LinksMIT BeaverWorksDraperCollegeSensei
I started in competitive robotics in middle school, in Stoneham. In VEX a system either works on the field or it doesn't, and explaining does not change which one happened.
That got me into MIT BeaverWorks, and BeaverWorks got me a Draper internship, six of us out of four hundred. I expected a summer. I stayed fifteen months, two summers and the semesters between them, shipping code to people whose work depended on it.
The other half started with admissions. Guidance in this country is a market, and it serves whoever can pay. The students who most need someone to explain how any of it works are the ones who cannot afford that person, so I started CollegeSensei. Since then I have built a scholarship portal for undocumented students and a voice agent for people who cannot use a mouse.
I study economics because I want to know whether a thing is viable as well as whether it runs.
Linkgrid.harvard.edu
I am building robots that run at actually useful speeds and get better on the job. The project made the top fourteen of Harvard Grid's Applied AI Incubator.
I want to talk with operations leads in manufacturing, logistics, agriculture and anywhere else about the tasks where robots have been too slow to be worth it.
LinksFlashRTπ0.5 paperOur paper, soon
A robot policy is only as fast as the computer on the robot. FlashRT is the fastest open runtime we found for running π0.5 there: 27.17 milliseconds per inference on a Jetson Thor. Its own profiling says a large share of that time goes to starting GPU kernels rather than doing arithmetic, about 2,840 launches per inference, 2,340 of them in the ten-step denoising loop.
Three of us are removing that overhead in two steps. First, early launch: a CUDA feature that lets each kernel set up while the one before it finishes. Then a persistent megakernel that replaces the denoising loop with one resident kernel, working through every layer of every step as tiles from a fixed schedule, with no launches in between. The arithmetic does not change, so the outputs have to match FlashRT's. The target is 20 milliseconds.
The reason to want it is speed at the robot. A policy trained on demonstrations moves at the pace of the person who teleoperated them, and methods that play actions faster than the demonstrations reach three to four times the speed. At those speeds the delay between sensing and acting becomes the limit.
A real estate brokerage north of Boston had a website with no real estate on it. The search form threw away whatever you typed and sent you to the MLS. The homepage loaded three copies of Bootstrap, and there was well over a hundred megabytes of video on it. The brokerage works in English and Brazilian Portuguese, and the Portuguese half lived behind a query string.
They contracted me to rebuild it, and I did, from the database up. A background worker pulls the brokerage's own listings from the MLS feed, so the houses are on the site with their photos. A lead is assigned to an agent the minute it arrives. English and Portuguese each get their own URLs. Next.js and Postgres, in Docker behind Caddy, on one server.
The feed was the hard part. The data is licensed for the brokerage's own listings only, the standard dictionary describes fields this MLS does not carry, and one wrong field name makes the whole sync fetch nothing, quietly. I ran the mapping against the live metadata before I trusted a number on the page.
Private admissions counseling runs into the thousands of dollars, so the students it would help most are the ones who go without it. CollegeSensei connects them to expert mentors at a price that isn't absurd.
We passed a hundred users in the first quarter and partnered with five local schools and organizations. I raised over $10K pre-seed. I ran strategy, product and marketing with a team of ten volunteers and contractors, and I took students myself, across a dozen or so universities from Ivies to small liberal arts colleges, because I did not think I could build it well otherwise.
Good mentors are scarce and they have better-paying options. Each school partnership is its own negotiation with its own gatekeeper. The software was easier than any of that.
Linkdraper.com
GPS can be jammed, spoofed or simply unavailable. At Draper I worked on an alternative navigation system for the Air Force's positioning, navigation and timing capabilities, writing the operator interfaces in Python and C++. That software went out to clients who were not going to read my code to work out what I meant.
I also built a full-stack tool for the microelectronics teams that parses sensor data and does schema matching with LLMs. Matching two schemas is a judgment call, and engineers had been spending days on it. With a language model doing the judging it takes hours.
The summer intern project is the one I would show first. Eight of us built a mini lunar rover prototype sized to fit an Artemis payload, so the mass and volume budgets were fixed before anyone wrote a line of code and the decisions after that were negotiations against them. I led the robotics programming: sensing, autonomous navigation, drivetrain control and the state machine, in ROS2 and Python.
Then we defended it for two hours in front of forty domain experts and company leads. Nobody had taken my design reasoning apart that thoroughly before, and I have never learned as much in an afternoon.
Linkbigtincan.com
The first thing I put in front of real users, at Bigtincan, was a serverless document summarization tool on AWS. Clients upload internal documentation and get summaries back. It was small, and it shipped, and it had to keep running at three in the morning without me.
LinkAGI House
Onboarding flows assume a mouse. They assume you can see where the button is, click it, and read the error when you miss. For a lot of people none of that holds, and the workaround is asking someone to sit next to you and do it.
Florence is an MCP-enabled voice agent that runs an entire website by speech. It is not a chatbot bolted onto a page. You talk, and it navigates and drives the site. It was built for a disability tech startup, at an AGI House hackathon.
The second largest egg producer in the world runs a meaningful share of its production, inventory and distribution through spreadsheets. That is more common than people outside operations think. Spreadsheets scale well right up until they don't, and then they fail silently: a stale copy, a broken reference, a number nobody can trace.
I built the system to replace that, working directly with the CEO and the operations team. The aim was to take more than ten hours of manual work off them a week. The number I cared about more was how much of the process became legible to the people running it.
The technical work was the smaller part. The spreadsheet held a decade of decisions nobody wrote down, and I had to find every one of them before I could replace it.
I wrote autonomous drone racing software and entered solo against teams. It fuses camera and IMU data into a state estimate, and plans the racing line from that, fast enough to fly on.
I entered alone on purpose, so I had to do every part of it myself. It went through to the second round.
Linksdreamersroadmap.orgTech for Social Good
DREAMer's Roadmap helps first-generation, undocumented and DACA students find scholarships they are eligible for. That information is scattered, published inconsistently, and keeps changing.
I rebuilt their admin portal with Tech for Social Good: architected the backend, migrated a legacy Django schema to Prisma, and wrote the scraper that keeps the listings current. The stack is React, Next.js and FastAPI. The codebase had been written fast, under real deadlines, by whoever was available, and people were depending on it every day, so I had to improve it without breaking it.
LinkGitHub
A ComfyUI engine built so that every stage of image generation is its own step. You can change one stage without touching anything downstream of it.
I wrote up what each part does and why, so someone other than me can run it.
Ray-Ban Meta glasses stream what is in front of you into a facial recognition pipeline. It matches the faces in frame against what has already been published about those people on the open web, and gives it back while you are still standing there.
It works, which is the uncomfortable part. None of it needed privileged access. It runs on consumer hardware and an off-the-shelf recognition model, against sources that were already public. Latency was the hard engineering. The more interesting finding was that putting it together took an afternoon and no budget, and I would rather have found that out by building it than by reading about it later.
Let's talk.