Projects
An archive of things I’ve built, am building, want to build, or stopped building.
My driving interests are model compression (quantization, pruning, distillation) and on-device AI deployment: squeezing capable models onto hardware with limited compute, especially where privacy, latency, or cost rule out the cloud. If something here overlaps with what you’re working on, or you have a half-baked idea in the same space, reach out. I’m always up for comparing notes or collaborating.
Small models as simulated patients
Finding the smallest AI that can realistically play a patient, so training simulators work fully offline.
Compressing local coding agents
Every team rents coding agents by the token and ships code to someone else's servers. What if the agents lived on-prem instead?
Master's thesis: post-training quantization
Picking AI models today means trusting bar charts with no context. This thesis builds the understanding to choose well.
Whisper encoder layer pruning
Cutting whole layers out of a top speech recognition model, then getting the performance back without labeled data.
On-device transcription for product
Some products cannot afford cloud AI per use but still need the feature. On-device is the only math that works.
In-browser offline transcription
Open a web page, get transcription, nothing ever leaves your device. Built in days, shared widely.
MLOps mentor
Every one of 450 students deserves feedback while they work. TAs cannot read 450 repos, so the reading got automated.
Whisper quantization sweep
The obvious first try at shrinking speech recognition: smaller numbers, sparser weights. It mostly failed, instructively.