About
From mathematics to machine intelligence
I am an engineer and a mathematician. I have spent more than twenty years building systems that learn from data, as a consultant, a startup co-founder, a director of machine learning in a large company, and a prolific open-source developer.
Today I work on putting AI to work: setting up AI agents around the way people and teams really work (see the home page), and leading technology for startups as a fractional CTO.
Open source
An active developer, in the open
I have been writing reusable tools for well over a decade, and I publish them. As of November 2026 that is 290 Python packages on PyPI, with more than 4,000 releases in the past twelve months alone, and 73 JavaScript and TypeScript packages on npm, all published this year. Rust crates on crates.io are next. Most of them start the same way: I solve something for a project, see that I will need it again, and pull it out into its own package.
Python: the enabling layer, AI integration, sound and image
The packages I care about most are the enabling layer, the small tools everything else stands on. dol gives one dict-like interface to any storage: your code reads and writes store[key] and does not care whether the bytes live in files, S3, MongoDB or a vector database, and more than twenty *dol packages plug specific backends into it. meshed composes plain functions into DAGs, wiring them by their argument names, and i2 handles the signatures and wrappers underneath both. Then qh turns a Python function into an HTTP service, and py2mcp turns the same function into a tool an AI agent can call over the Model Context Protocol. You write the function once and expose it however it is needed. wads and epythet take care of packaging, CI and documentation, which is how one person keeps this many packages released.
Most of my recent work is AI integration. aix puts one Python interface over many model providers. imbed and vd handle embeddings and vector search across about fifteen databases, and ir and raglab build retrieval and agentic search on top. A newer group is tooling for coding agents themselves: crowsnest watches all my Claude Code sessions and tells me which one needs me, openloops tracks what agents left unfinished, and annals lets an agent publish a report to a page I read on my phone. More and more of my packages also ship agent skills and machine-readable docs, so an agent can use them without me explaining.
The same habits carry into sound, image and music. voxy and scribed put one interface over speech synthesis and transcription engines. an turns a written scene into rendered animation, and muvid makes music videos. tonal, arioso and accompy cover music analysis, AI music generation and backing tracks from chord charts.
JavaScript and TypeScript: zodal and acture
When my work moved into the browser, the Python habits came with it: keep one source of truth, build in small layers that compose, and put a plain facade in front of anything complicated.
The largest family is zodal. You describe your data once, as a Zod schema, and @zodal/core works out the rest: the columns of a table, the fields of a form, the filters that make sense, and the state that holds it all. A storage provider then reads and writes that data wherever it lives, in browser storage, files, S3, Supabase or a REST API, and its interface is a direct translation of dol. The same idea now covers settings (@zodal/dials-core), items that sit in several folders at once (@zodal/groups-core) and graphs (@zodal/graph-core).
The other family is acture, a command dispatch layer. In most apps one action, say "apply a filter", gets wired by hand to a button, then a keyboard shortcut, a command palette, a test, and now an AI assistant. Five copies of the same thing drift apart, and the assistant ends up with a stale idea of what the app can do. With acture you declare the action once, as a named command with a typed schema, and every surface reads that one declaration: shortcuts, the palette, the assistant's tool calls, an MCP endpoint, end-to-end tests and undo. It also works on existing code, one handler at a time. I wrote the reasoning up in a paper.
The rest are smaller tools. Several are TypeScript twins of my Python packages, so the same logic runs on both sides, such as chordsym for chord symbols. Others stand alone: sigilsmith draws deterministic avatars and icons, and pyfiddle builds pages with runnable Python cells.
Code: github.com/thorwhalen and github.com/i2mint. I am also the main contributor to Cosmograph's Python package, for visualising large graphs and embeddings.
Career
Twenty years, mostly at the start of things
- 2023 – present
Cosmograph (Figiri)
Co-founder, Director of AI
GPU-accelerated visualisation and machine-learning tools for exploring graphs and embedding spaces: interactive interfaces for semantic navigation in retrieval, recommendation and similarity search.
- 2023
Addix Group
AI integration consultant
Helped organisations adopt language models. Proofs of concept for résumé matching, a no-code AI platform and retrieval-based assistants.
- 2018 – 2023
Analog Devices
Director of Machine Learning
Led machine-learning platforms and tooling for IoT and signal processing. Built the Python frameworks that unified data access, feature engineering and experimentation, mentored data science teams and shaped the company's AI strategy.
- 2016 – 2018
OtoSense
Co-founder, CTO
Sound and vibration recognition. I led the architecture and the machine learning, and wrote the models behind four patent filings. Acquired by Analog Devices in 2018 (more below).
- 2014 – 2016
Commanders Act (TagCommander)
Lead data scientist
Attribution models, behavioural targeting, clickstream analytics and KPI time series for marketing.
- 2011 – 2014
Expedia / Venere
Senior data scientist
Language processing and semantic systems for search advertising; pricing, forecasting and the quality of geographic data at scale.
- 2000 – 2011
Methodic Solutions
Independent consultant
Machine learning for companies in Europe and the US: recommender systems, Bayesian inference, duplicate detection, semantic analysis, forecasting and optimisation. Clients included Sanoma, Easyvoyage, Ricoh, Metron, iVivity and Scentric.
The venture
OtoSense: teaching machines to hear
I co-founded OtoSense, a company that taught computers to recognise sounds and vibrations. We treated acoustic signals the way linguists treat language: break the sound into primitives, find a syntax in how they combine, and map the patterns to meaning. The technology served industrial machine monitoring, home security and accessibility for deaf people. Four of my patent filings come from this work, two of them granted.
Analog Devices acquired OtoSense in 2018, less than five years after it started. I stayed on as Director of Machine Learning until 2023, building the platforms and open-source tools the company's data scientists used to get data, engineer features and run experiments, and mentoring its data science teams.
Some of the open-source code behind that work: slang (a structural approach to signal machine learning), hear (audio data for machine learning pipelines) and creek (simple streams of live data).
Intellectual property
Patents: three granted, three more published
- Employing user input to facilitate inferential sound recognition based on patterns of sound primitives
- Syntactic system for sound recognition
- Sound-recognition system based on a sound language and associated annotations
- Facilitating inferential sound recognition based on patterns of sound primitives
- Intelligent general duplicate management system
- System and method for implementing advanced RAID using a set of unique matrices as coefficients
Research
Fifteen peer-reviewed publications
In two fields: pure mathematics (graph theory, combinatorics) and applied social science (drug-use trajectories, measuring change in longitudinal data, harm reduction).
- Language preferences and impact of labels on the lived experience of people who use heroin
- Measuring discontinuity in binary longitudinal data: applications to drug use trajectories
- New Ore-type conditions for H-linked graphs
- Drug use trajectory patterns among older drug users
Seven more, in graph theory and combinatorics
The full list is on Google Scholar and LinkedIn.
Foundations
Education and languages
Ph.D. in Mathematics
Emory University, Atlanta, 2003. Graph theory and combinatorics, with Ronald J. Gould.
M.S. in Computer Science
Emory University, Atlanta.
Licence de Mathématiques
Université de Nice Sophia Antipolis, Nice. Exchange studies at the Technische Universität Berlin.
Languages
English and French, both native. Italian and German, working knowledge.
Writing and teaching
Thinking out loud
I write about Python architecture, machine learning tooling and the craft of building systems that last, on Medium, Substack and in GitHub discussions. I have taught probability, statistics, data science and software design at universities and in industry workshops.