A virtual ₹1,00,000 portfolio priced off real NSE and BSE quotes: 4,324 equities, 351 ETFs and 8,800+ active mutual funds (as of 1 Oct 2026). Nine behavioural-bias detectors run on every trade, and an AI coach explains what just happened in the user's own numbers without ever giving a tip. Crash Replay re-runs real Indian market crashes day by day: three hand-built from real Nifty 50 closes, ten featured, and a generator for any other event. Placed 2nd nationally at the Masters' Union AI Buildathon; 170 accounts since, all by word of mouth.

• Vanilla ES modules, no build step
• Python on Vercel (one ASGI entrypoint)
• Supabase Postgres + RLS + RPCs
• Gemini 3 Flash / 3.1 Pro on Vertex AI
• GitHub Actions data + order jobs
• Yahoo + Tickertape + AMFI feeds
• Service Worker PWA
• Kotlin + Jetpack Compose (Android port)

Live: stocksaathi.co.in · Source: github.com/thealiarbab/StockSaathi

Started as my Class XII Computer Science practical — Python, raw parameterised SQL, one operations module shared by a console app and a web app — then rebuilt into a product: a React + TypeScript client over a Flask JSON API on Postgres. Budgets against actuals, a ledger with accounts, transfers and recurring rules, and a portfolio priced live from StockSaathi's public API. Open demo, no sign-up.

• React 19 + TypeScript + Vite
• TanStack Query + React Router
• Flask JSON API
• PostgreSQL (Supabase), raw SQL
• psycopg2 connection pool
• pytest against a real database
• StockSaathi public API for prices

Live: spendincheck.com · Source: github.com/thealiarbab/SpendInCheck

Tap the mic, say “Ram ne paanch sau udhaar liya,” and the ledger entry writes itself — confirmed back in a natural Indian voice, with one-tap Undo. Llama 3.1 8B parses intent in about 200ms, Sarvam Saarika transcribes Indian voices natively, and one Postgres function inserts the row and returns the new balance in a single round trip. Shaped by interviews with local shopkeepers. The web prototype is complete; the app is now being redesigned and rewritten natively in Kotlin.

• Next.js 16 + React 19
• TypeScript strict
• Tailwind CSS v4
• Supabase + anonymous auth + RLS
• Llama 3.1 8B (Groq) primary
• Llama 3.3 70B fallback
• Sarvam Saarika v2 STT
• Sarvam Bulbul v2 TTS
• Whisper-large-v3-turbo fallback
• Kotlin (rewrite in progress)

Source private until the rewrite ships

ESP32 firmware drives two fail-secure magnetic-lock relays; an ESP32-CAM streams MJPEG from the door; a Flutter app ties them together. No vendor cloud and no account to cancel — the lock lives on the home network, and the only way in from outside is a Cloudflare tunnel the owner controls. Maggy, an optional voice assistant with persistent memory, can lock, unlock and set the auto-lock timer on request. Now being rebuilt from the ground up.

• ESP32 (Arduino, ArduinoJson)
• ESP32-CAM (hardware JPEG, MJPEG)
• Flutter (Dart 3, Material 3)
• Provider state
• Cloudflare Tunnel (remote access)
• speech_to_text + flutter_tts
• Grok-3 (Maggy's reasoning)
• Inno Setup Windows installer

Source publishes with the rebuild

Three modules in one closed loop: an LLM reads a decade of filings, Fed statements and news under versioned analyst personas and writes an Alpha Ledger; a Polars backtest engine trades on it with slippage, commissions and partial fills; an Unreal Engine 5 city visualises the result live. The rule that everything else serves is the temporal firewall — no simulated moment may see a row from its own future. The harness is built and tested on synthetic data; full-scale runs wait on hardware.

• Python 3.12 + uv
• Polars (as_of_join only)
• Pydantic + Pandera contracts
• Hypothesis property tests
• ZeroMQ + MessagePack bridge
• Unreal Engine 5 (twin, planned)
• DeepSeek-R1 32B (planned)

Source: github.com/thealiarbab/sovereign-alpha

A home machine behind carrier-grade NAT (so no public IP and no port forwarding) serving my own sites to the internet anyway, through a Cloudflare tunnel it dials out itself — FastAPI on uvicorn, no hosting bill. A Python control panel, packaged as a Windows executable, watches its processes and network. And the Command Nexus design system: one JSON spec of colours, type and motion rules that MagLock's app and my study pages are both built from. It's live at lamecraft.org.

• FastAPI + uvicorn
• Cloudflare Tunnel
• Python control panel (psutil)
• PyInstaller Windows build
• Command Nexus design system (JSON spec)
• Standalone HTML + KaTeX pages

Live: lamecraft.org · Source not public

Turned my Class XII Computer Science practical into a live personal-finance product: a React client over a Flask and Postgres API, with holdings priced live from StockSaathi.

Built and run a live paper-trading simulator for Indian teenagers: real NSE and BSE prices, nine behavioural-bias detectors and an AI coach that never gives tips. 170 accounts so far, all by word of mouth.

Voice-first Hindi and Hinglish ledger for shopkeepers. The web prototype is complete; the app is now being redesigned and rewritten natively in Kotlin.

Two-door ESP32 smart lock with an ESP32-CAM and a Flutter app, reachable from outside only through a tunnel I control. Being rebuilt from the ground up.

Designing and building a fully local, reproducible market simulation to test whether a language model reading history can find real, friction-adjusted alpha.

Built StockSaathi end to end during a 36-hour national buildathon while a teammate pitched it to the panel. It placed second nationally.

Interviewed shopkeepers about how they actually keep their books before designing BolHisaab — credit, cash, customers named in Hindi, and a paper notebook.

Run my own server for my sites and tools behind a Cloudflare tunnel, with a Python control panel and the Command Nexus design system my other projects borrow.

Devansh was already shouting StockSaathi's name everywhere, unasked, at the risk of his discipline-coordinator badge, so I made it official. Every sign-up since has come by word of mouth.

Built a biomimetic bionic hand and hand-exoskeleton prototype as part of a robotics group project; the hand was shown at a science exhibition.

Teaching a younger neighbour, now in Class X: programming, hosting servers, how packets and networks work. After a two-day rescue of his locked-out PC, he conceded I'm a "PC goat".

The brief was set on the day, so StockSaathi was built from scratch in 36 hours; a teammate pitched it to the panel.

Three tie-breakers deep, I had made my peace with third place — then my house was announced first. December 2025.