AI is not one thing.
A language model can read ten years of email and tell you how your organisation actually works. It cannot tell you what you will sell next quarter. That is a forecaster’s job. Extracting meaning, extrapolating signals, classifying, or sometimes just applying a rule - we build across the range, and wire the parts together so the whole does what none of them could alone.
Four jobs, four kinds of tool.
Prediction and forecasting
Demand, risk, capacity, churn, what happens next. The models that move a number rather than describe one, built to be retrained and rerun rather than presented once.
- Demand and capacity forecasting
- Risk and propensity models
- Automated model and feature search
Extraction and classification
Structure pulled out of email, tickets, contracts, documents and transcripts. How the organisation actually works rather than how the org chart says it does. What a support queue is really telling you. What a document commits you to.
- Entity and relationship extraction
- Document and message classification
- Real-time language and content detection
Integration and tooling
The wiring between the neural and the deterministic: MCP servers, agent skills, and tool contracts that let a model reach your systems and your models without being trusted to behave itself.
- MCP servers and agent skills
- Tool contracts and model interfaces
- Feature and model infrastructure
Evaluation and monitoring
Knowing whether it works, and knowing when it stops. Tests a model cannot quietly pass by accident, and the instrumentation that tells you something has drifted before a customer does.
- Evaluation harnesses that cannot be gamed
- Point-in-time and leakage-safe testing
- Drift and degradation monitoring
Models work best in company.
Almost nothing useful is one model doing one thing. It is several, each picked for what it is actually good at, wired together so the whole is checkable.
- 01
A language model is the best tool ever built for turning text into structure.
Ten years of email will tell you how decisions actually get made in your company, which is rarely what the org chart says. A support queue will tell you what your product is doing to people. Contracts will tell you what you have agreed to. None of that was practical to extract at scale five years ago, and now it is routine.
- 02
It is not what should forecast your demand or hold your source of truth.
Those are different jobs with better tools: a forecaster, a classifier, a ranker, a rules engine, a query against something authoritative. Using a language model for them is expensive, slower, and impossible to audit. Most of the disappointment with AI comes from asking one component to do all four jobs.
- 03
The value is in the interplay, and the wiring is the work.
MCP servers and agent skills are how a model reaches a tool. Contracts and evaluation are what let you rely on the result. Get that layer right and each part does what it is good at: the neural part reads and proposes, the symbolic part defines and checks, the deterministic part executes and keeps the record. Get it wrong and you have a very expensive chat interface.
The longer version of this argument, and where it came from, is in Symbolic, testable tools for neural agents.
Two of these, in anger.
One a forecasting engagement where the platform underneath decided how fast the team could work. One a deep learning system running live on air, where a late answer and a wrong answer are the same thing.
Automotive parts
Forecasting collision-parts demand
A five-person team forecasting which parts a dealer network would need, and when. Rudolph built the shared-library platform, the Features API and the automated search that let the team evaluate many models safely. The modelling was the team's work.
Method only, data under NDARead itLive broadcast
Real-time language detection at the Olympics
Deep learning applied to live audio, detecting language mismatches across broadcast feeds while they were on air. Built and run through the French Open, Wimbledon and the Tokyo Olympics in a single season, where being right an hour later is being wrong.
Delivered, 2021Read it
Twenty-five years in production, and the mathematics to match.
A career spent building the data systems that live broadcast, elite sport, regulated finance and global education depend on, plus certifications from Cambridge and Imperial College in data science, machine learning, and the mathematics of machine learning. Plenty of people can train a model. Rather fewer have found out what happens to one at three in the morning during a live event.
The thinking, at length.
Written from the work rather than about it.
- Essay3 min read
Symbolic, testable tools for neural agents: a small neurosymbolic example
Frank Coyle argues that neural agents need to be tethered to symbolic, testable tools. The platform behind the Inchcape collision-parts engagement is a small, concrete instance of exactly that, long before an agent is in the loop.
Read article - Case study4 min read
Rails for data science: the platform beneath a forecasting engine
A shared library, a model contract and a Features API let a distributed team evaluate many forecasting models safely, and turned feature search into something you can automate. Notes from the Inchcape collision-parts engagement.
Read article - Case study3 min read
AI-powered broadcast monitoring at the Tokyo Olympics
In 2021 we realised deep-learning was entering mainstream development. This real-time AI NLP system detected language mismatches across live broadcast feeds and came together for the French Open, Wimbledon and the Tokyo Olympics that year.
Read article
Got a question your dashboard cannot answer?
Whether that is a forecast you do not trust, a pile of documents nobody has time to read, or an AI system you would not yet let a customer near, that is the conversation we want to have.
- rdd@palefire.io
- Studio
- London · Remote worldwide