Now: Accessibility AI research @ NYU · JR Rizzo Lab◈ 9+ products shipped solo · 5 live right now◈ MS Computer Science, GPA 4.0
scroll▾
01 · Present
Builder first, researcher again
I'm a computer-science entrepreneur: for the past decade I've designed, built, and operated complete
products solo — the models, the frontend, the payments, the ops, the 2 a.m. production debugging.
I published cited research early, spent the years since shipping things people actually use, and now
I'm bringing that shipper's toolkit back into the lab.
Current role
Accessibility AI Engineering & Research — NYU, JR Rizzo Lab
I work on AI for accessibility at NYU in the lab of Dr. John-Ross (JR) Rizzo — the group behind
VIS4ION, a wearable assistive-technology platform that fuses cameras,
sensors, and AI services to help people who are blind or have low vision perceive and navigate their
environment. The lab's work spans NYU Langone's Rusk Rehabilitation and NYU Tandon, and its
commuter-assistive research is backed by a
$5M NSF Convergence Accelerator grant.
My focus: applying modern machine learning — computer vision, multimodal models, and LLM-driven
systems — to real assistive outcomes, where "does it work?" is measured in human independence.
DomainAssistive tech for blind & low-vision users
StackComputer vision · multimodal AI · wearables
WhereNYU Langone Rusk Rehabilitation × NYU Tandon
9+
Products shipped solo
5
Live in production
10+
Years building with ML
320+
Research citations
02 · Trajectory
A decade of signal
From GAN research at MILA to satellite signals at Apple to accessibility AI at NYU — each entry expands. Click around.
Accessibility AI Engineer & ResearcherNYU · JR Rizzo LabPresentNew York, NY — NYU Langone Rusk Rehabilitation × NYU Tandon ＋
Engineering AI systems for people who are blind or have low vision, within the lab behind the VIS4ION wearable navigation platform.
Research at the intersection of computer vision, multimodal AI, and human-centered assistive technology.
Translating state-of-the-art models into deployed tools measured by real-world accessibility outcomes.
Machine Learning EngineerMealMate Inc.2023 — 2024Los Angeles, CA ＋
Built a cosine-similarity matching system over embeddings in a vector database, powering semantic food search.
Wrote advanced preprocessing pipelines to augment and load USDA food data into AWS Postgres for the app's frontend and backend.
Owned database work in pgAdmin — views, tables, and schemas for custom MealMate data.
Full-stack multiplayer AI image-guessing platform · guessprompt.com
Live for 20 months: 8 real-time party modes (phone-as-controller with a synced TV screen), a daily puzzle with embeddings-based similarity scoring, and Elo-rated head-to-head matches — FastAPI + Next.js + Supabase, designed, built, and operated solo.
970+ consecutive daily AI puzzles and 4,200+ generated images across four image-model providers (Gemini, FLUX, Qwen, OpenAI) behind a unified engine with automatic fallback and per-image cost accounting.
Monetized in production: Stripe checkout with per-seat dynamic pricing, idempotent webhook fulfillment, credit rollover via claim links, and a generated-image-to-merch print pipeline.
Operated like a product, debugged like an engineer — 266k-email retention system, and a production P95 of 180s diagnosed down to sub-second (a sync CPU-bound spell-check blocking the async event loop).
Automation-native e-commerce for a master luthier · mybowexpress.com
Rebuilt a bow restorer's business (rehairs & repairs for string-instrument bows) as an automation-native storefront: Stripe checkout, automated customer emails, and automatic shipping-label creation.
Replaced a fully manual workflow — payments, buyer communication, labels, and order records now flow through with zero hands-on work.
Embeds the real site, scaled down — and still clickable.
Featured · Autonomous agent
Fable Swing-Trader
A Robinhood swing-trading agent with an LLM as its brain
An autonomous agent that reads live market news and makes swing-trade decisions on Robinhood, with an LLM (Fable) as the decision-making core.
After every closed trade it updates a live calibration card — a running scorecard of its own performance: win rate, expectancy, streaks, drawdown.
The card is dynamically injected into future prompts, so the model reasons against its own verified track record — a self-calibrating feedback loop that turns past results into future judgment.
→ The card on the right is a live simulation of that feedback loop. Watch it trade — or press "run cycle" to step it yourself.
“Chipmaker beats earnings; guidance raised”
→ decision: LONG · confidence 0.72 · sizing: ¼ Kelly
simulated demo · not financial advice
MLBot
Algorithmic trading system for Kalshi prediction markets
Live-trading bot for MLB player-prop contracts: MCMC price simulation + Vegas-odds divergence + ESPN stats, fused by an XGBoost meta-model trained on 17K+ historical markets.
Fractional Kelly sizing, real-time exposure caps, and a per-minute backtesting engine — 36% validated ROI over 2,400+ simulated trades.
Found and exploited a structural favorite-longshot bias on NO-side contracts (339% ROI in backtest); crash-recovery rehydrates all state from the exchange API.
PythonXGBoostKelly criterionasyncio
ConnectIn
AI-assisted telehealth EMR platform
HIPAA / PIPEDA / Quebec Law 25-compliant EMR on Firebase + GCP: 30+ Cloud Functions, RBAC, App Check, signed Stripe webhooks, PHI-redacted logging.
Gemini-powered clinical intake generating personalized triage questions across 500+ reasons-for-visit, fully localized in EN/FR/ES.
FHIR R4 interoperability, Cloud Vision OCR health-card scanning, RxNorm + Health Canada drug search, and auto-issued Google Meet visit links.
CNN that reads chessboard images into FEN positions, plus engine analysis and NLP explanations of the position — Flask backend, React frontend.
Grew so popular on Chrome that Google ultimately removed it: chess.com players were using it to gain an unfair edge online. (Build something people shouldn't want this much.)
Aggregates the same story from multiple outlets and fuses them into one unified summary, deliberately averaging out each source's bias.
TypeScriptsummarization
04 · Published
Research bookends
I published in Pattern Recognition, Plastic & Reconstructive Surgery, and JAAD at the very start of my
career — work that's been cited 320+ times while I was off building products. Now the two threads meet:
production-grade engineering, applied to accessibility research at NYU.
Hou, Nguyen, Kanevsky, Samaras, Kurç, Zhao, Gupta, Gao, Chen, Foran, Saltz. A sparse convolutional autoencoder for unsupervised nucleus detection — state-of-the-art on four datasets at just 5% of the annotation cost.
Kanevsky, Corban, Gaster, Kanevsky, et al. Charted how machine learning can support surgical decision-making across burn, micro-, craniofacial, and aesthetic surgery.
Journal of the American Academy of Dermatology, Vol. 78(3) · 2018
Safran, Viezel-Mathieu, Corban, Kanevsky, et al. Analyzed 29,971 histologically proven skin lesions across 50 ML screening techniques for melanoma detection.
JAADjournal
05 · Recognition
Early proof
№1
Intel ISEF — First Place, Computer Science
May 2014 · out of 1,700 finalists worldwide
First place in computer science at the world's largest pre-college science fair, for detecting abnormal vascular flow in free-tissue transfers using speech-recognition techniques.
ai
AI Grant Fellow
October 2017 · awarded by Nat Friedman & Daniel Gross
Fellowship funding for research on detecting abnormal vascular flow in free-tissue transfers using audio recognition — machine listening for surgical outcomes.
06 · Toolbox
The instruments
Click a skill to light up the projects that use it.
AI / Deep Learning
Languages & Web
Data & Infrastructure
A site about accessibility should practice it.
Fully keyboard-navigable — try Tab, or ⌘K for the command menu.
Semantic landmarks and labels throughout for screen readers.
Honors prefers-reduced-motion — or toggle animations yourself with the ✦ button.
Light and dark themes, with contrast-checked colors in both.