Adam Sohn

A small corner of the web. Projects below, links to find me elsewhere at the bottom.

Writing and data visualization

CLAP, step by step: An interactive explainer of Contrastive Language-Audio Pretraining — how a sound and a phrase land in one 512-dimensional space, worked out one operation at a time with every audio value computed live in the browser.
Isolating the Engine Audio: Pulling clean engine and mechanical audio out of messy YouTube clips — filtering speech and music live in the browser on WebGPU to build labeled training data.
GRPO, step by step: An interactive explainer of Group Relative Policy Optimization — the RL objective behind modern reasoning models, walked through piece by piece with click-to-pronounce math.
Reliably Incorrect: Notes and visualizations on the agent capability threshold — when LLM agents stop being reliably wrong and start being occasionally right.
λ-bench variance: Five Sonnet 4.6 runs on the same LamBench task, classified by Opus 4.6, rendered as flame charts — what variance looks like when you stop averaging it away.

car-diagnosis: Diagnose a car fault from its sound — an honest, end-to-end audio-ML pipeline (scrape → clean → CLAP → calibrated triage).
intercept: Turn any website into a typed JSON API using self-improving agents.
agent-tuning: Using recursion to achieve predictable agent output.
alphadidactic: An iterative research agent: searches academic research, applies it to time-series data, and probes it to find novel discoveries.
HNswered: Notifies you when someone replies to your Hacker News posts and comments.
goldseam: Bring-your-own-model healing and authoring for Cypress.

Food: Cooking was once my profession and coding my hobby; now coding is my profession and cooking is my hobby.
Street food and BBQ: A photo gallery pinned to a spinning globe — street food stalls and barbecue joints, plotted where I ate them.

adamsohn.com