Custom models. Trained, evaluated, shipped.

We fine-tune vision and language models on your data, build the retrieval and tool-using agents around them, and train the unglamorous forecasting models that often do the real work — then compress them to run on-premise or fully offline. Everything is measured against a baseline, gated for safety, and traced in production. The weights are yours.

  • Fine-tuning & QLoRA
  • Embedded vision — EfficientNet, MobileNetV4, DINOv2
  • Vision & OCR
  • Model councils
  • Forecasting & anomalies
  • RAG & agents
  • Quantisation & distillation
  • Edge & offline
  • Evaluation & tracing

We run this ourselves, before we sell it.

School nutrition · in build

Fine-tuned vision

Closed-set dish recognition for school meal monitoring — a compact vision model adapted on real program photos across successive QLoRA training rounds, paired with purpose-built seven-segment OCR that reads serving weight straight off the scale in frame.

Health · running

Private model councils

For a cancer-wellness product: MedGemma, OpenBioLLM and Meditron answer independently on-premise, a Gemma synthesiser combines them into one warm, non-diagnostic reply, and a crisis gate runs before any model is called. Nothing leaves the building.

Health data · running

Forecasting + insight agents

Two layers over personal health data: per-subject gradient-boosted models forecast trends and flag anomalies, then a local LLM turns those signals into plain-language daily insight — reading every metric through a tool interface rather than a fine-tune.

Small models, owned outright, beat rented giants at the task.

Most enterprise AI work does not need a frontier model — it needs a model that knows your data, runs where your constraints allow, and costs nothing per call. We fine-tune compact open models to the task, evaluate them against a baseline you can see, and hand you the weights.

  • Your data stays yours — training and inference can run fully on-premise
  • Evaluated against a measured baseline, not a demo
  • Fine-tuned and compressed to fit the device — from data centre down to offline edge
Fine-tune run

Adapted to the task, measured at every step.

training losseval gate
QLoRA

Fine-tune

Quantised

Runs offline

Weights

Yours to keep

Four systems we built, shipped and still run.

Not pilots or proofs of concept — working systems, on our own infrastructure, with the operational scars that come from keeping models running after the demo.

School nutrition · social impact

A plate, a scale, and a model that reads both

Mid-day meal programs feed millions of children with almost nothing measured. We built a vision system that turns one kiosk photo into a verified record: which dishes were served, how many grams were on the scale, and the nutrition that reached that child.

  • Training data labelled by an ensemble of vision-language models, then corrected and verified by people before any round of training
  • Embedded architectures trained and compared head to head — EfficientNet and MobileNetV4 classifiers, plus DINOv2 self-supervised features — to find what holds accuracy at the size the device allows
  • Fine-tuned on real program photography, iterating through successive QLoRA rounds against a held-out evaluation set
  • Built seven-segment OCR to read the weighing-scale display from the same frame — no second device, no manual entry
  • Nutrition computed per serving from the recognised dish and measured weight
  • The model is compressed and quantised to fit the device it ships to — accuracy held against the evaluation set at every step down
  • Runs offline on a Jetson or Raspberry Pi class device in the school, so children's images never leave the building
Cancer wellness · consumer health

A council of medical models, entirely on-premise

For 4CancerWellness — an oncologist-founded platform — we built a companion that answers patient questions without shipping a word of it to a third-party API.

  • A council of open medical models — MedGemma, OpenBioLLM, Meditron and a PubMedQA-tuned Qwen — each answering independently on local hardware
  • A Gemma synthesiser composes the individual answers into one warm, non-diagnostic reply
  • A crisis risk gate evaluates every message before any model is invoked
  • Graceful backend fallback: local council first, hosted frontier model second, static guidance last — the product never hard-fails
  • Served through Ollama, with every model, expert set and endpoint environment-configurable — the deployment shape is a decision, not a rewrite
Personal health data · longitudinal

Forecasts underneath, language on top

Fifteen years of labs, wearable streams and clinical records feeding a system that has to say something useful every morning — without hallucinating about someone's health.

  • Per-subject gradient-boosted models forecast metric trends and flag anomalies against personal baselines
  • A local Qwen model converts those numeric signals into plain-language daily insight
  • The agent reads every metric through a tool interface — retrieval over live data, not a fine-tune that can go stale
  • Lab PDFs parsed into structured results; genome variants annotated locally against pharmacogenomic and clinical-variant references
  • Full tracing on every model call, so behaviour can be audited and regressions caught
Product features · SaaS

AI where it earns its place in the product

For Missing Page, a digital-legacy platform, AI is not the headline — it removes friction at the two places people stall.

  • A support assistant grounded in the product's own documentation
  • A help-me-write feature for people facing a blank page at the hardest moment
  • Keys and configuration held server-side as container environment, never in the repository

Eleven disciplines behind an AI engagement.

Use-case & data assessment
Data labeling & human reviewModel-assisted labeling, verified by people
Model training & fine-tuningQLoRA and full fine-tunes, vision and language
Forecasting & anomaly detectionGradient-boosted models on your own baselines
Retrieval & tool-using agentsEmbeddings, RAG and structured tool calls
Embedded & edge vision modelsEfficientNet, MobileNetV4 and DINOv2 features, quantised to fit the device
On-premise & on-device deploymentYour servers, or Jetson and Raspberry Pi in the field
Evaluation & benchmarkingMeasured gates before anything ships
Safety gating & fallbacksRisk checks ahead of the model, graceful degradation behind it
Tracing & monitoringEvery call observable, regressions caught
Enterprise workflow integration
Proof of work

See what we build when nobody is paying us to.

Meal-recognition AI for school nutrition programs, a robotics simulation lab, and the ventures we back — on the Innovation page.

Let's discuss how we can make your business better.

Tell us what you are running and what is in the way. We will tell you honestly whether we are the right people for it.

937-660-4748