_
Machine Learning Engineer — Robotics & Autonomous Systems
I build systems that run themselves — multi-agent pipelines running 24/7 in production, RL policies training live on this page, and spatial AI on iOS.
MODULE.01 // PROOF.LOG — VERIFIED OUTPUT, NOT ADJECTIVES
WORK CELL 01 // LIVE — COMPUTED IN YOUR TAB
DQN POLICY FROM MY TOH_RL REPO, RUNNING LIVE IN YOUR TAB — IT PICKS EACH MOVE, THE IK ARM CARRIES IT OUT. DRAG A DISK TO INTERFERE. lib/toh/policy.ts
MODULE.02 // MISSION.HISTORY
LIVE4–10 LLM agents deep-analyze every shortlisted property — crime, comps, news, income — turning 3,000 listings into 4 deals. Runs unattended with per-call cost telemetry, an API-error watchdog, and 36/36 tests gating every deploy.
A FastAPI service wrapping a PyVRP solver — dispatch endpoints (next-3-stops, mark-done, live re-optimization), config-driven scheduling, Docker, pytest. VRP is the same math as multi-robot task allocation.
A body of algorithmic-trading work built to be falsified, not sold — ARIMA/LSTM forecasting, FinBERT news-sentiment, a full technical-indicator stack, and RL trade policies.
The capstone is an analyst-consensus equity screener exposed as an MCP server, paired with a 35-report program that concluded the naive high-upside strategy loses once you correct for survivorship bias and analyst optimism — so it specified a sector-neutral composite validated with walk-forward + a Deflated Sharpe.
Three reinforcement-learning studies in one — Tower of Hanoi with curriculum learning + Hindsight Experience Replay, Atari agents, and an A2C trading policy on gym-anytrading.
Extended Apple's RoomPlan sample into a persistent room-scanning app — scan storage, human-detection guard, localization, and unit/UI test suites.
A full-funnel marketing OS I designed, built, and operate — six subsystems wired into one command center, deployed 24/7 on a droplet with nightly backups.
One command turns a JSON spec into a published video — image gen, TTS, music, programmatic Remotion render, auto-post to four platforms.
A recommendation engine in two regimes — a PySpark ALS + MinHash-LSH pipeline for large interaction matrices, and an SVD/KNN collaborative-filtering model for smaller sets — evaluated with MAP / NDCG / MRR.
Low-latency voice agents on the OpenAI Realtime API — chat-supervisor and sequential-handoff architectures (WebRTC streaming, server-side session management, output guardrails), extended from OpenAI's realtime-agents patterns.
A coursework project training an overcomplete sparse autoencoder to compare how learned features are organized across GPT-2 and LLaMA. Write-up and walkthrough on request.
Frontend of a short-term-rental management platform — dashboard, multi-calendar, unified inbox, crews, automation, pricing, reports; runs in demo mode on mock data.
Small operational tools shipped for real businesses — payroll, insurance CRM, lead auto-responder, lead-gen CRM, email keyword parsing.
full inventory: github.com/Alonerism — 22 public repos
WORK CELL 02 // LIVE — COMPUTED IN YOUR TAB
A FRESH REINFORCE POLICY TRAINS CARTPOLE ON EVERY PAGE LOAD — WATCH IT GO FROM FLAILING TO SOLVED. lib/rl/reinforce.ts
MODULE.03 // CAPABILITIES.MATRIX
I did the degrees — an M.S. in Data Science at NYU (May 2025) and a joint B.A./M.A. in Behavioral Science (2023), publishing behavioral-science research (PsyArXiv) along the way.
I also taught the recitation sections for NYU's graduate data-science courses — Big Data and Intro to Data Science — running my own classroom, office hours, and grading for 300+ master's students. I've been tutoring quantitative reasoning for years, GMAT math prep included; it's still the fastest way I know to find out whether I actually understand something.
Then I spent a year running my own AI studio: designing, shipping, and operating autonomous pipelines solo — from the model to the server to the production metric. I own the whole loop.
Point me at a domain I've never touched and a deadline that's too short — that's my favorite kind of problem.
Now I want to build robots with people smarter than me.
PERSONAL // OFF-DUTY LOG
AT THE HELMBefore any of the tech, I ran math and English classrooms in a Cape Town township as Director of Volunteers, coordinating a 40-person international teaching team.
I've been a licensed skipper since 2016 — I like systems where reading the conditions and trusting your instruments is the whole job. I've traveled most of the world; Southeast Asia won.
On land or water I'm usually moving — surfing, windsurfing, kiting, whitewater rafting, and the occasional ill-advised pickup soccer game — and currently trying to suck less at guitar.
MODULE.04 // ESTABLISH.LINK
Hiring for robotics, autonomy, or ML infrastructure? Grab 20 minutes — I'll bring demos, not slides. Or grab the resume and email me.