I build ML systems that real people use, and side projects that get a little out of hand.
Right now I'm a founding engineer at G8R Security, working on cryptographic security for AI agents. I'm also an ML engineer at GDX, where my code runs behind sites with well over a million visitors.
On the side: a game with friends became a multiplayer app with its own geographic data pipeline, and a Wordle habit became a solver for 32 boards at once.
The short version: hand me a messy problem and I'll get it all the way to something people can use.
I like owning the whole path: cleaning the data, building and evaluating the model, and making the screen someone actually uses. Polaris went from raw engine telemetry to a crew dashboard over one hackathon.
$30K Stanford grand prize, 1,500+ applicants
I treat compute like a budget. At Verbit I rebuilt speech-model evaluation so it needed 80% less of it, and when Python got too slow for StatAtlas, the hot path moved to C.
80% less evaluation compute at Verbit
At GDX I work directly with executives in English and Japanese. A lot of the job is asking the right questions early, then turning a loose idea into a roadmap and a production system.
Code behind sites with 1M+ visitors
Six things I'm proud of: the security work I'm doing at G8R, what I ship at GDX, a game I built for my friends, a hackathon winner, a puzzle I refused to lose at, and a planning tool from my day job.
NowFounding engineer · g8rsecurity.com
Zero trust for AI agents, with receipts.
AI agents are starting to take real actions: calling tools, moving data, touching production systems. G8R sits in the path between those agents and the outside world. Deterministic policies stop violations before they execute, and if something goes wrong, it fails closed.
I'm a founding engineer there, working on the cryptographic side of agent security: every verdict gets signed into a tamper-evident, per-tenant audit chain that stays inside the customer's own environment, so there's evidence nobody can quietly rewrite.
The hard path has a p99 budget under 50 ms with the signed audit commit included, so security never gets to be the slow part. The team is also working with NIST on how its AI risk framework should apply to agents.
NowML engineer · Nov 2024 to now
ML that has to work on a Monday morning.
At GDX I build the ML and automation that real businesses run on: an LLM agent that Shimamura's corporate teams use to standardize operational manuals and summarize internal data, plus the automation, testing, and tooling behind sites with well over a million visitors.
I also built and validated repurchase-prediction models for Onitsuka Tiger / ASICS, security tooling that drafts patches for engineers to review, and Gap Engine, which you'll find further down. Most of it starts in a meeting with executives, in English or Japanese, with a requirement that isn't quite a spec yet.
LivePersonal project · live at jetlag.drori.dev
Hide and seek, but the whole city is the board.
My friends and I play hide and seek across city transit networks, inspired by Jet Lag: The Game. I started building a companion app to help us play. It got out of hand, and now it's a full multiplayer game with live maps, curse cards, radar, and photo challenges.
The hard part lives under the map. A clue about elevation, a river, or a train line has to turn into a shape you can cross off. I built the geographic pipelines that do that with OpenStreetMap, Copernicus elevation, and Overture water data, with C++ doing the heavy lifting, so country-scale questions come back in seconds.
A Socket.io server keeps everyone in sync, the interface is built for phones, and Playwright tests catch regressions. The real QA, though, is running around outside with friends.
$30KStanford Sushi Hackathon · grand prize
Spotting engine trouble before it strands a ship.
Polaris reads a ship's engine telemetry and flags parts that look like they're heading for failure, early enough for the crew to do something about it. It won our team the $30,000 grand prize at the Stanford Sushi Hackathon in October 2025, out of more than 1,500 applicants.
My piece was the time-series pipeline, end to end: cleaning and windowing the sensor data, transformer-based anomaly detection, Wiener-filter estimates of component health, and a calibrated risk score for the crew's dashboard. It was my second hackathon. I've won both.
DemoPersonal project · browser demo + Chrome extension
How close can a solver get to a perfect game?
Duotrigordle is Wordle times 32: thirty-two boards and 37 guesses, all played at once. A word that's great for one board can be useless for the other 31, so picking the next guess is the entire game.
My solver scores every candidate word by how much it reveals across the boards that are still open, and a small value network breaks ties. It usually finishes in 33 to 35 guesses. Over a 2,000-game benchmark on the daily puzzle's own generator, it went 2,000 and 0.
WorkBuilt at GDX · internal tool
The plan is written. What did we miss?
Most planning tools stop once the task list exists. Gap Engine starts there. A project manager gives it a goal, and it drafts a plan from the team's own documents.
Then it comes back every day to look for holes: questions nobody has answered, and documents that quietly contradict each other. The point is to catch those while there's still time to fix them. I built the whole workspace, including the document checks, streamed plan generation, and a Discord bot for the team.
Research, side projects, tools, and a few things that only live on my hard drive. Click any row for the story.
Founding engineer at G8R Security, a self-hosted enforcement layer that blocks agent policy violations before they execute and signs every verdict into a tamper-evident audit chain.
The multiplayer game my friends and I actually play. Live maps, radar, curse cards, and photo challenges, sitting on geographic pipelines that turn clues into areas you can rule out.
$30,000 grand prize at the Stanford Sushi Hackathon. I built the pipeline from raw telemetry to the crew's risk score.
Drafts a plan from the team's documents, then comes back daily to flag missing answers and contradictions.
Usually done in 33 to 35 of 37 guesses, with zero losses across a 2,000-game benchmark.
Second-purchase prediction for Onitsuka Tiger / ASICS. Most of the work was making the data trustworthy: reconciling online and in-store records across 491,000 members, then evaluating with leakage controls, stratified validation, precision/recall, and top-k metrics.
Choosing where to live in California meant juggling pollution, commute, disaster-risk, and air-quality data. StatAtlas puts four public datasets on one searchable map. When Python got too slow for statewide scoring, I rewrote that part as a C extension.
How does a device prove it belongs in your fleet, report its health, and safely take orders? A Go prototype of that whole lifecycle, with a simulated device on the other end.
A GDX tool that reads daily reports from staff and flags duplicate work, stalled tasks, and recurring blockers. Managers can ask follow-up questions in plain language. The analysis core has zero dependencies and sits apart from the Discord bot, so it can plug into other chat tools.
SJSU research on how well multilingual BERT handles translation between languages with very different structures. Adding intermediary models cut errors on complex translations by about 15%.
I kept needing demo data that looked real, and every tool for it wanted me to sign up. So: about 50 field types, null rates, seeds, a live preview, and export to CSV, JSON, JSONL, or SQL. Nothing leaves your machine.
Built at GDX. It scans a codebase, takes a closer look at anything suspicious, and drafts patches. An engineer reviews each one and accepts, edits, or rejects it.
Batch image processing, scheduled reports, product-ranking updates, product-data collection, storefront checks, Playwright end-to-end tests, and an email server on Azure. The unglamorous automation behind sites with well over a million visitors.
A course project for SJSU CS185C. I rotated the GEBCO seafloor data so today's equator runs through the poles, coarsened the grid until a global circulation model would actually run, and spent the semester watching where temperature, salinity, and currents settled.
A two-pane desktop app: drag tracks onto the watch, pull them back off, and convert FLAC to MP3 on the way. Packaged for Debian, RPM-based distros, Arch, and AppImage.
Turns a script into narration, matched stock footage, and an edited short with word-timed captions, then uploads it. Each channel runs from one config file, and the repo is honest about the costs and limits.
I've been building and modifying RepRap printers since 2013: custom firmware, contributions to open-source slicers, and printers that print parts for other printers. In 2020–21 I built a camera-guided drone with a 3D-printed frame that plans its own path and avoids obstacles.
Hand-written HTML, CSS, and canvas, no framework. The contour map up top is simplex-noise terrain redrawn every frame, with a hill that follows your cursor. It's served from my desk in Saratoga through a Cloudflare Tunnel.
Nothing in that category yet.
A founding role, two ML jobs, three internships, and a perfect hackathon record.
Zero trust for AI agents
San Jose, CA & Japan
New York, NY
San Jose, CA · part-time
San Francisco, CA
Mountain View, CA
Minors in Mathematics and Japanese
I've been building and modding RepRap 3D printers since 2013, including a few whose whole job is printing parts for other printers. That's still how I pick projects: find something I want to figure out, then keep going until other people can actually use it.
I finished my B.S. in Computer Science at San José State in May 2026, with minors in Mathematics and Japanese. These days I'm an ML engineer at GDX, working between San Jose and Japan. Off the clock, there's a decent chance I'm on a train somewhere, hiding from my friends.
Python, C++, TypeScript, Go, Java, C#, Rust, OCaml, SQL, Shell
PyTorch, TensorFlow, JAX, Hugging Face, NVIDIA NeMo, LangChain, RAG, LLM agents, time-series
React, Next.js, Node, FastAPI, Postgres, Docker, Playwright, Azure, AWS, Cloudflare