UPPi
Ask a forecasting question in plain English and get a tested answer with an honest uncertainty band. The prototype has been benchmarked on 10 public datasets across five sectors.
View projectRESEARCH LAB · AGENTS / PREDICTION / RETRIEVAL
Building AI that shows its work.
FutureCraft Research Lab studies how to make AI systems people can check: agents that learn from memory, forecasts that prove themselves on your own data, and assistants that answer only from sources they can point to.
4
Research programmes
+287
Datasets catalogued
12
Forecasting methods
+4
Research notes published
WHAT WE WORK ON
Three directions, one open question
Can an AI system's answer be checked against evidence, rather than simply trusted? We approach that question from three directions at once.
Learning Agents
How groups of AI agents can improve by remembering past work instead of retraining, and how to decide which agent gets to remember what.
.svg){kind=small}
Trustworthy Prediction
Forecasts chosen and tested on each customer's own data, with uncertainty ranges that actually hold up in practice.
.svg){kind=small}
Grounded Retrieval
Search and answers across specialist collections such as legal documents, CAD models and technical drawings, where every statement points back to its source.
Research Areas
Engineering the infrastructure of trustworthy AI
Six areas where our research groups are building the methods that make intelligent systems checkable, from agent memory to engineering search.
.svg){kind=small}
Memory-Based Agent Learning
Frozen AI models that keep improving by storing past cases and recalling similar ones, with no changes to the model itself.
.svg){kind=small}
Market Mechanisms for AI Fleets
Testing whether agents bidding with earned credits allocate work & memory better than a central controller, once the market's own running costs are counted.
Automatic Model Selection
Picking the right forecasting model for each customer from their own data, then re-selecting automatically when the data shifts.
Rigorous Evaluation
Pre-registered experiments, leakage checks and benchmarks built from realistic data, so results can't be tuned after the fact.
Multimodal Engineering Search
Finding existing parts and drawings by description, rough sketch or 3D geometry, ranked by the dimensions and tolerances that actually matter.
Source-Grounded Answers
Assistants that tie every statement to a named source and say clearly when the answer isn't there.
Featured Work
Active projects
A sample of what our research groups are shipping right now.
MECH-RAG
AI search across CAD models and technical drawings that helps engineers find and reuse existing parts instead of designing duplicates.
View projectAgent Credit Market
Eight identical AI agents bid for work and memory space using credits they earn. The study tests whether this beats a strong central router once the market's costs are counted.
View projectFrom the lab notebook
Latest from the blog
We publish as we go: findings, design notes, and the occasional dead end.
September 29, 2026
Visit the BlogWhat AI Can and Can't Do for a Law Firm Yet
Extracting facts and tracking deadlines already work well. Legal judgment and predicting outcomes don't. A practical map of seven kinds of legal work.
September 29, 2026Scaling Zero-Knowledge Proofs for Neural Networks
Eighteen months ago, proving a single inference pass took 4,000x longer than the inference itself. Here is how we got that down to 40x.
September 29, 2026Why an 80% Forecast Band Covered Only 61%
On retail sales data, our forecast bands promised 80% coverage and delivered 61 to 65%. Here's why the most accurate model isn't always the most honest one.
Research & Publications
Behind the scenes of our research
Retrieval-Augmented AI for CAD Libraries and Technical Drawings
A research plan for searching engineering libraries by text, sketch or geometry, organised around four topics, from reading drawings to grounded answers.
AI for Law Firms: Task Types, Importance and Difficulty
Seven kinds of legal work rated by importance and by how far AI has solved them. The rule of thumb is to let AI do work that can be checked against a source and keep judgment with lawyers.
A Credit Market for Work and Memory Among Frozen LLM Agents
A fully pre-registered design for testing whether agents trading work and memory beat central routing, with the market mechanics already verified by 31 simulation tests.