Selected client work.

Built a full-stack live options trading system running a portfolio of 7 to 10 concurrent intraday strategies on Nifty and BankNifty with ₹1.5 Cr deployed. Three strategy families ran simultaneously: Greeks-based strategies using real-time delta and vanna thresholds for entry; momentum strategies with IV rank and VIX filters; and non-directional multi-bucket straddles and spreads with layered exit logic covering fixed SL, trailing SL, target exits, time exits, and end-of-day square-off.

Execution layer built in Python and Rust connected to three brokers simultaneously through a custom OMS with position reconciliation, pre-trade margin checks, and automatic retry logic. InfluxDB stored full tick data, signal events, and order lifecycle logs. Gradio dashboard provided live PnL per strategy, margin utilisation, drawdown tracking, and risk breach alerts.

Delivered end-to-end: backtesting infrastructure to validate strategies before deployment and a live execution system across three strategy families. Covered call logic handled dynamic strike selection by delta target and IV rank, with rolling triggers on premium decay. Deep OTM strategies used delta thresholds below 0.10 with VIX and IV percentile regime gates. Equity long-short used futures for short-side execution to bypass cash market short-selling constraints.

XTS integration handled in Rust for the critical execution path. Backtesting engine modelled realistic NSE transaction costs including STT on exercise and expiry-roll logic for weekly and monthly contracts. Separate standalone dashboard built specifically for Tech Options with live PnL, margin tracking, and EOD confirmation.

Deployed a modified version of QuantEngine's proprietary MFT strategies directly onto Kyra's on-premise server infrastructure, calibrated to their capital base and prop account structure. The sub-10 ms latency target drove all architectural decisions: Rust handled market data ingestion, order construction, and XTS dispatch on the hot path with zero garbage collection and lock-free data structures between threads. Python managed orchestration and session lifecycle above the Rust core. Cython compiled signal pre-processing and Greeks computation utilities to eliminate interpreter overhead on high-frequency function calls.

All heap allocation on the order dispatch path eliminated via pre-allocated buffers at session initialisation. System benchmarked under simulated live load before production deployment and validated against the latency target on Kyra's actual network path to XTS.

Built a real-time decision-support platform for Alaric's discretionary trading desk. The scanner engine written in Rust evaluated the full US equities universe continuously against four scan types: price action and breakout detection, volume spike identification using relative volume multipliers, gap up and gap down tracking with fill detection, and technical indicator scans covering RSI, MACD, and moving average relationships, all computed incrementally on live tick data.

Unusual Whales integration added options flow intelligence and dark pool print data alongside standard price and volume scans, giving traders institutional activity context in a single view. Delivered as a native desktop application keeping all processing local. Position tracker provided live mark-to-market PnL on manually entered positions. Alert system used the same Rust scanner engine ensuring alert triggers were consistent with dashboard state.

Built a multi-layer statistical pairs trading system scanning 100 NSE futures pairs daily. Pair universe was Hurst-filtered from a broader candidate set — only pairs with Hurst exponent consistently below 0.5 were retained, indicating mean-reverting spread behaviour. Kalman filter was used for hedge ratio estimation rather than static OLS, allowing the ratio to evolve continuously as new price data arrived without periodic recalibration, critical in Indian markets where FII flows and corporate actions shift pair relationships gradually.

Z-score computed on the Kalman-adjusted spread with configurable entry and exit thresholds. Half-life estimated via Ornstein-Uhlenbeck process to characterise mean reversion speed and inform holding period logic. Hurst exponent also served as a live regime gate during scanning, suppressing signals on pairs whose exponent had drifted above 0.5 even at technically valid z-score levels. Scanner output ranked pairs daily by composite signal score for the trading desk.

Built an AI agentic research platform that automated the production of comprehensive equity research reports. Five specialist agents ran in parallel via CrewAI: a fundamentals agent covering financial ratios, earnings trajectory, and balance sheet health; a valuation agent computing multi-method valuation ranges and peer comparisons; a news and sentiment agent sweeping global web sources for recent developments and analyst activity; a technical analysis agent assessing trend, momentum, and support and resistance; and a sector and competitive positioning agent mapping the company's market position and industry dynamics.

All agents used GPT-4 via LangChain tooling with web search for global data retrieval. Output parsers enforced structured JSON from each agent before passing to a synthesis agent that assembled the parallel outputs into a coherent report with executive summary, bull and bear case, and key risks. Final reports generated as PDF and Word documents.

Built a gold inventory risk management and hedging platform for a physical gold consultancy carrying continuous price exposure across its inventory. Three live price sources were consolidated into a single risk view: RSBL domestic spot for inventory valuation, international gold spot via Zerodha Kite as the leading price signal, and MCX futures prices for hedge mark-to-market. A manual override channel handled feed outages and off-market transactions.

Risk measurement layer tracked per-client inventory in grams against live RSBL prices and separately tracked open transactions with deferred settlement as counterparty exposure, computing mark-to-market on each as prices moved between deal date and settlement. Hedge engine converted total unhedged exposure to MCX Gold Futures lots and exposed FastAPI endpoints for hedge placement, modification, and close. Internal web dashboard gave operations staff a live view of inventory value, total exposure, hedge adequacy ratio, and counterparty risk register.