The best way to prepare for an interview is to interview. With Carl Liu, I built a voice agent that interviews you, so you can practice whenever you want.
I benchmarked speech-to-text, text-to-speech, and LLM combinations to find the best balance of quality, latency, and cost. I also built the harness the agent runs on: an interviewer LLM paired with a validator LLM that keeps the interview on track.
With Prof. Jef Caers, I'm building an agent that helps the government of the Democratic Republic of Congo (DRC) navigate 505 mining contracts, 4,618 ministerial legal acts, a set of environmental and sector reports, and the 2018 Mining Code and Mining Regulations. The documents can't leave the country's own hardware, so every model runs locally.
I first built a 150-question benchmark to compare models on accuracy and speed, which led us to LFM-2.5-2.6B. I adapted the Pi agent harness and created custom tools for this small model to navigate the full corpus. It now answers ~60% of the benchmark correctly. I am now fine-tuning the model aiming to get it to ~90% accuracy.
A paper with Prof. Hari Subramonyam that reviews the impact of AI on tasks at different levels of Bloom's taxonomy. Our hypothesis: AI boosts learning on creative tasks, but hurts it on typical school tasks like summarizing or analyzing a system.
A chess tutor built entirely on small local models. I fine-tuned LFM2.5-Encoder-230M to extract chess moves from natural-language explanations, and I'm now fine-tuning an LLM to be the tutor itself.
I trained Qwen2.5-Coder-3B to turn a plain-language brief into Manim code, the library behind 3Blue1Brown's videos. Out of the box, only 2.5% of its code rendered into a video.
I built the instruction dataset, then generated examples with gemini-3-flash to improve the model manim generation via SFT, then ran DPO on preference pairs sampled from the SFT model. Together, SFT and DPO raised the compile rate 10×, to 25%.
I collected close-up and wide-area satellite imagery for every high school in Colombia, fixing wrong official coordinates along the way. Then I fine-tuned two vision transformers, combined by a stacking model, to predict whether a school scores above average on Colombia's national exam. It's right 75% of the time, and 66% when comparing only schools with similar socioeconomic backgrounds.
Can a model's internals tell you it's about to be wrong? With Carl Liu, I quizzed Liquid AI's LFM2-1.2B on TriviaQA, recorded its activations and token probabilities for every answer, and trained classifiers that predict from those signals alone whether the answer is a hallucination. At the end token probabilities was the most accurate one.
My mom wanted to practice English, I wanted to practice Japanese, and OpenAI had just released its first realtime voice model. So I built a tutor you learn languages by talking.
It launched as the #1 education product of the week on Product Hunt and won its first paying subscribers. I paused the project when I started at Stanford. It taught me how to ship a consumer product, listen to users, and build software that actually teaches.
Right after my undergrad, just as ChatGPT came out, I co-founded Educo and led everything technical. We shipped generative AI tools for schools, from a lesson-planning assistant for teachers to a WhatsApp tutor for students that gained early traction.
We raised $100k from Platanus Ventures, a Chilean accelerator that backs ~1% of applicants. After a year and a half I moved on, with a hard-earned lesson in selling to schools, where the buyer (the principal) isn't the user (teachers and students).
I taught math in German to 11-year-olds and 16–17-year-olds. It taught me empathy, and how to hold the attention of an audience that will do anything but listen. I loved it.
Alongside my CS degree, I co-founded Techdalo. We raised $15k to fund technical courses for young Colombians, then helped them land jobs afterwards. I led an international team of 11 volunteers: 3 in Colombia and 8 in Germany.