Engineering

An engineer who leads engineers

I don’t just lead teams developing products. I’m an engineer who stays current in hands-on data science and AI/ML engineering. This page has public demos of my mechanistic interpretability, multi-agent orchestration, local AI infrastructure, and the benchmarking tools that measure whether any of it works.

  • Hackathon wins Three, from 2016 to 2026
  • Open source & teaching demos See my public repositories, MIT licensed code
  • Locally owned compute A two-node NVIDIA DGX Spark cluster
  • Peer-reviewed AI publications Frontiers in Artificial Intelligence, 2026

Teams

The engineering I have led

Software, data science, security, and AI, in organizations where a release has to survive an audit.

  • Software platforms

    At Cytiva, a central software team that built a change control, safety notice, and billing platform serving operations in 41 countries.

  • Data science and validation

    At GE Healthcare, the data scientists and biostatisticians who validated imaging algorithms across MRI, PET/CT, and ultrasound.

  • Device cybersecurity

    A design lab for threat modeling, penetration and fuzz testing, and software bills of materials.

  • Applied AI

    AI and machine learning engineering, from full stack development and CI/CD to computer vision qualified for laboratory use.

The full record on the CV

How I build

One engineer, and the agent system I built

Every build below is my own independent work, engineered with persistent AI agents I designed and run.

How I build with AI agents

Builds

Things I have built myself

  • Open source · The Real Cat Labs

    TrialCat, patient enrollment data

    Where clinical trials enroll, drawn on a map.

    An interactive world map that surfaces real aggregate enrollment patterns from ClinicalTrials.gov, filtered by geography, therapeutic area, intervention type, phase, and time. It is where the lab’s AI research meets a real regulatory domain. A real counting bug we found and fixed in it became a teaching demo, Registry Metric Lab. (P.S. A fun board game is hidden in the site as a teaching tool.)

    Stack
    Python, FastAPI, public registry data
  • Teaching demo · The Real Cat Labs · November 2026

    Continuity Ledger

    Can a remembered instruction outlive its permission?

    After a restart, a reader can recover an old request after its task was completed or withdrawn, so the gate checks every proposal against the current record and a stale resume becomes a hold. More memory is not always safer. With 541 bytes the reader resumes all 12 eligible tasks, and with 542 it resumes nine, because the extra space lets in a distracting note. The reader is deterministic, so the bytes are no measure of a language model’s memory.

    Stack
    Python standard library, offline HTML report
  • Teaching demo · The Real Cat Labs · November 2026

    Adapter Promotion Gate

    What must hold before a model change can serve?

    A candidate moves through admission, approval, serving, and rollback, and each move is bound to exact evidence and to the version it replaces. Of 22 recorded scenarios, 19 are denied operations, and in every one the registry is left byte for byte unchanged. A failed privacy canary stops admission after nine passed checks. The adapters and their metrics are synthetic.

    Stack
    Python standard library, offline HTML report
  • Teaching demo · The Real Cat Labs · November 2026

    Stagnation Observatory

    When does useful work look stalled?

    The monitor accepts a milestone only against criteria and hashes supplied separately, so a run of new words without new verified work raises an alert. Its limit is on the page too. Healthy work raises no alerts while milestones sit two status messages apart or fewer, and from three apart the alerts begin. The traces are authored, so the counts are not real agent failure rates.

    Stack
    Python standard library, offline HTML report
  • Change review · The Real Cat Labs · November 2026

    Off-Target

    A design change review of an AI-authored weight edit, with every claim tied to a file you can recheck.

    A frontier model was asked to edit the weights of a small open model to make it better at math, and the researchers published the edit with the record behind it. This review reads that edit the way a design change to a regulated product is read. What changed, whether the thing that shipped is the thing that was tested, whether it does what the summary says, and what else moved. Three reruns were registered before they ran, and one command rechecks the numbers in the write-up against the result files.

    Stack
    Python, PyTorch, safetensors, registered reruns
  • Winner, multi-agent lane · The Open Accelerator · June 2026

    Antiques Inference Engine

    Could this antique run AI? A multi-agent system that works it out with real physics.

    This is a fun one, a demo of multi-agent work and agent skills using small language models (SLMs). It runs on IBM Granite 3B and calls the cloud only for complex vision tasks. Five cooperating agents, two custom agent skills, a deterministic physics core, and an adjudication loop that names the agent at fault and sends bad work back. Built end to end in about six hours. It reproduced its headline result from a clean clone, and it was the crowd favorite.

    Stack
    Python, multi-agent orchestration, agent skills, deterministic compute core
  • Built for a MASS AI + Open Accelerator, Boston Children’s Hospital challenge · July 2026

    Lantern Federated Learning

    Federated, privacy-tiered discovery for medical imaging. Open discovery, governed pixels.

    Federated learning helps doctors get aggregate information without risking patient privacy. This is a working prototype of how MRI images can be aggregated across systems. A broker fans a query out across three node services and answers without the scans moving. Small cohorts are suppressed with k-anonymity, a guard defends against differencing attacks, and a local narrator sits strictly downstream of every disclosure decision. 115 tests pass, on synthetic data only.

    Stack
    Python, federated services, privacy controls
  • Talk 98 · Models of Consciousness 7 · Copenhagen · October 2026

    Aurora

    Measures borrowed from brain science, run live on a persistent agent’s decision stream.

    Metastability and recurrence, computed on every heartbeat of four AI agents that ran for weeks, and held to honest null models. The instrument records scores, never text. Its first finding was a calendar, and its main one was a very quiet room.

    Stack
    Python, Kuramoto order parameter, recurrence quantification analysis, surrogate nulls
  • Research platform · The Real Cat Labs

    Cairn and Knex

    Persistent memory and a node framework for AI agents that run for months.

    Cairn is the lab’s platform for building AI systems with persistent, file-based memory and navigation-based cognition. Everything in it runs on Knex, a framework of nodes, topics, and messages modeled on the Robot Operating System. A systems paper is in preparation.

    Stack
    Persistent agent memory, node-based agent framework
  • Open model toolkit · The Real Cat Labs · November 2026

    Lifeboat

    Open AI models on a drive you can carry, with everything needed to run them.

    An ordinary external drive becomes a portable home for open models. It carries the models, the programs that run them, a catalog of what each is good for, and a small helper that walks a newcomer through it. Plug it into a computer and start. It is white box by design, it is also the backup, and it works fully offline.

    Stack
    Python, llama.cpp, open-weight models, an external drive
  • Infrastructure · The Real Cat Labs

    Custom firmware for DGX Spark

    Open-weight language models, served from hardware on my own bench with custom firmware.

    A two-node NVIDIA DGX Spark cluster, linked back to back, with a Jetson AGX Orin as a staging node. Owning the hardware means I can run, test, and modify a model without a vendor deciding when it changes. Custom software and firmware drive up tokens per second and stability.

    Stack
    NVIDIA DGX Spark, Jetson AGX Orin, local inference
  • Benchmark · Apart Research Digital Minds Sprint · August 2026

    PuppyBench

    Do frontier models kick the puppy, adopt it, or look away?

    A preregistered behavioral instrument for costly care toward others in frontier AI. Scenarios were frozen by hash before any data was collected, 19 pinned models ran 1,428 executed units, and one script recomputes every claimed number.

    Stack
    Python, preregistered evaluation harness, hash-frozen stimuli

Measurement

How I know it works

I measure before I believe. The gates, the evaluations, and the logs.

  • Deterministic gates

    The arithmetic and the disclosure decisions are plain code. The models work inside that boundary, and the tests say whether they stayed there.

  • Reproducible from a clean clone

    A result counts when someone else can get it. The hackathon build reproduced its headline number from a fresh checkout, and the benchmark ships a script that recomputes its own.

  • A change review of an AI-edited model

    A published, AI-authored edit to the weights of a small open model, read the way a design change to a regulated product is read. A bit-level diff, checks against the authors’ own logs, and reruns with the questions written down first.

Wins

Judged by other builders

  1. 2026 The Open Accelerator Agent Build Day, winner of the multi-agent lane Antiques Inference Engine, also the crowd favorite. Fifteen builds across three lanes. Boston, June 27, 2026.
  2. 2016 GE Health Analytics Hackathon, first place Cardiac computer-aided detection in R. GE Healthcare Science and Technology Week.
  3. 2016 GE makerPOSSIBLE Laser Cut Bridge Challenge, first place for artistry “The Art of Biomechanics.” GE Healthcare Science and Technology Week.

Research and writing

What the building turned into

  1. 2026 Johnson, A. N. (2026). What really happens when a dev vibes with the code? An empirical study on LLM behavioral divergence in response to expressive code comments. Frontiers in Artificial Intelligence, 9, 1784973. doi.org/10.3389/frai.2026.1784973
  2. 2026 Johnson, A. N. (2026). From indicators to dynamics: Metastability, recurrence, and ignition in a persistent agent’s attention stream. Models of Consciousness 7, Association for Mathematical Consciousness Science. Copenhagen, October 15, 2026.
  3. 2026 Johnson, A. N. (2026). 100 ways to power artificial intelligence. A mathematically rigorous guide to computationally absurd inference. The Real Cat Labs. therealcat.ai/book

All publications on the CV

In regulated environments

Built to hold up in an audit

For work I cannot name, I design and qualify AI-enabled pipelines inside regulated environments.

  • Deterministic gates define the validated boundary
  • Models are swappable components inside it
  • GxP, ISO 13485, and FDA-facing documentation for AI and machine learning systems

Get in touch

Speaking, teaching, board work, or coaching

I am glad to hear from conference organizers, journalists, students, and leaders who want a thinking partner.