From Olympiad math to building AI startups.

Tablet-based learning programs in low-resource schools have real results. Randomized trials show months of additional learning gain at under $7 per child per year. They all share the same flaw: they deliver content, not conversation. When a child is stuck or losing confidence, the tablet cannot slow down, reassure, or adapt.

A fully offline, voice-based tutoring system running on a tablet. A small language model runs on-device and controls a digital whiteboard in real time, displaying visuals and lesson content while it speaks, listens, and adjusts. The first module taught counting from one to ten in both directions, with the tutor guiding each child at their own pace.

Version one worked: the tutor introduced itself, guided the child, and wrote each number on the whiteboard. Version two was field-tested in a rural school in Pakistan, and the excitement faded fast. The tutor misheard unclear pronunciation and kept asking children to repeat themselves. The voice was robotic, the pacing too fast, and students got frustrated, which was the exact outcome the tool existed to prevent. Version three was rebuilt around teacher feedback: relationship before content, slower pacing, a softer voice, and a shift from “instruction” to “emotionally attuned learning partner”.

Personalization is not difficulty level. It is pronunciation, comprehension speed, attention span, and the emotional state of a child who has already decided they are bad at math. The tool had to feel like someone who believed in them.

Presented at the T522 Impact by Design showcase at Harvard GSE. Met with UNICEF’s Learning Passport team in New York, who had been working toward a similar solution and had not yet cracked the offline, conversational layer.

A team of twelve teachers and staff. Partnerships with 20+ schools. More than 1,000 learners across 15+ countries, mostly students in the US and Canada, plus international schools in Hong Kong, Singapore, China, and Europe. Bootstrapped from day one, profitable, growing around 30% a year.

No ads. Distribution ran on trust: partnering with education consultants and parents, offering pilots instead of pitches, and converting pilots into long-term school contracts. I owned product, pricing, and distribution end to end.

How to operate when failure is not an option, and that the teaching method mattered more than anything else we did. Problem-first, question-driven, no judgement on wrong answers. The students who stayed longest were the ones who had been told elsewhere that they were not math people.

In-service teachers are time-poor, often skeptical of training, and working in wildly different contexts: public schools in the US, exam-prep centers in China, national curricula in Kyrgyzstan. Existing professional development is theory-first and designed for a classroom that does not exist.

A full redesign of Harvard GSE’s foundational How People Learn course for in-service teachers, centered on an AI mentor. The bot is grounded in 50+ authentic classroom scenarios and the core principles of learning science: motivation, prior knowledge, feedback, transfer, metacognition. A teacher brings a real challenge; the bot retrieves similar cases and responds with reasoning explained in learning-science terms, making the invisible logic of expert teaching visible. Four modules, backward-designed, deployed in Canvas.

Teachers had to critique and adapt the bot’s suggestions, not accept them. Intelligence augmentation, not substitution.

Adopted by Harvard Graduate School of Education as a live product, now used in the online module for training teachers. Built with Katrina Kim, Taylor Yao, and Shutong Liu.

Every employee gets a link each morning and has a short adaptive conversation, not a form. The system extracts six behavioral signals from what they say: ownership, specificity, energy, blockers by source and severity, risk flags, and the gap between the job on paper and the job in practice. The manager opens one digest at 7am: who is on track, who is blocked, what needs attention today. Skips are signal too.

Every conversation is data no other tool is collecting. Over time it becomes a queryable company memory: institutional knowledge that used to walk out the door when people left, captured. The record compounds with every check-in.

Designed the product architecture, the signal extraction framework, the check-in conversation engine, and the manager digest. Led customer discovery, pilot scoping, the investor narrative, and cross-functional operations with the founding team.

Built on Dr. David Dockterman’s cognitive partnering research at Harvard GSE. Most tools measure what you produce. AIQ measures how you think: foundations (what you understand about AI), behaviors (how you work with it day to day), and disposition (how you reflect on your own use). The framework separates using AI as a shortcut from using it as a thinking partner.

The full scoring framework, the K/B/D axis model, normalized scoring, and the archetype assignment system. All 20 questions and all archetype copy. A mobile-first web app: no login, auto-advances on answer, email capture before results, a shareable badge generated on the spot. Loads in under two seconds.

The Orchestrator · The Accelerator · The Skeptic · The Explorer · The Automator · The Newcomer

A voice-based AI mentor for students heading into international education, humanitarian work, and global development. Two modes: a Socratic phase that surfaces what the student values and where they want to go, then a tactical phase with concrete networking strategies, realistic entry points, and next steps. Every session ends with two to four action items.

Testing with Harvard students showed voice produced far higher engagement and completion than text. A calm conversational voice made students open up and stay. That finding drove every iteration after.

Getting an AI to behave like a trusted mentor instead of a generic coach: knowing when to stop asking and start advising, handling vague or discouraged students without being pushy, and delivering practical clarity no matter how the conversation started. I designed the full prompt architecture encoding Prof. Jenkins’s frameworks, ran multiple testing rounds, and iterated prompt and voice configuration each time.

Selected as a Student Partner in the LEAP Bot Club. Presented at the HGSE Bot Club Showcase, May 2026.

Ask “Is Apple a good long-term investment?” and instead of a confident hallucination you get an answer built only from retrieved SEC 10-K filings, analyst recommendations, and live market data, with every source shown and relevance-scored.

FastAPI backend. FAISS and ChromaDB vector search over sentence-transformer embeddings. GPT-4 generates from retrieved documents plus live Yahoo Finance data. Streamlit front end with expandable source panels and a clear indicator of RAG versus fallback mode.

A comparative study of four models (neural network, Random Forest, AdaBoost, XGBoost) on 2,913 confirmed exoplanets from the Planetary Habitability Laboratory catalogue. Built the data pipeline from scratch: 86 raw planetary and stellar features reduced to 11 reliable observable ones, incomplete records removed, 70/30 train-test split.

XGBoost reached 99.89% accuracy, misclassifying one planet in the test set. More useful than the accuracy was what the model exposed: planet radius alone carried over 50% of predictive weight, and radius plus solar insolation together carried 68%. Habitable planets cluster tightly between 0.8 and 2.8 Earth radii and 0.2 to 1.6 Earth flux. That gives future surveys two measurements to prioritize.