Sole PM, designer and engineer. The non-goals were written before any code, which is the only reason a week was enough.
Product Management · Berkeley Haas
I write the PRD, then I build the thing.
MBA / MEng candidate at UC Berkeley Haas & IEOR. Three years atAWS building analytics products for an $80B sales org — then I started shipping my own.
Seeking Summer 2027 PM internship
Focus consumer · AI/GenAI · enterprise data
Based Berkeley, CA
01 — Shipped alone · Aug 2026
TripMatch
A 400-person cohort coordinated rides by scrolling a WhatsApp chat, where requests get buried and matching seats go unused. I shipped a verified, Berkeley-only rides board — problem framing through production.
Read the case study →Verified sign-in, structured board, seat claims with capacity enforcement, comments, audit log, feedback inbox.
The day before launch, a race condition let two simultaneous posts silently delete each other. You post, you see it appear, on refresh it's gone — so you post again, which makes the race more likely. I rebuilt the storage layer rather than ship it.
Three feedback-driven iterations after launch: owner-scoped delete, in-place edit that preserves comment threads, and date chips that only appear once the board spans more than one date.
The product decision
Non-goals are the product decision.
TripMatch shipped in six days because I wrote down what it would never do, before writing any code. Every line below was a deliberate cut, and each one bought the schedule.
Payments / cost-splitting
Venmo already works. Adding money introduces trust and liability complexity that buys no additional matches.
Real-time dispatch
The job is surfacing the match. Logistics between two classmates is already solved — they text each other.
A WhatsApp replacement
Final coordination stays in chat. Owning the whole conversation would mean competing with an app at 100% adoption.
Public / open matching
Opening to strangers re-introduces the exact trust problem the design exists to avoid.
A native mobile app
A link is the right form factor — zero install friction, one tap from the group chat where the users already are. Distribution determined the form factor, and later produced the nastiest bug in the project.
02 — 0 → 1 at scale · AWS, 2023–2026
Action Hub
A sales org running on roughly 200 dashboards, with no agreed place to start the day. I owned the technical chapter's first tab: every open issue across a seller's accounts, ranked, with the next action named on each one.
Read the case study →Everything else
The full shelf.
7 more — agentic tooling, forecasting, data governance, go-to-market, full-stack engineering, and reinforcement learning.
Pull-Request Automation
Agentic AIBuilder and product owner · Amazon's internal OpenClaw platform · 2023–2026
Getting a finished change into code review took eight mandatory steps and 30–60 minutes, none of it thinking. A skill built on Amazon's internal OpenClaw platform now does all eight, leaving about five minutes of genuine decisions — and stops precisely where a human reviewer starts.
80% less manual processing time5 min human input, from 30–608 ceremony steps automated
GenAI Target Setting
ForecastingAnalytics and target-setting methodology · AWS Global Sales Strategy & Analytics · 2024
The first year AWS carried a GenAI goal, the metric being targeted had barely existed twelve months earlier. Forecasting it directly was impossible, so the model forecast a leading indicator that did have history and converted through it — then shipped with a quarterly mechanism to correct itself.
55,000 data points4 scenario simulations3-month cycle-time lag
Allowlist
Data governanceBuilder and product owner · AWS Global Sales Strategy & Analytics · 2023–2026
Exceptions to sellers' data access lived in a spreadsheet anyone could edit, with no record of who changed what. Allowlist replaced it with a constrained application: read everything, add, remove — but never edit in place, and every action logged.
200+ data owners and ops partners0 in-place edits permittedEvery action logged
GenAI Adoption Program
Go-to-marketProduct owner · with the LATAM sales director and a principal seller · 2024
GenAI tooling existed; usage didn't. Rather than mandate it, I partnered with the LATAM sales director and a principal seller on a points-and-leaderboard programme that reached over 75% participation in 60 days — because the rewards were designed for the people who would never finish first.
1,100 LATAM sellers>75% participation in 60 daysWeekly and monthly ranking cadence
Sole designer and engineer · Sep 2026
A networking tracker for Berkeley contacts where one user's list is unreachable to another even if the API layer were bypassed entirely — because the boundary lives in Postgres, not in application code.
3 independent ownership mechanisms28 validation tests404 returned, not 403
Ms. Pac-Man DQN
Reinforcement learningMBA 290T — Fundamentals of Agentic AI · Sep 2026
Mean evaluation score rose from 492 to 906 (+84%). The more interesting finding was that the untrained baseline wasn't random at all — it was stuck, repeating one move 95.4% of the time, which means the headline number flatters the result.
+84% 906 vs 492 baseline8 search runs1.6M agent decisions searched
Sole engineer · Sep 2026
Scans roughly 29,400 postings in about 50 seconds every weekday, and opens a GitHub issue only when something new appears. The hard problem turned out not to be finding roles — it was deciding which two per company were worth your attention.
~29,400 postings per scan50s full scan time2 roles shown per company