Five-layer quant research pipeline built incrementally from raw tick data to a market-making backtest. Connects to Bybit's L2 WebSocket feed, validates sequence numbers on every delta, and buffers snapshots to Parquet partitioned by instrument and UTC date.
Feature layer extracts bid-ask imbalance, queue pressure, and signed trade flow at each snapshot. A direction predictor trained on rolling windows feeds an Avellaneda-Stoikov-style backtest with configurable spread and inventory limits.
European option pricing under stochastic volatility via the characteristic function. Implements Heston-FFT and the Lewis (2001) contour-integral method. Calibrates (κ, θ, σ, ρ, v₀) against implied vol surfaces using differential evolution, fixing κ and θ to reduce dimensionality.
Systematic backtest of options strategies on NSE via DhanHQ. Fetches historical chain data, simulates entries and exits with real strike selection and expiry logic, and generates per-year P&L attribution from 2021 through 2026.
Rule-based strategy selector on NSE weekly derivatives. Classifies market regime from prior-expiry data, then selects from bull spreads, straddles, strangles, butterfly, condor, box spread, and synthetics. Black-Scholes pricing for out-of-sample entry.
Rolling mean-variance optimizer. A neural network predicts forward returns on a 504-day window; Ledoit-Wolf shrinkage estimates Σ̂. MVO solves for weights with risk-aversion parameter λ and 40% position cap. 21-day rebalance with proportional transaction costs.
Ingests Spotify plays, GitHub commits, weather, movies, gym sessions, dreams, calendar, F1 results, and more into a SQLite database. Surfaces all of it in a bento-grid dashboard — ten-plus pages, each one a different slice of the same life.
NoteRAG: a local vector search layer over the full log history. Ask it anything about your own data. Answers privately, never leaves the machine.
Tracks bets on college sports. Log a bet, see the outcome, know where you actually stand.
Drop in a WhatsApp export. Get back a report: who responds faster, vocabulary divergence, how sentiment shifts over months.
Ctrl+Space, from anywhere on the desktop. A popup with a tool registry — AI that can reach into apps, files, and clipboard. No browser tab required.
Business management system for a steel trading company. Role-based access (sales, manager, admin), user approval workflow, quotations, product catalogue. Running in production.
Real-time sign language recognition from webcam feed. MediaPipe hand tracking, trained classifier, live interpretation overlay.
Flask app around a fraud classifier. Enter claim details, get a risk level: low, medium, or high.
Most of what I build starts with a question about how markets work. The Heston implementation came from wanting to understand stochastic volatility at the level of the math, not just the formula. The orderbook pipeline came from wanting real tick data to work with. The derivatives backtester came from wanting to know what actually holds out-of-sample on NSE weekly expiries.
The infrastructure work — the self-hosted dashboard, the schedulers, the data pipelines — is the same precision in a different domain. I'm interested in problems where model assumptions matter: pricing under incomplete markets, microstructure effects on execution quality, what's actually in the risk premium. Based in India.