Rocketable (YC W25) acquires profitable software products and builds the AI agent systems to operate them. I worked three months on the customer experience and support systems of one of the company's products. The work covered prompt engineering and LLM evaluation, agent development across OpenAI and Anthropic models, AI cost analysis, and customer conversation analysis at scale.
View projects →
Contracted by a physician partnering with Simply Balanced, a movement studio in San Francisco, to build the practice's first patient-facing app as sole designer and developer, translating her clinical protocol into product specs, interface design, and production code. The app administers and scores a 9-test battery covering balance, memory, and reaction, used across cognitive and motor conditions, from Parkinson's and age-related cognitive decline to pediatric ADHD and post-concussion recovery. Next comes the prescription engine, which will convert those scores into the movement and exercise plan she currently writes by hand, encoding her decision rules, tuned against the doctor's past prescriptions. A first beta with a youth cohort launches December 2026, testing how younger patients interact with the app and gathering cases to sharpen the engine.
Project manager for a five-person ML team predicting walking movement from brain signals, setting the roadmap, assigning work, and managing delivery timelines. A hybrid TCN-LSTM model predicted gait from 64-channel EEG alone, separating true heel-strikes from false ones with 0.924 AUC and accounting for 74% of the variation in ankle angle. Presented at the California Neurotech Conference in April 2026, the research targets prosthetic control for patients with spinal cord injuries.
View Poster →A personal assistant built on a cloud-hosted Hermes agent, running as four scoped workspaces in Slack. Each workspace has its own system prompt, skills, permissions, and working directory, compiled from a single config file. The agent's full definition lives in a git repo that it pulls daily and applies to itself. Obligations mentioned in chat become reminders and calendar events, and the OAuth scope limits its writes to calendars it created itself.
Automates picking the AI news worth reading from a set of RSS sources. A crawler pulls new items daily, a Claude summarizer writes one or two sentences on each, and a Curator agent ranks them and assigns sections. A second pass cuts the shortlist to twelve to fifteen items, recording a reason for each cut. Picks matched hand-picked selections 86% of the time on a 25-item validation test.
Automates the daily process of checking six job boards by hand. Scrapes YC, Simplify, Ashby, Greenhouse, Workday, and company career pages, filtering each posting against a defined set of role criteria. Matching roles are added to a Notion database and sent in a weekly email with deadline reminders for programs on fixed application windows.
View on GitHub →
Mirror Progress is a small team building AI data infrastructure, including searchable knowledge layers and document retrieval, for firms in architecture, engineering, and professional services. Built an account map for a target enterprise, profiling and scoring stakeholders on title, seniority, and history with comparable work, then identifying a subsidiary as the lower-friction entry point into the parent company. The founder adopted the approach for enterprise outreach.
A short documentary from junior year of high school, interviewing family and friends about what it's like to grow up speaking more than one language. Chacun avait une réponse différente.
Watch on YouTube →