AI that answers from your data, safely
The most useful AI feature we build is also the simplest to describe: a text box where a manager types a question in plain English and gets an answer from the company's own data. "Which horses have been on site for more than 30 days without a bill?" "Show me every runner that won first-up over 1200 metres at this track." Behind the box, a large language model writes a database query, the platform runs it, and the result comes back as a table and a sentence.
Getting that to work reliably is the job. The model only sees an allow-list of tables. The query runs in a read-only transaction with a hard time limit. Sensitive columns, such as owner names, are blocked at the database layer, not by asking the model nicely. The feature is only offered to roles that are already permitted to see the underlying data, it is rate-limited, and there is a daily spend cap so a runaway loop cannot cost more than you agreed to.
We have built this twice in production: "Ask the Stable" inside a racing operations platform, and "System Builder" on top of a horse racing database, where the model also backtests the resulting runner list across years of results.
What we build
- Natural-language query assistantsText-to-SQL over PostgreSQL with allow-listed tables, read-only transactions, timeouts, PII blocking, role gating, rate limits and spend caps.
- Document and image readingVision-language models that read brands, markings, labels, forms and markups from photos and video frames, cross-checked against a registry.
- Explainable scoring enginesDeterministic, rule-weighted scorers that write their own reasoning lines, for when "the model said so" is not an acceptable answer.
- Machine-learning modelsGradient-boosted rankers and classifiers (LightGBM) with proper holdouts, retrained on a schedule, with the training pipeline versioned and automated.
- Evaluation harnessesQuestion banks with known answers, scored automatically on every change, so you know whether a prompt or model update made things better or worse.
- Cost engineeringModel routing (small model for easy questions, large for hard ones), prompt caching and tool chaining to bring per-question cost down to cents.
Honest AI
A lot of "AI" projects are a chatbot bolted onto a product that did not need one. We take the opposite line. Machine learning and language models go in where they measurably outperform plain code, and they are tested the same way: against a holdout set, a question bank, or a ground-truth label set, before anyone relies on them.
That testing discipline comes from research work in equine biomechanics, where we have spent years separating signals that replicate from signals that only look good once. The same scepticism applies to every model we ship for a client: if an effect does not survive a held-out test, it does not go on a dashboard.
Platforms and tooling
We build primarily on the Anthropic Claude API, with model routing between smaller and larger models to balance cost and quality. Classical machine learning uses LightGBM and scikit-learn; training jobs run on a schedule on Render or GitHub Actions and write versioned artefacts. Analytical workloads use DuckDB and PostgreSQL. Vision models are covered under computer vision.