Enterprise grounding, and beyond LLMs.
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. It cannot reliably add up what you sold last quarter either. That is a database’s job.Grounding is how you close that gap: in your own documents and data through retrieval, in symbolic tools that define what may be asked and check what comes back, and in the forecasters and classifiers that already hold the answer. We build across that 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, retrieval, and tool contracts that let a language model reach your systems, your forecasters and your classifiers within guardrails that balance power, speed, safety and reliability.
- MCP servers and agent skills
- Retrieval and grounding pipelines
- Tool contracts and model interfaces
- Feature and model infrastructure
- Systems administration, integration and DevOps
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 parts, each picked for a single job, and most of the work is in how they are joined.
- 01
Each part has one job.
The neural part reads and proposes. The symbolic part defines what may be asked and checks what comes back. The deterministic part executes and keeps the record. Ask a language model to build a shift rota and you get one that looks right and breaks three of your rules. The right tool would have told you in a second that no valid rota exists.
- 02
The interesting failures are at the joins.
Not inside any one component. An extraction that is correct, in a shape nothing downstream can consume. A model that keeps answering confidently after the world it learned from moved on. A check that fails quietly because nothing was listening. Every part passes its own test. The system does not work.
- 03
So the guarantees have to live in the wiring.
Not inside the model, where you cannot inspect them. This is what grounding actually means in an enterprise: the retrieval says what it may draw on, the contract says what may be asked and what must come back, the evaluation says whether it held, the instrumentation says when it stopped holding. Get that layer right and you have something you can put in front of a customer. 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.
The work came first.
The insight comes from practice: years of picking up new capabilities early and taking them into production to find out what they can actually carry. A deep learning system on live broadcast in 2021, and the engineering discipline that let a forecasting team move faster and more safely.
Automotive parts
Forecasting collision-parts demand
Predicting which parts a dealer network would need, and when. The interesting part was underneath: a Features API and a shared library that let a distributed team try ideas without tripping over each other, and an automated search across a space too large to walk by hand. The modelling belonged to the team.
Data science platform, method onlyRead itLive broadcast
Real-time language detection at the Olympics
Deep learning listening to live audio and flagging language mismatches while the feed was on air. Built in 2021, when this was still a novel thing to attempt, and run through the French Open, Wimbledon and the Tokyo Olympics in a single season. No second takes.
AI engineering, 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