LLM Trading Systems
Deployed decision pipelines across FX, commodities, and options using DeepSeek-V3.1, MiniMax-M2, gpt-oss-120b, and a DPO-QLoRA fine-tuned Llama-3.1-8B served on vLLM.
DSPy/GEPADPO-QLoRAvLLMOANDA
About
From research idea to production system
I build and operate live ML- and LLM-driven systems end-to-end—from multimodal retrieval and tool-using agents to quantitative trading platforms running against real brokers. My work spans model fine-tuning and serving, evaluation, APIs, cloud infrastructure, and the production reliability required to keep automated systems running unattended.
PhD in Engineering Mechanics from UT Austin; author of 10 peer-reviewed publications with 400+ citations.
Selected Systems
Production AI across models, markets, and infrastructure
ML Trading Systems
Built Bayesian volatility-arbitrage signals, a live HMM/NSGA-II equity portfolio, and a Chronos-T5 LoRA plus TabPFN interday options platform.
PyTorchChronosTabPFNNSGA-II
Agents & Multimodal Retrieval
Built an MCP tool-calling signal-mining agent and a full-stack PDF research assistant with CLIP page embeddings, Qdrant retrieval, and a two-pass Qwen vision-model RAG workflow.
MCPCLIPQdrantFastAPI
Production Infrastructure
Shipped containerized, scheduled systems with persistent state, cached model artifacts, broker integrations, failure recovery, and automated deployment.
AWS LambdaModalDockerGitHub Actions
Experience
Building systems that leave the notebook
Zebra Capital Management LLC · Stamford, CT
1.5+
Sharpe, agent-mined strategies
5%
Weekly returns on capital at risk
18
Live options underlyings
LLM-backed trading systems
- Deployed an intraday FX pipeline that synthesized RSS news and price action into structured market narratives and generated trade decisions with Meta-Llama-3.1-8B-Instruct, executed through OANDA.
- Fine-tuned Meta-Llama-3.1-8B-Instruct with DPO-QLoRA—rank-16 BF16 adapters over a frozen 4-bit NF4 base—promoting the adapter only after held-out base-versus-adapter evaluation, and served it through vLLM on on-demand Modal L4 GPUs.
- Deployed a commodity workflow combining news and narrative on price action with sovereign yield-curve differentials, fed to DeepSeek-V3.1 and MiniMax-M2 for mixture-of-experts trade signal generation.
- Developed an intraday options strategy that calibrates a stochastic-volatility model to live option chains and supplies volatility, jump-risk, relative-value, and premium-repricing trajectories to gpt-oss-120b for structured entry and exit decisions.
- Implemented evaluation-gated prompt optimization (DSPy/GEPA) with gpt-oss-120b as the reflection model, promoting a new prompt only when it outscored the incumbent.
ML-backed trading systems
- Deployed intraday options strategies using Bayesian anomaly-detection signals derived from stochastic-volatility calibrations, generating weekly returns of 5% on capital at risk.
- Built and operated a live equity portfolio using Hidden Markov Models for regime detection, Jensen–Shannon divergence for rebalancing, ElasticNetCV for forecasting, and NSGA-II bi-objective optimization—producing 5% annualized alpha versus the benchmark in a multi-year backtest.
- Built an interday options backtesting engine generating trade-level P&L outcomes across 18 single-stock and ETF underlyings.
- Fine-tuned Amazon Chronos-T5 with LoRA on backtest-derived labels, generated rolling time-series embeddings, trained per-asset TabPFN classifiers on the adapted representations, and deployed them for live inference translating predicted profitability into automated entries.
Agents and infrastructure
- Built an MCP-compatible, tool-calling agent for mining interday options entry signals, identifying candidate strategies with Sharpe ratios above 1.5.
- Built a multimodal PDF research assistant with FastAPI and Next.js: CLIP page embeddings indexed in Qdrant, plus a two-pass Qwen vision-model RAG workflow for grounded, page-cited answering and response self-evaluation.
- Designed and deployed containerized serverless systems on AWS (Lambda, ECR, S3, DynamoDB, EventBridge, CloudFormation) and Modal, integrating Groq and Fireworks AI inference with HMM, Bayesian/Heston, Chronos, TabPFN, Qdrant, DSPy, and broker APIs.
- Engineered DST-aware market scheduling, persistent workflow and position state, model-artifact caching, API services, monitoring dashboards, execution recovery, and fail-closed error handling across multi-asset trading sessions.
VISIE Inc. · Austin, TX
I joined during the early integration phase of a surgical navigation platform combining imaging and robotic actuation. My work centered on implementing TCP/UDP communication protocols for robotic arm motion control, supporting end-to-end product packaging and deployment, and helping stage a live demonstration that anchored the company's successful $8.2M Series A.
Sophelio · Austin, TX
I adapted physics-informed modeling originally developed for fusion experiment data to financial time series. The resulting production pipeline used sparse regression for PDE construction and signal generation, paired with a CAGR-maximizing Bayesian TPE optimizer driving a swing-trading system deployed on AWS Lambda.
Los Alamos National Lab & University of Southern California
I generated hydrocarbon production estimates for a real shale site by replacing 3D discrete fracture network geometry with a graph-based reduced-order model of flow and transport. I also built a computational framework for fast estimation of fault stability and ground deformation for field-scale CO2 storage in faulted oilfields.
Education
Foundations
PhD, Engineering Mechanics
University of Texas at Austin
Austin, TX. Advanced training in numerical simulation, scientific computing, and mathematical modeling that continues to inform my work in ML and quantitative systems.
MEng, Mechanical Engineering
Indian Institute of Science
Bangalore, India.
BEng, Mechanical Engineering
University of Mumbai
Mumbai, India.
Life
Beyond the code
