Work
What I shipped, and what it changed.
Production AI inside live operations. 0→1 infrastructure in rates, FX, and cross-currency. Then RideLens, Daylight, RailDrop, and Gridiron: independent products built end-to-end and shipped live. Filter by Agents, Markets, or Labs for the path that matches the role you are hiring for.
01/11 · Setup Agent
-
Operations agent that collapsed a 3.5-hour bottleneck
Lead Product Manager
A production AI agent inside a live operations workflow, not a side chatbot. Bounded actions, operator visibility, and zero AI-initiated configuration errors in its first six months.
3.5h → to 8msetup time3×run volume0AI config errors / 6 mo
Read the case
Before Gate After
-
Incident agent that returned 750+ engineering hours a year
Lead Product Manager
When something broke in production, ops still paged senior engineers by default. The agent gathers domain evidence first, sharpens the diagnosis, and escalates only when the evidence is incomplete.
4.5h → to 11minvestigation−78%escalations750+eng hours / year
Read the case
Incident Evidence Diagnose
-
Multilateral compression for trillions in trapped notional
Product Manager
Banks held large cross-currency books that could only net in pairs. A central settlement path made network-wide offsets executable, so notional stuck in pairs became eligible to compress.
+34%notional reduction / run18 banks12 currency pairs$X.XTeligible notional
Read the case
Gross Run Compressed Risk envelope
-
Dual-source valuation that caught discrepancies 48 hours earlier
Product Manager
A late valuation mismatch can halt a multi-trillion compression cycle. An independent second source removed that single point of failure and moved failure upstream, while it was still cheap to fix.
−91%resubmissions48hearlier detection$XT+cycle notional
Read the case
Split Diff Lock
-
FX compression with margin intelligence in the optimizer
Product Manager
A mathematically good FX proposal still fails if margin or data disagree. Margin checks moved inside the optimizer, and four-source validation gated inputs before the live window.
40+live runs100%proposal acceptance−94%live failures
Read the case
Book Gates Accept
-
A reusable control plane for five enterprise agent systems
Lead Product Manager
One-off agent integrations do not scale. A shared control plane for context, typed actions, approval, and evals made the next agent a reuse problem, not a prototype from zero.
5enterprise systemsMCP + APIsreusable interfacesHITL + evalsproduction controls
Read the case
Context Actions Services HITL Eval
-
Pre-trade simulator that made capital a decision, not a surprise
Product Analyst
A simulation engine that showed the next trade’s capital cost across venues before the desk committed. Comparing incremental with standalone cost made portfolio offsets visible.
0→ to 1originated and shippedup to −30%initial margin20+enterprise clients
Read the case
Base + Trade Simulate Compare
-
RideLens: every ride, one comparison
Product, design, and build
Live roads. Real rate cards. Uber, Lyft, Empower, and Curb ranked before you open four apps and guess.
4providers, one board3ways to rank a tripHonestranges stay ranges
Read the case
Route Price Soonest Value
Open RideLens (opens in a new tab) -
Daylight: your screen, through the day
Product, design, and build
Adaptive display lighting for macOS. Warmth and brightness on a schedule you set: gradual, offline, and honest about what it is doing and why.
6layers, one fixed order3levels of certainty263tests, no dependencies
Read the case
Schedule Resolve Explain Confirm
Open Daylight (opens in a new tab) -
RailDrop: know when your train gets cheaper
Product, design, and build
Book the trip. Watch every bookable Amtrak option across your window. One email when a listed fare actually drops. Never an invented price.
±1 daydefault watch window1 emailonly when it improves0invented prices
Read the case
Booked Watching Board Alert
Open RailDrop (opens in a new tab) -
Gridiron: every game, every drive, one view
Product, design, and build
A live NFL and college football command center. Every game gets its own 3D field with the reported ball spot, beside the score, win probability, and odds from named sources.
3sources, each named0guessed ball spots669unit and integration tests
Read the case
Slate Drive Odds Touchdown
Open Gridiron (opens in a new tab)
1-4 facets · j / k ledger · Enter opens · y link · ⌘K jump