Co-Founder & CTO at ArbwickLos Angeles, CA
I build AI agents finance can trust.
I'm the co-founder & CTO of Arbwick, an AI research platform for institutional investors, after 3.5 years as a software engineer at Goldman Sachs. I ship agentic workflows end to end, from scoping ambiguous problems with customers to the tool, evaluation, and audit layers that make them production ready.
whoami
4yrs
in regulated finance
300+
agent functions, loaded on demand
130M
rows of market data from 17 vendors
200k+
client reports a day at Goldman Sachs
01/Experience
Four years in regulated finance: from the reporting engine behind Goldman Sachs' institutional clients to an AI research platform I'm building from the ground up.
Co-Founder & CTO
Apr 2026 — Present
AI research platform for institutional investors
Under the hood
Describe an investment idea in plain English. The agent plans the analyses, loads only the tools it needs, runs independent branches in parallel against live market data, and returns a backtested report where every figure traces back to the query that produced it.
▸ report ready · re-verify any number in one click
Sole engineer on the entire product: FastAPI backend, React/TypeScript SPA, 40+ pipelines ingesting 17 data vendors into a 130M-row time-series store, plus multi-tenant access control, usage-based billing, and production monitoring.
Work directly with portfolio managers and analysts to identify which parts of their research an agent can reliably take over, agree on what counts as working, then build and ship it.
Scaled the agent to 300+ analytical functions across 50 modules without the selection-accuracy collapse that caps most agent systems, by having it load tool definitions on demand instead of carrying the full catalog in context.
Built the quantitative function library the agent reasons over, implementing valuation, risk, and factor logic as tested code rather than prompt instructions, so every number in a report is reproducible and defensible to a client.
Keep a fixed set of validated queries the system is re-run against before every prompt or model change, so accuracy regressions surface before customers see them.
Architected the orchestration layer: the system takes an investment thesis written in plain English, plans the analyses needed, runs them in dependency order against live market data, and returns a backtested report. Running independent branches in parallel cut runtime 27%.
Made agent output auditable for a compliance-sensitive audience. Every figure in a report traces to the query that produced it and re-verifies against live data in one click, with per-request tracing across the full agent trajectory.
Passed large intermediate results between stages by reference rather than through the prompt, keeping context small across long chains where naive prompt-chaining silently drops data, with caching and per-call cost attribution over the same path.
payloads stored once · prompts carry pointers · cached · cost per call
Software Engineer, Associate
Sep 2022 — Apr 2026
Client reporting and document automation
200k+
reports a day
Designed and shipped the v2 client reporting engine (25+ tables) rendering 200k+ reports daily for institutional clients, working with stakeholders to add dynamic grouping and sorting for accounts at 20k+ pages/day.
98%
less handling time
Engineered an Apache PDFBox parser extracting structured data from unstructured financial documents, cutting handling time 98% (10m to under 10s) and eliminating manual entry.
3% → 60%
test coverage
Led a legacy modernization initiative: removed 400+ dead files and raised test coverage from 3% to 60%, unlocking daily deployments. Mentored new hires and interns through onboarding and production debugging.
02/Project
A side project that turns any LLM client into a careful trader.
An MCP server that lets any LLM client place real orders in plain English, with guardrails on anything destructive.
Runs onWeb UIWhatsApp
Long ETH perp, $2k notional at 3x. Stop-loss 4% below entry.
place_order
{ "symbol": "ETH-PERP", "side": "buy",
"notional": 2000, "leverage": 3,
"stop_loss": "-4%" }Destructive action. Confirm before sending?
Confirm
Order submitted with stop-loss attached.
03/Toolkit
From model APIs and agent patterns to the data, backend and observability layers that keep them honest in production.
04/About
I studied computer science and economics at UCLA, and I've worked where the two meet ever since: three and a half years building client reporting at Goldman Sachs, and now Arbwick, where agents do real research work for institutional investors.
The part I care about most is the unglamorous one: evaluations, audit trails and tested code underneath the model, so the answer holds up when a client asks where a number came from.
Education
Sep 2018 — Jun 2022 · Los Angeles, CA
Off the clock
05/Contact
Working on agents, data platforms, or anything where AI meets finance? I'd love to hear about it.