Developer tools
Build systems · AI infrastructure
Stripe
Previously Google
Games built in Luau, with assets modeled in Blender.
Aim in secret, launch together, stay out of the lava
For two to six players or solo against bots. Choose your direction and power before each launch. The disk shrinks every round.
Roll, squish, and stack a tower of giant squishies
Stack collectible squishies into a tower that earns coins. I generate the meshes with Python scripts in Blender.
IN DEVELOPMENT
Steal an egg, outrun a giant squishy boss, and hatch it back at your plot. Your collection earns coins; treadmill training helps you reach the next floor.
Playable in Studio · private staging · not publicly released
PLAYABLE SKETCH
You are the yellow Verity. Aim, launch, and stay on the disk.
Drag from your Verity and releaseor ← → aim · ↑ ↓ power · Space launch
Last Verity on the disk wins
Browser prototype
Unpublished prototypes
Steal food from a shared pantry, carry it to income plots, and raid other players. The exit has a safe zone.
Roll twelve collectible variants and place them in six income bays. Coin-funded rarity packs show their odds. The server determines rolls and income.
Choose a ball, aim, and watch it battle for collectible unlocks. Fixed-step simulation and recorded inputs make matches reproducible.
Choose a character and lane for a twelve-second obstacle trial. Earn medals and unlocks.
Print memes, collect earnings, upgrade four machine tiers, and rebirth. The factory runs through an animated interface.
Make words from six letters in sixty seconds. Collect modifiers that change your score.
Describe a target without using forbidden words. A rule-based judge scores your clue without a live AI model or stored clue text.
Studio prototypes, not published. Verity Knockout and My ASMR Squishy Tower are the two released games.
Prediction-market research and automated trading. One EC2 host.
Python / SQLite / systemd / EC2
Bitcoin above $X at 5 PM?
Highest temperature in Miami today?
What are the biggest accounts buying?
Forecast research and market-data collection continue. These are cumulative research observations, not trades. Forecast validation has not cleared promotion.
EXECUTION Funded executor paused pending reconciliation. Continuous paper executor paused at its loss limit.
DATED HOST SNAPSHOT · THE SETTLEMENT READOUT BELOW UPDATES AUTOMATICALLY
No verified snapshot loaded
| DATE | MKTS | NET | FEES | CUM |
|---|---|---|---|---|
| AWAITING SNAPSHOT | ||||
Reference prices and public market data
Forecast, validation and relative-value studies
Separate paper state; loss-limit gate reached
Pre-trade blockers decide whether live may run
Reconciliation required before funded execution
Settled results, published on a 15-minute delay
AS INSPECTED 11 OCT 2026 · RUNNING SERVICES DO NOT IMPLY ACTIVE TRADING
Can a model beat the market's own price? Each one starts from that price and learns only a correction to it.
| MODEL | WHAT IT IS | STATE |
|---|---|---|
| CatBoost | Gradient-boosted trees fitted on market log-odds as an offset | EVALUATED |
| TabICLv2 | Tabular foundation model, run locally from a pinned checkpoint | FELL BACK TO MARKET |
| TabPFN | Adapter written; no licensed weights | NOT RUN |
| BASELINE | 0.0958 |
|---|---|
| MARKET | 0.0928 |
| CATBOOST | 0.0928 |
| BASELINE | 0.3178 |
|---|---|
| MARKET | 0.3082 |
| CATBOOST | 0.3078 |
VERDICT Effectively ties the market. Not promoted.
DEVELOPMENT DATA, NOT AN UNTOUCHED HOLDOUT · ISOLATED RESEARCH ENVIRONMENT · NO EXECUTION INTERFACE
Settlement revenue minus reported position costs and fees. Open positions, unrealized P&L, deposits, withdrawals, and paper trading are excluded.
Primary Kalshi account. May include manual trades; these are account settlements, not bot-only performance.
A zero is a verified reading: the exchange returned no cash settlement inside the window. Missing, stale, or failed data is labelled as such and shown as dashes, never as zero.
Shared model access and device-data systems at Google, 2022–2024.
ENGINEER
A shared API for Google’s generative AI experiences
One of two engineers on the Imagen and Bard gateway used by Android AI Wallpaper, Android Auto, and Messages. I owned design, implementation, quotas, experiments, CI/CD, and on-call.
Powered demos at Google I/O 2023 and CES
Metrics / quotas / experiments
One interface for image and text generation
ENGINEER
Measuring how models, fonts, and other packages reach devices
Before Aratea, I built rollout analytics for Google apps, including Search and YouTube. HLL/KLL sketches aggregated device events for metrics, anomaly alerts, and A/B tests.
ONE BILLION EVENTS A DAY IS≈ 11,574 every second
Language upgrades, editor tooling, and builds at Stripe.
PLATFORM LEAD
From quarterly migrations to continuous language upgrades
I led the platform that checks live Bazel dependencies, groups eligible upgrades into reviewable PRs, and reports blockers.
26M-line JVM repository · 700+ Java services
A library can compile on a newer JDK while a service consuming it still runs an older runtime. Eligibility accounts for reverse dependencies, runtime caps, source/test coordination, and framework guards.
The platform partitions eligible targets into bounded PRs. Repeatable runs work continuously or on a selected package; skipped targets retain visible blocker categories.
In the interactive example, both apps consume //lib/common. Moving //app/b to Java 17 clears the runtime constraint. A production upgrade still needs the remaining graph and test checks.
ILLUSTRATIVE / AFTER ELIGIBILITY CHECKS− target_jdk = 11
+ target_jdk = 17Coauthored with Cara Mecozzi.
app/b still runs Java 11. The shared library’s language upgrade is blocked.
IMPLEMENTATION
Less build noise in coding-agent context
I implemented --config=llm across Stripe’s Ruby, JVM/Python, and Go monorepo. It strips ANSI escapes and verbose build/test output before that text reaches coding agents.
Used by Minions, Codex, and Claude Code at Stripe
Loading: 0 packages loaded Analyzing: target //demo:parser_test INFO: Found 1 test target... [14 / 15] Testing //demo:parser_test FAIL: //demo:parser_test ParserTest.java:42: expected 3, got 2 at ParserTest.parsesItems(ParserTest.java:42) Executed 1 test: 1 fails.
ROLLOUT LEAD · UPSTREAM CONTRIBUTOR
Java tooling in VS Code and Cursor
I led Stripe's Metals V2 rollout and built the BSP server connecting Bazel to VS Code and Cursor. I worked onsite with Cursor and Databricks engineers on Java support.
26M-line JVM repository
#8800Prefer target classpaths over Turbine headers; resolve generated protobuf builder methods and Java packages.#8590Resolve BSP-generated Java sources to their build targets, including symlinked roots; prevent stale diagnostics.#8607Java Bazel test code lenses, backed by current SemanticDB test-class information#1963Custom server launcher for the VS Code extension, enabling integration with existing LSP proxy and observability tooling#1998Protobuf and prototext document selection, enabling proto-to-proto navigation through Metals#2039Reverted language-based proto/prototext activation in the extension.A protobuf class can exist on the build target’s classpath while the editor resolves a less useful version from Turbine headers. My fix prioritizes the target classpath, keeps headers as a fallback, and tests generated builder-method resolution.
Generated files and symlinked roots must map back to their build target. That gives the editor the right classpath and prevents stale diagnostics. Java test actions similarly need current SemanticDB information.
Metals + Bazel on GitHub ↗01 var receipt = Receipt.newBuilder() 02 .setId("example") 03 .build();
Receipt.BuilderGenerated Java class
Builder methods available to the editor
Build-target classpath first; Turbine headers remain a fallback.
TOOL AUTHOR · AI + COMPILERS
Migrating Scala to unblock JVM upgrades
I built llmigrate, a Go CLI that coordinates five Claude agents through the migration steps below.
Generated bridge code and cross-language tests validate Java implementations against Scala behavior.
BUILD INFRASTRUCTURE
Remote execution, caching, and versioned dependencies
I improved remote execution for JVM development, enforced versioned dependencies in Starlark, and introduced Java quality gates with Error Prone.
700+ Java services
Unchanged inputs can reuse a stored result
Flight software, mobile experiences, and deployment automation.
INTERN · SUMMER 2017
Contributed to FlightHorizon, which combines aircraft position data to help unmanned aircraft detect and avoid collisions.
2017 R&D 100 Award finalist
INTERN · SUMMER 2018
Won the Nike Digital Intern Hackathon with Elefit, a live workout-tracking app. Built dynamic string localization infrastructure for Nike apps.
OpenText · Software Engineer · 2019–2022Built reusable Ansible and Makefile deployment templates adopted across most team services.UC San Diego · B.S. Mathematics–Computer Science · 2019
I built a bot for underground music lyrics with AWS Lambda and Python
16K FOLLOWERS AT PEAKEarly implementation on GitHub ↗Cofounder & CTO
Turned smart TVs into digital billboards with WebSockets, scheduling, and payments. Top 10% of our Y Combinator applicant batch
trillboards.comBuilt the initial portfolio website for Alex Headlam and hosted it on AWS EC2.
aleqth.com/portfolio
Explicit dependencies, reproducible builds, and reusable build results.
Graphs, actions & cachingPersistent terminals for builds, logs, and coding agents.
Source ownership, classpaths, and a working “go to definition.” Upstream contributor.
metals-lsp.orgJava · Scala · Go · Python · Kotlin · Rust · SQL · AWS · Docker · Observability
