Work

Everything I've built

The full archive — shipped products, experiments, and the small things in between, newest work first.

Featured

dún firmware

Nov 2026

The firmware for the dún earbud, written as an ear-fatigue tracker. Save a baseline on a quiet day; after a gig, press once and it checks your ear's echoes at six pitches against it, then answers steady, dip or retest over the light, USB and Bluetooth. Each speaker plays only one tone so the earbud can never fake the ear's note, every tone sits on an exact FFT bin of one shared clock so blocks average cleanly, calibration starts quiet and every tone is capped at 70 dB SPL, and a loose fit or a noisy room gives retest rather than a false dip. The test itself is hardware-free C99 that runs against a simulated ear: no false dips on 300 reseated ears, 92% of 8 dB drops caught. The Zephyr app builds in CI and has not run on a board yet. A screening experiment, not a medical device.

Zephyr RTOS
nRF52840
Embedded C
Audio DSP

sotto

Sep 2026

A voice that has to sound unhurried has a problem: every text-to-speech speed control slows the words themselves, which sounds like a tape running slow. So leave the speed alone. Render the sentence once at its natural rate, then find the silences the model already produced and lengthen only those, in proportion to how long each one already was. Every register in the ladder runs at speed 1.00 and still lands on its target words per second, 2.33 down to 1.25, because the silence does all the work. Getting there meant learning that a stop consonant contains a real 40 to 120 millisecond gap of nothing, which a naive pause detector cannot tell from a breath, and on one line it stretched a 60 millisecond consonant closure by 3.1 seconds, inside the word. Around it sits the loop: the model streamed token by token and cut into sentences as they complete, so audio starts after the first one rather than the whole reply; a pre-screened holding line if that first sentence misses its budget; and barge-in that stops the audio but keeps reading the stream, because letting it go marks the wrong line as the interrupted one. Rendering is capped at two slots that queue rather than shed, on the principle that a late sentence is still spoken and a refused one never is.

Python
FastAPI
Kokoro TTS
Self-hosted

mote

Aug 2026

A complete transformer inference engine in C with nothing underneath it. No PyTorch, no NumPy, no BLAS, so RMSNorm, rotary embeddings, grouped-query attention with a KV cache, SwiGLU and sampling are all written out by hand and every shortcut is closed. Quantization is where it got interesting. The obvious improvement, searching for a slightly clipped scale for each group of weights, made the mathematical error smaller and the model worse, because the outliers it clipped were exactly the ones that mattered; plain rounding won. int4 turns out to decode slower than int8, since unpacking two weights from a byte costs arithmetic you do not get back, and it only wins on disk. And a gently trained behaviour can be smaller than 4-bit rounding noise and simply vanish, so training has to round bit for bit the way the engine does. A converter that rounds slightly better than training is rounding differently, and different is broken. Under 90 KB of WebAssembly, a real chat model offline on a phone, a $6 microcontroller next.

Embedded C
Transformer
Quantization
Open Source

ràdá

Aug 2026

A pocket 60.5 GHz radar that measures in microns, by reading the phase of the returning wave rather than its timing: one degree is 6.9 micrometres of movement, about a tenth of the width of a hair. Two things come out of that. Count a heartbeat across still air, and recover sound from a surface the device never touches. The whole architecture then hinges on one number nobody can calculate honestly in advance, which is how fast the link to the pre-certified module really sustains, because buying the module rather than the bare die costs three to four orders of magnitude of bandwidth and the audio you can recover is only ever half the sweep rate. So the sensor is made to answer for itself: every frame carries a flag meaning I could not deliver that one on time, and the first program written for the project does nothing but push the rate up until that flag starts firing. The board passes its electrical checks with zero errors and its traces are deliberately not laid yet, because the answer decides whether they get routed for the easy job or the hard one. My first hardware project, and the line between proven and assumed is drawn on purpose.

60.5 GHz radar
nRF52840
Embedded C
KiCad

àyò

Jul 2026

A research platform for systematic trading, arranged around one refusal: the language model is never allowed to emit code. Claude researches and writes each strategy as structured data, a content-hashed artifact, and a deterministic engine trades only the frozen artifact. So you validate the artifact rather than the model, a non-deterministic author yields perfectly reproducible behaviour, and if you removed Claude entirely the thing would keep trading safely on its frozen rules. It would just stop getting smarter. Nothing reaches the market without clearing an out-of-sample gate that raises its own bar as more ideas are tried, and that gate is the point: an in-sample Sharpe of 3.72 became minus 6.99 out of sample, and a strategy that passed at 1.39 was still rejected for a 65% drawdown. On efficient large-cap markets it has so far turned away every simple strategy put to it, which is the correct result, so it trades paper until something survives. The same discipline points inward. In July the scoreboard was caught flattering itself, with after-hours benchmark prices up to 3% stale quietly overstating the book for the entire live window. It was found by auditing a number that looked too good instead of enjoying it, and the smaller honest figure replaced it the same day.

Python
Claude
Alpaca
pandas

dimeji

Jul 2026

A schema and a discipline, given a renderer. It lays an emotional history out as a typed graph, and the rigour is entirely in what the types refuse to let you write down rather than in anything the machine concludes. Thirteen node kinds, six levels of evidence from observed through reported to hypothesis, and edges that either support or contradict, so a claim cannot be recorded without declaring how it is known. A hypothesis is never promoted to a fact, competing explanations sit side by side with no winner, every tension lists unknown among its valid readings, and softer provenance is drawn with a dashed border so you can see the uncertainty before you click. There is no inference here and it does not pretend to have any: nothing is scored, nothing propagates, and every judgement is authored by a person. The only real computation is the layout, a force simulation run to convergence once and then frozen, with the repulsion capped so 200 nodes stay framable on a phone.

Framework
React Flow
TypeScript
Epistemics

vibe

Jul 2026

A Claude Code skill that hands the assistant a whole persona, an accent, an attitude and an energy that all correspond, and holds it in character for an entire session while the engineering stays senior grade. The code, the comments and the commits stay clean; only the conversation takes on the voice. Five deeply rooted voices written as love letters rather than caricatures, an invent-your-own mode, and a rule set for staying human instead of sounding like an AI in a costume. One Markdown file, no dependencies, open source.

Claude Code
Skill
Open Source
Markdown

the scribe

LIVE SITE ↗Jun 2026

A ghostwriting studio for ministry authors, where the interesting part is how an agent is stopped from inventing. Every margin note the editor writes has to carry an anchor quote, and the anchor is checked as a verbatim substring of the manuscript before the note is allowed to exist. What makes it work is where the rejection goes: rather than discarding a failed note, the checker hands the model back the string anchor_quote does not appear verbatim in the chapter, copy the exact text and retry as the tool result, so it corrects itself inside the same loop with no orchestration code at all. The author's voice is a 45-field record built by an interview that merges monotonically: identity fields are write-once and the weights only ratchet up, so contradicting yourself at question seven cannot erase question one. Fifteen end-to-end evals gate the pipeline on real tokens. The verbatim gate proves a note is anchored to real text; it does not prove the model's claim about which sermon that text came from, and the drafting pass itself is governed by prompt rules rather than a verifier.

TypeScript
Next.js
OpenAI
Postgres
Prisma

Experiments

dún

Nov 2026

A healthy cochlea answers a click with a far fainter sound of its own, a few milliseconds later, and clinic screeners that listen for it cost thousands. A University of Washington team showed an earbud can do it for about $28. dún rebuilds theirs from the paper: a XIAO nRF52840 so there is no radio to design, two amplifiers and two speakers because one cheap speaker playing two tones fakes the very note the ear should add, and two microphones, one for the ear and one for the room. Both boards are routed and the schematics generated to match. Nothing is ordered yet, and the first test is a sealed plastic tube that should answer nothing. An experiment, not a medical device.

Otoacoustic emissions
nRF52840
Audio DSP
KiCad

rondo

Nov 2026

A rondo is Barcelona's keep-ball drill, and the passes before the last one look like nothing on a scoresheet. A two-layer net, written in NumPy and trained on 985,628 touches from 524 Barcelona league matches in StatsBomb's open data, gives each touch a chance that the possession ends in a goal. The 2010/11 season was held out and the net never saw it. On the 5-0 against Real Madrid, Xavi's opener sits under 5% for almost sixty touches until one Iniesta ball sends it to 32%, Villa's two goals are both Messi through balls, and Madrid's best moment all night reaches 13%. It is calibrated on the unseen season but only barely beats a logistic regression, and it sees only the ball, because 2010 event data says nothing about where anyone else stood.

Neural networks
Python
NumPy
StatsBomb

gbó

Nov 2026

A model that watches an F1 stint the way a pit wall does and calls the stop. Every green-flag lap of the 2025 season from the public timing feed, 20,669 of them over 22 dry races, is corrected for fuel and measured against the stint's best so far. The tyres mostly fade rather than fall: a soft loses about seven tenths in ten laps and keeps going, and two thirds of stints run a second off their best before they stop. A gradient-boosted model, using only what was known at each lap and tested on races it never saw, forecasts the loss ten laps ahead to 0.50 s, against 0.58 for assuming nothing changes. Replayed against all 698 stops, a quarter of which came under a safety car, it wants to stop a median three laps before the teams at a 3 s fuel effect; at 2 s the teams look nine laps late and at 4 s a lap early, because inside a stint fuel and tyre age move together and lap times cannot split them.

Machine learning
Python
scikit-learn
FastF1 timing

rán

Nov 2026

Every 2025 qualifying session, its fastest pieces stitched into one lap and raced as a ghost against the pole lap on the track map. From the three official sectors, timed by loops in the track, the stitched lap beats pole by a median 0.16 s, and by nothing in Austria and Mexico. Cut the lap finer from telemetry and it claims 1.7 s at 25 pieces, because telemetry rebuilds an official sector only to a median 40 to 50 ms and the minimum of many noisy pieces collects the error. In 9 of 24 sessions another driver's own best sectors beat the pole lap; five of those were Verstappen poles, out of his eight.

Telemetry
Signal processing
Python
NumPy

aas

Nov 2026

Show me your friends and I'll tell you who you are, asked of the data a person leaves behind. Round one reads one part of a life at a time against each person's own personality test. Forum posts read thinking against feeling at AUC 0.83 after every type word is removed (the type words alone reach 0.90). Phones and Fitbits both hint at extraversion near r 0.24 but with 50 people neither can settle it, and Steam profiles carry nothing. v1 reads a person's own exports on their machine and states how accurate it is for their kind of writing, which on tweets is close to chance. It started with listening histories, where whether a new artist will stick reads from the first session at AUC 0.792. Round three looks for statements true of one listener and rare in others. On 333 new listeners, checked once under a rule set in advance, three of them pick the right person out of five 67.5% of the time, against one in five by chance.

Python
Statistics
Personality

vouch

hackathonBLI Legal Tech, Oct 2026Oct 2026

AI now drafts investment research, and it cites filings the way it cites court cases: mostly right, sometimes invented. Each claim is posted on-chain as one figure, the filing sentence behind it and a stake. A deterministic check, run through a Chainlink CRE workflow against SEC EDGAR, decides it: the sentence must be in the filing and every figure, period and direction must be as filed. A wrong claim loses its stake in the same transaction as the verdict. Change 50% to 56% in a real Tesla 10-Q sentence and it is still 99.3% similar to the filing; in one round all ten doctored quotes scored between 99.1% and 99.7%, so a 95% fuzzy match would have passed every one. The figures rule slashed all ten.

Solidity
Chainlink CRE
Python
SEC EDGAR

kickflip

Oct 2026

Tesla rolled a Roadster over in mid-air on compressed air, and gave only a few numbers: ten nozzles, two tanks holding 600 L and about 250 kg of air, 80 kN of thrust. From those alone, an engine concept, drawn to scale: four rear nozzles to launch, four under the corners to lift, land and roll, two at the front for yaw, all fed at tank pressure down one trunk with a regulator beside each group, and 6.7 cm-throat cones that stop just short of frosting the air. Checking it works: the air sits at 437 bar, a quarter above what an ideal gas would need, and the tanks and air together weigh a fifth to a quarter of the car. A real-gas model of the blowdown shows the catch: the gas left in the tank cools towards −120 °C as it empties, so a regulated system holds 80 kN for about a second and strands 74 kg of air. The roll itself is cheap, around 10 kN a side and an eighth of the tank for a half turn in a second; launch, hover and the landing burn are what cost. One run uses roughly all of a cold tank, and a refill takes at least 36 kWh.

Mechanical engineering
Thermodynamics
Python
CoolProp

mpemba

Oct 2026

Does hot water freeze before cold? In ordinary water the moment ice forms is random, so the useful question is how often the hot cup wins. If the containers are interchangeable and heating leaves nothing behind, the answer is at most half the time, whatever the cooling law or the physics of nucleation. Under Klimov and Finkelstein's nucleation model there is also a floor, set by the ratio of nucleation rate to cooling rate, and with identical containers the effect is visible only within about half a kelvin of freezer temperature. Realistic differences between containers stretch that across several kelvin. Every published hot-first rate above one half comes from containers that were not interchangeable; Burridge and Hallstadius's sanded beaker froze first in 18 of 20 pairings whatever the starting temperatures. One setup in Janni et al. (tins on a metal sheet) beats a no-memory simulation and may point to memory in the container's contact with the shelf. The paper is in draft for the American Journal of Physics.

Physics
Probability
Python
NumPy

thrum

Oct 2026

A string lab in the terminal, played with your hands through a webcam and drawn entirely in characters. Pinch the air to make a particle and flick to throw it; pinch with both hands and pull to stretch a string, tuned by its length; hold a particle to charge it until it bursts where it lands. Two things had to be measured rather than assumed. MediaPipe’s metric landmarks put touching fingertips 6 to 8 cm apart, so the pinch is read in the image instead, where pinched and open hands sit at 0.10 to 0.19 and 0.85 to 1.45 of a palm. And a throw leaves at your hand’s real speed, so gravity has to be real too: the scale comes from your palm in the camera, and a particle now leaves within 3% of the hand’s speed.

MediaPipe
Python
Terminal
NumPy synth

unsaid

hackathonAssemblyAI Voice Agent, Sep 2026Sep 2026

Every transcription tool records a turn that was talked over as a finished turn. This reads a live multi-speaker stream from AssemblyAI and reports the ones that did not survive: disrupted, where somebody came in within 0.3 seconds and the thought died, and abandoned, where the speaker stopped mid-thought. Then it acts: it waits until speaking would not itself be an interruption and offers the floor back, quoting the words that were lost. It never speaks over anybody, because a tool about people being talked over cannot talk over people. Measured on 200 blind hand-labelled turns from the AMI meeting corpus: precision 0.74, recall 0.48 at the shipped threshold. It misses about half of what is there.

Python
AssemblyAI
WebSockets
AMI corpus

ika

Sep 2026

Shadowbox in front of a laptop; it finds your habit and says it out loud a beat before you do it. Not “you threw a jab” but “when you finish on a right, your left hand comes down”. Timing is the whole problem: a conventional pipeline recognises an action 250 ms after it ends, and the next begins 263 ms after that. The first reader measured a palm growing on its way to the lens, the right physics at arm’s length and the wrong physics standing back, where it caught 1 punch in 1,448 on real footage. The finished drill reads punches from the arm and the guard from its height after each combo, scored on fifteen public rounds held out by person: 47% of punches, and a phantom habit in 3% of shuffled rounds. The habit and the call are measured in simulation at those error rates, and it says so.

MediaPipe
PyTorch
Python
On-device

node

Aug 2026

An intelligence that lives in the building rather than a browser tab, and the machine specified to hold it. The want is continuity: one system with several bodies that has been indexing my work rather than being handed a summary of it, so that what we decided last week is retrieved instead of reinvented. Most of it never reaches a large model. Around 80% of requests are a nearest-neighbour match against an embedded command registry, another 15% are a single tool call whose reliability comes from masking generation to a grammar rather than from trusting the model, and only the rest escalates. But every tier has to be resident at once or the whole thing is just an app with extra steps, which is a problem when everything I run today time-shares one 12 GB card. So: four 96 GB Blackwell cards in one enclosure, specified from part numbers, in two purchase phases, with the power ledger and the places the platform genuinely cannot match a datacentre machine written down rather than glossed.

Threadripper PRO
384 GB VRAM
Local inference
Hardware

mak

Aug 2026

The single most common operation inside mote, the from-scratch C inference engine, is multiply an activation by a weight and add it to a running total. A whole neural network is mostly that, done billions of times. mak is that operation built as actual hardware, a signed int8 multiply-accumulate lane taken from Verilog all the way down to a SkyWater 130nm silicon layout on an open flow. Eight int8 lanes multiplied at once, summed through an adder tree, and accumulated across cycles into a 32-bit register. Verified against a software reference across forty dot products, synthesized to roughly 4,500 cells, then placed and routed to a GDSII, the same file a foundry is handed. Designed and laid out, not fabbed, one tile rather than an accelerator, and honest about being exactly that: the atomic operation of something I built, carried down to the layer where it becomes a physical object.

Verilog · RTL
Sky130
OpenLane
Silicon

jùwọ̀n

Aug 2026

How do you measure the error rate of a change detector without labelling anything? Point it at a desert. Two cloud-free dates over ground that has not moved in a century, and whatever area it reports as changed is, by construction, its own false-positive rate. That test started at 7% on the Aswan desert and 38.9% on Lagos harbour, and it found four real bugs on the way down to 0.23% and 1.99%. The best one: Otsu's method has no absolute floor, so a relative threshold always splits something, and handed pure sensor noise it will confidently split that too. The next best: the engine reported the sensor's 10 metres per pixel while quietly resampling every area into a fixed 256 pixel tile, so a 6 km box was really 26 metres a pixel and every square-metre figure downstream of it was wrong. Run the same harness in reverse, painting a structure of known size into real imagery, and it finds 60 metres and misses 30, which is physics rather than a bug. On radar a whole-scene brightness threshold found zero vessels, because land backscatter swamps the statistics; taking them from the water class alone found 42. Free Sentinel optical and SAR, no API keys, no GPU, and a detector that is still a brightness baseline which has never seen a human label. It says so itself, before you have to ask.

Sentinel-2 · SAR
NumPy
Python
Calibration

Archive

atom

LIVE SITE ↗Jun 2026

A command-line tool that quantizes AI models so they run in far less GPU working memory (VRAM). It reads the GPU and picks the method for it: NVFP4 or INT4 via Nunchaku on modern NVIDIA cards, NF4 via bitsandbytes as a universal fallback. Every run is validated with real before/after numbers for VRAM, speed and output similarity, and it runs entirely locally, so models and data never leave the machine.

CUDA
Nunchaku
bitsandbytes

orator

Jun 2026

An AI public-speaking coach with real-time conversational voice agents (ElevenLabs) and an LLM feedback pass that scores delivery and coaches each rehearsal. Credit-based billing, Google auth, the works.

Next.js
FastAPI
ElevenLabs

greenlight

Jun 2026

Agentic prior-authorization for healthcare payers: LLM agents verify medical necessity with paragraph-level citations, gated by a deterministic citation verifier that blocks fabricated quotes outright — agents you can actually trust.

Python
LangGraph
Anthropic

glean

Jun 2026

An AI editing studio that turns a long recording into a cut master plus captioned vertical clips — Claude tool-use proposes the cut list, ffmpeg renders it. Background jobs, transcription, and storage wired end to end.

FastAPI
Claude
ffmpeg

nūr

May 2026

A shipped, premium AI Qur'an companion. A SwiftUI iOS app on a FastAPI/Postgres backend with Claude-powered reflections behind a safety + eval layer, geolocated prayer times, multi-reciter recitation audio, subscriptions, home-screen widgets, and an automated, personalized retention/notifications engine.

SwiftUI
FastAPI
Postgres
Claude

whisk

May 2026

An "aesthetic" AI recipe app where a single codebase ships as six niche App Store apps via an env switch. SwiftUI (WidgetKit, Live Activities, StoreKit) on a FastAPI/Postgres backend, with Claude recipe generation and fridge-photo vision.

SwiftUI
FastAPI
Claude

medula

May 2026

A multi-agent AI platform — a Next.js studio over a distributed microservice backend (API gateway, agent engine, context engine), with Postgres and Qdrant powering long-term vector memory.

Next.js
FastAPI
Qdrant

folio

LIVE SITE ↗Apr 2026

A decision layer that sits on top of your product analytics. Connect PostHog or the Folio SDK, describe a growth goal in plain English, and Folio drafts the in-app nudges — banner, modal, and dashboard copy with CTAs — that you approve before they ship, then serves them through a headless decisions API with serve→click→convert attribution on every recommendation.

TypeScript
Next.js
Supabase
React

nuj

Mar 2026

A relationship-wellness app for couples — AI daily moments, quizzes, check-ins, and shared journaling. A native Swift app (widget + share extensions) on a FastAPI/Supabase backend with pgvector and push.

SwiftUI
FastAPI
Supabase

lumi

Oct 2025

An AI voice-companion iOS app — real-time conversational agents (ElevenLabs) with a lock-screen Live Activity, on a FastAPI backend.

SwiftUI
ElevenLabs
FastAPI

netforge

GITHUB2024

A lightweight, cross-platform networking abstraction layer designed to simplify HTTP networking operations across iOS, macOS, and Linux.

Swift
SwiftNIO
SPM

notely

2023

A podcasting app in SwiftUI: record, and move the audio between Apple devices without losing it. Early work, and plainly that rather than dressed up as more.

Swift
SwiftUI
AVFoundation

veil

A competitive-intelligence OSINT platform: large-scale, resilient web scraping on Bright Data feeding structured, queryable intelligence — proxies, layout-drift tolerance, dedup, and freshness handled at scale.

Python
Bright Data
Postgres

hydrate

A health-focused iOS and watchOS application that helps users track their daily water intake in correlation with their health metrics.

Swift
HealthKit
CoreML
WatchKit

pixels

Advanced image processing iOS application leveraging CoreImage and Metal for high-fidelity editing capabilities. Features real-time photo synchronization across Apple devices using CloudKit, delivering exceptional performance for complex visual tasks.

Swift
Metal
CoreImage

itype

Smart keyboard utilizing transformer-based language models and federated learning for privacy-preserving text predictions. Implemented efficient on-device inference with model quantization.

Swift
Python
CoreML

kept journal

AI-powered journaling app featuring sentiment analysis, topic modeling, and personalized content recommendations using fine-tuned BERT models optimized for iOS.

Swift
Python
TensorFlow

kagaku

Content recommendation engine using collaborative filtering and neural network-based deep learning models, with custom loss functions for improved accuracy

Python
PyTorch
FastAPI

docgen

Command-line tool that automates comprehensive documentation generation through static code analysis. Features intelligent source code parsing, custom template engine for flexible documentation formats, and support for multiple programming languages with extensible architecture.

Python
AST
Jinja2

annotate

Advanced iPad note-taking platform leveraging machine learning for real-time handwriting analysis and semantic content organization. Implements on-device natural language processing for intelligent note categorization, featuring contextual learning algorithms for personalized content suggestions and study pattern optimization.

Swift
CoreML
PencilKit
NLP

lab controller

Sophisticated laboratory control system utilizing Bluetooth Low Energy (BLE) for precise Arduino microcontroller management. Features real-time bidirectional communication protocol, custom GATT service implementation, and hardware abstraction layer for seamless iOS-Arduino integration in laboratory environments.

Swift
CoreBluetooth
Arduino
C++

menudancer

macOS menu bar application featuring real-time audio visualization through an animated character. Implements low-latency audio processing using AVFoundation, with custom Core Animation choreography synced to frequency spectrum analysis of microphone input.

Swift
AVFoundation
CoreAnimation