Experience
The full timeline. Newest first.
–
Independent research
BEMA: stress-testing a “can’t-hallucinate” decision model
An open, MIT-licensed reproduction of a typed-decision model design, built to test the claim that a model with no free text to generate can’t hallucinate. I specified the experiments and directed Claude Code agents through the build and write-up.
- Two adjacent-character typos cut BANKING77 accuracy from 81.9% to 51.2% (n = 3,080), while mean confidence only fell from 0.79 to 0.61.
- The gap held on a second domain: CLINC150 went from 71.8% to 45.0% accuracy while confidence fell from 0.75 to 0.56 (n = 5,500).
- Reported the negative result: on the current run, temperature scaling made the 77-way head’s calibration worse (ECE 0.0210 → 0.0371); isotonic recalibration brought it to 0.0113.
- BANKING77 accuracy under two typos (n = 3,080)
- 81.9% → 51.2%
- Mean confidence under the same typos
- 0.79 → 0.61
- CLINC150 under typos (n = 5,500): accuracy · confidence
- 71.8% → 45.0% · 0.75 → 0.56
- 77-way head ECE: raw → temperature → isotonic
- 0.0210 → 0.0371 → 0.0113
Pitched Redl to investors
Redl, my local-first AI workbench
Pitched Redl to investors in San Francisco (Jul 2026).
Pitched to investors in San Francisco
Intern, team build with Fluxxion88
Loop Engineering Hackathon, AWS Builder Loft
Competed on a team with my collaborator (Fluxxion88). We built Intern, which trains an agent like a new hire: from one hand-made example, an LLM loop writes, runs, scores and repairs a script until it matches, then hands back a script with no model inside.
- The team’s recorded run climbed from a 68% to a 96% to a 100% match in three attempts.
- Recorded run, attempts 1 → 2 → 3
- 68% → 96% → 100% match
Warrant + Forge, two-person team with Fluxxion88
Estate-settlement AI hackathon
Built Warrant + Forge with my collaborator (Fluxxion88) over two days, with Warrant forked from Concord’s verifier, ingestion and provider layers: an estate-settlement engine where no fact reaches the ledger without a verbatim quote verified in its source document. It didn’t place, but I opened the AccountWard repo the day after.
- I directed AI coding agents through Warrant, the engine half; Forge, the form-filling half, is my collaborator’s.
- Warrant caught 41 of 41 rewritten source quotes before they reached the ledger.
- Rewritten source quotes caught before the ledger
- 41 / 41
–
Software Engineering Intern
early-stage fintech startup (stealth)
Worked across the stack on data workflows and user-facing features, directly with the founders on architecture and product decisions.
NEXUS, solo entry, Wargaming track
SCSP AI+ Expo Hackathon (Phase 1, San Francisco)
Entered the Wargaming track solo with NEXUS, an AI wargame adjudication engine for supply-chain resilience analysis, and demoed it live to the track’s national-security judging panel, who probed the simulation logic and the AI adversary.
- Solo entry: I architected and specified NEXUS and directed AI coding agents (Cursor and Antigravity) through the build: a Monte Carlo cascade simulation argued over by a chain of AI agents, streamed to a 3D globe, plus a companion Android app.
Demoed live to the Wargaming track’s national-security judging panel
–
Riptide: live AirSim drone mode and volunteer Android app
Hack for Humanity 2026, Santa Clara University
My team built Riptide, a platform for coordinating ocean-pollution cleanup. I architected and directed an AI coding agent through a live Microsoft AirSim mode behind its drone console, and through a volunteer companion app for Android on the final morning.
- A Python Flask bridge flies a simulated quadcopter over RPC: telemetry, camera, depth and LiDAR feeds, and threaded waypoint missions.
- The simulator’s own log records three bridge-driven missions.
- Both pieces ran in my build; the team’s submission kept its map-based simulation.
- Bridge-driven missions in the simulator log
- 3
- Bridge endpoints
- 14
SlugBites: 2 awards
CruzHacks 2026, UC Santa Cruz
Our team’s SlugBites won two single-winner category awards: Most Ambitious Project and MLH Best Use of Solana. It’s drone-delivery ordering for UC Santa Cruz dining halls, built on live-scraped menus, with a Python flight layer written for Microsoft AirSim.
- I wrote 20 of the repo’s 23 commits, including the scraper that turns a human-facing nutrition site into a menu API, meal-period pricing and a SOL-priced checkout, plus the AirSim flight layer.
- Round two of an idea: three months earlier it was AeroBites at the AWS x INRIX Hack.
Most Ambitious ProjectMLH Best Use of Solana
- Commits (mine / total)
- 20 / 23
- Submissions at CruzHacks 2026
- 88
AeroBites: drone food delivery, v1
AWS x INRIX Hack 2025, SCU ACM
Built AeroBites on a five-person team: a React ordering app for campus drone delivery, with an AWS serverless back end written as infrastructure-as-code (Lambda, DynamoDB, S3 and a Bedrock function). I was the top committer, including every commit that changed code.
- We had no drone, so on the final morning I stood up Microsoft AirSim’s Unreal Engine environments and flew a quadrotor in simulation.
- Three months later the idea came back as SlugBites, on live data, and our team won two awards with it at CruzHacks 2026.
- Team size
- 5
- Commits (mine / total)
- 13 / 18
–
Triple major: Finance (BS), Computer Science (BA), Economics (BA)
Santa Clara University, Leavey School of Business
Class of 2029. Most finance people can’t ship software, and most engineers don’t follow where the money moves. I’m studying all three sides: finance, computer science and economics.
- Statistics and conditional probability (OMIS 40)
- Trade and subsidy economics (ECON 3)
- Capital-structure strategy in a shareholder-value simulation (BUSN 70)
- Information systems and web development (OMIS 34)
- Fall 2026: writing on how generative AI works and why it fails, with my own AI systems as the case studies (STS)
- Competed at SCU ACM’s AWS x INRIX Hack (Oct 2025) and SCU’s Hack for Humanity (Feb 2026)
- Majors
- 3
- Class of
- 2029
–
Vice Branch & Fundraising Lead
DevFinTech
Ran fundraising and corporate-partnership outreach for financial-literacy programs.