I'm a computational engineer who designs, ships, and operates full-stack ML and LLM-based systems end-to-end. A PhD at the University of Texas at Austin grounded me in feature engineering, optimization, stochastic modeling, and scientific computing — the toolkit I now turn on quantitative finance, LLM pipelines, and multimodal document intelligence.
In practice, that means systematic trading systems running in production — some driven by classical ML and stochastic models, others steered by LLM inference shaped to how the underlying model actually thinks: chain-of-thought scaffolding for small non-reasoning models, decision policy and abstention rules for reasoning models. Markets grade the prompts, so the iteration loop is honest. Alongside that sit multimodal LLM agents and the serverless cloud scaffolding around them. I'm comfortable owning research, infrastructure, and production code as a single workflow — feature engineering, model selection, calibration, prompt optimization, deployment, evaluation, and the unglamorous infrastructure that keeps any of it from breaking at 3 a.m.
LLM-backed systems. Built three intraday LLM-driven trading systems across asset classes, each with its prompt architecture shaped to the model class doing the reasoning. An options strategy where an LLM reasons over live stochastic-volatility calibrations every few minutes: state-dependent prompts for entry versus position management, verdict history against flip-flopping, and the entry thesis read back verbatim until exit — the prompt is the strategy, versioned and evaluated over fixed windows. An FX system with a staged pipeline that synthesizes news and price action into market themes and a volatility narrative, then conditions per-pair long/short/no-trade signals through explicit chain-of-thought scaffolding sized to a small non-reasoning model — with the loop closed by a DSPy (GEPA/MIPRO) optimization workflow that relabels archived live runs against executable market prices as verifiable rewards, guarded by temporal splits and a paired-bootstrap promotion gate. And a commodity ticker system distilling macro news and multi-country sovereign yield curves into a daily directional call via a tiered evidence hierarchy with conflict-resolution and abstention rules, plus a required disconfirming-evidence field stating what would flip the call.
ML-backed systems. Built intraday options strategies on Bayesian anomaly-detection signals, with stochastic-volatility calibration via constrained optimization and FFT options pricing, generating 5–10% weekly returns on peak capital at risk. Deployed a live equity portfolio with genetic-algorithm optimization and regime-adaptive rebalancing: a regressor→optimizer workflow (forward-looking rolling α/β targets → per-ticker predictions → bi-objective NSGA-II → Pareto-optimal weights) paired with HMM regime detection and Jensen–Shannon rebalancing triggers, reaching 5%+ annualized alpha-to-benchmark on a multi-year backtest. Also deployed multi-asset interday options strategies on pre-trained models: TabPFN classifiers gauging trade profitability on LoRA-adapted Amazon Chronos embeddings over technical indicators.
LLM agents & infrastructure. Built an MCP-wrapped, tool-calling LLM pipeline that mines trading signals against backtested PnL, surfacing entry/exit rules above 1.5 Sharpe. Built a Qwen2.5-VL multimodal PDF Q&A system with contrastive-embedding fine-tuning and Qdrant indexing, fronted by a FastAPI + Next.js + Streamlit stack. Ran it all on a serverless-first AWS stack (Lambda, EventBridge, DynamoDB, S3, CloudFormation) and Modal, with GitHub Actions CI/CD pushing digest-pinned ECR images across multiple bots, DST-aware EventBridge scheduling, Lambda tuned for ML inference with S3-backed HuggingFace/Torch caches, and React/static dashboards over DynamoDB/S3 served via FastAPI behind Lambda Function URLs.
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 Series A round.
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.
ML & Optimization
Model development, feature engineering, and optimization across classical ML, probabilistic models, time-series foundation models, and evolutionary search.
Cloud & MLOps
Serverless-first production systems with CI/CD, infrastructure-as-code, containers, and model deployment workflows.
Programming
Python-first engineering with scientific computing, data wrangling, and web scraping and feed ingestion for research tooling.
APIs & Dashboards
Production APIs and app backends connecting market data, broker integrations, feeds, and interactive analytics interfaces.
LLM & RAG
LLM pipelines and prompt optimization for trading inference, plus vector retrieval and multimodal document intelligence.
Beyond the code
