I've always been drawn to problems where the answer isn't obvious and the data is messy. That instinct has taken me from pure mathematics at Stanford to the trading floor at Morgan Stanley to leading an AI team at Lenovo. Each chapter built on the last, and each one taught me something I couldn't have learned any other way.
I chose math because I wanted the deepest toolkit I could get. While most of my peers were optimizing for a specific career path — CS, pre-med, consulting — I bet on versatility. Stochastic calculus, convex optimization, machine learning, distributed systems. I wanted to be the person who could look at any quantitative problem and know how to break it down from first principles.
That bet paid off faster than I expected. Stochastic calculus became the language I used for volatility modeling at Morgan Stanley. Convex optimization became the backbone of my portfolio construction code. The machine learning coursework gave me the foundation to build production AI systems at Lenovo. Every class that felt abstract at the time turned out to be exactly what I needed next.
Morgan Stanley was where everything clicked. For the first time I was working with data that had real stakes behind it — TAQ feeds with millions of rows per day, timestamps at the microsecond, every entry a decision someone made with actual capital. The pace and the precision were on a different level from anything I'd seen in academia.
I worked on equity stock selection, developing cross-sectional signals under purged combinatorial cross-validation. They improved simulated returns by 15% over the production baseline while keeping turnover below 80% and tracking error under 30 bps. But the microstructure work was where I really found my footing — I built TAQ-level pipelines for spread, order imbalance, and trade informativeness, then trained a GRU volatility model on order-book snapshots and intraday volume profiles. Under walk-forward validation it cut out-of-sample MAE by 8% on stressed weeks.
Morgan Stanley gave me strong quantitative foundations. The next step was clear: I wanted to build full systems — not just models, but the infrastructure, orchestration, and deployment that turn a model into something people rely on every day. That's what I do now at Lenovo, where I'm tech lead on the company's internal AI platform for supply chain and lead a team of five engineers.
The platform has gone through two phases. In the first, we built each agent workflow ourselves: supply-chain agents wired to live business data, running in production where they have to be right, fast, and cost-efficient. They proved the value quickly, and demand soon outgrew what any central team could build by hand.
So in the second phase we made the platform self-service: business users assemble their own workflows from drag-and-drop components, and I built the backend that makes that composition possible. Orchestration and workflow state run on LangGraph; a versioned config center lets prompts, skills, and components ship as config changes instead of deploys; and MCP forms the tool boundary, so every component declares exactly what data it can reach. The first vertical on that stack was an AI Observation engine, a ReAct agent that pulls live business data through MCP tools and is grounded by design: every answer comes from a tool call, and every number it reports appears verbatim in the data it read.
Across the platform: 8 production agent workflows, 17% fewer order-fulfillment errors, and batching, caching, and model routing that cut unit inference cost by 32% and p95 latency by 41%.
Outside of work I am an active researcher, with two papers on arXiv. One is theory: what a market maker can learn when valuations don't even have a finite variance. The other is empirical: how the market reacts when it learns that a corporate insider has bought, built from raw SEC filings. Both are on the Research page.