + GSAP, no build step.
- If the animation CDN fails to load, the page renders a clean static version instead of a broken one
- Respects
prefers-reduced-motion— the whole zoom-through is swapped for the settled layout - Font-loading gate so text never flashes unstyled
- Live demo
HTML CSS JavaScript GSAP ScrollTrigger
Unattended-first. My content pipelines have run daily on schedulers for 5+ months with no manual intervention. That means atomic writes so a mid-run crash never corrupts output, single-instance locks so runs can't collide, never-reuse ID allocation, and quality gates that demand positive proof the output is good rather than just checking the job didn't crash.
Honest measurement over impressive numbers. I'd rather ship a system that reports 61% precision than one that claims 95% and can't show its working.
Python · Node.js · n8n · MCP servers · RAG · Ollama / local LLMs · FastAPI · Next.js · Playwright · FFmpeg · Supabase · Zoho Catalyst · OpenCV
Open to freelance work in AI automation, agent systems, and data pipelines.