How I Build
The projects below are what I ship. This is how they work and what went wrong on the way. Each card links to the full write-up, with the numbers.
Serverless on AWS
Python 3.10 Lambda functions deployed with the Serverless Framework, behind API Gateway, with one DynamoDB table and Cognito for accounts.
API Gateway gives a function 29 seconds and a twin conversation takes longer, so a "hunt" returns at once and a worker with a 5-minute limit does the real job.
Read the write-up AI Exodus and CompanionLLMs in production
Gemini Flash writes and grades the twins' conversations; 768-dimension embeddings give the twin a memory ranked by similarity, importance and recency.
Two real bugs from model "thinking": the grader scored every conversation exactly 50 because thinking used the whole output budget, and the Companion ran slow because it switched on quietly.
Read the write-up AI Exodus CompanionReal-time AI on the player's PC
An Electron app with one pack per game. Voice runs locally: Chatterbox Turbo in ONNX for speech, a local Whisper model for talk-back.
Screen reading uses the Windows built-in OCR at 2 to 18 ms per crop, after my first engine took 100 to 400 ms. A voice that arrives late is worse than silence.
Read the write-up Animal Life: SurvivalGame engineering in Unreal Engine 5.9
My own leg solver learns each animal's joints from its animation clips, because off-the-shelf IK failed on one skeleton in more than fifteen configurations.
Physics substeps at 1/60 s so the world runs in real time at any frame rate, and a server pose level of detail saved about 2.8 ms per frame in a six-player test.
Read the write-up SUTOKTesting my own product
A plain Node.js script of about 90 lines replayed the site's own engine over 27,709 next-day calls. The verdict was right 37.5% of the time; "always up" was right 43.3%.
I fixed the rule set and a pass/fail gate in advance before trying alternatives, and a nightly forward test records each day's verdict so it cannot be tuned after the fact.
Read the write-up Minai in World of WarcraftAn AI that acts on its own
A loop of look, weigh the options, refuse what has hurt her before, then act and say something aloud. The refusal list is her death map.
29,164 decisions logged so far. She is never told the words game, player or quest, and her lore on the site is the same text her mind receives.
Read the write-upThe stack I work in
- AWSLambda, API Gateway, DynamoDB, Cognito, S3, CloudFront, SES, Route 53, GameLift, Pixel Streaming
- Unreal Engine 5C++ and Blueprints, multiplayer, plugins, UEFN
- AIGemini, embeddings, local Whisper and Chatterbox
- LanguagesPython, Node.js, TypeScript, C++
- AppsReact, Vite, Capacitor for Android, Electron for Windows
- ToolingDocker, Serverless Framework, Playwright