What we build
AI agents with enforced limits. Knowledge systems that show their sources. Models delivered with an independent assessment of their limits.
-
Agentic AI for regulated operations
Agents that investigate, decide and act through controlled tools. Autonomy is set per workflow, from advisory to act-with-approval. Every step is logged, and consequential actions require human approval.
See the agent projects -
Auditable knowledge systems
Retrieval, knowledge-graph and ontology solutions in which each answer traces to a source or a rule.
See the knowledge projects -
Simulation and decision support
Deterministic engines with live operational interfaces, used to compare strategies before capital is committed.
See the simulation projects -
Applied machine learning
Vision, fraud and prediction models, reported on held-out data with stated limitations.
See the ML projects -
Advanced research
Computational biology, quantum computing and multi-agent reinforcement learning, taken from paper to working software.
See the research projects
How we work with enterprise teams
From a bounded pilot to a clean handover.
-
Step 1
Start with a bounded pilot
We agree the success measures, the data in scope and the autonomy level before work begins.
-
Step 2
Prove it on safe data first
Pilots run on synthetic or sanitised data until your security and risk teams have reviewed the controls.
-
Step 3
Transfer knowledge as we build
We work alongside your engineers and hand over documentation, evaluation suites, tests and runbooks, so the system does not depend on us.
-
Step 4
Plan the exit at the start
Ownership of code and IP, handover and transition responsibilities are agreed in the contract, not after delivery.
How we deliver
-
Autonomy is set to match the risk
Each workflow is assigned one of four levels: observe, advise, act with approval, or act autonomously. Higher-risk workflows get tighter controls rather than one blanket policy.
Seen in
-
Controls are enforced in code
Spend limits, approved-vendor lists, read-only data access and approval gates are applied by deterministic logic before any action runs. If one component fails, deterministic fallbacks, automatic re-weighting and cached responses keep the service running.
Seen in
-
Every output is traceable
Findings cite their evidence. Diagnoses cite the rule that produced them.
Seen in
-
Results are reported candidly
Where a simple rule outperformed a learned model, or a headline score was partly circular, we say so.
Seen in
-
Systems are built to be owned
Architecture is documented, evaluation suites and tests ship with the code, model providers can be swapped, and benchmark scripts let your team reproduce every figure. The aim is that your engineers can run and extend each system without us.
Seen in
-
Evaluation precedes data sharing
Demonstrations run on synthetic or public data, so behaviour can be assessed before any sensitive data is exposed.
Seen in
Security by design
Our agent work is designed against the risks in the OWASP Top 10 for Agentic Applications (2026). The controls below are built into the projects listed. They support your own security review and do not replace it.
-
ASI01
Agent goal hijack
Orchestration and limits sit in code rather than in the prompt, so instructions hidden in a document or message cannot change the limits the agent works within.
-
ASI02
Tool misuse and exploitation
Tools are narrow and guarded: read-only database access, spend limits, approved-vendor lists and tolerances.
-
ASI03
Identity and privilege abuse
Permission policy is explicit, and consequential actions pass an approval gate before they run.
-
ASI08
Cascading failures
Writes are idempotent and verified by read-back. Run state persists, and work replays safely after a forced shutdown.
The lab's periodic table
The tools we reach for most, grouped by what they do.
Agentic AI
Multi-agent orchestration, tool calling, structured outputs, human-in-the-loop gates, evaluation harnesses.
- 1LangGraph
- 2LangChain
- 3MCP
Retrieval and knowledge
RAG, GraphRAG, Cypher queries and vector search.
- 4Neo4j
- 5Qdrant
- 6FAISS
- 7Firestore
- 8Postgres
- 9PostGIS
- 10Redis
- 11OWL/RDF
- 12Pellet
ML and research
Including multi-agent reinforcement learning (DQN, MAPPO, MADDPG).
- 13PyTorch
- 14scikit-learn
- 15XGBoost
- 16OpenCV
- 17YOLO
- 18SHAP
- 19LoRA/QLoRA
Science
Docking, cheminformatics and quantum computing.
- 20AutoDock Vina
- 21DiffDock
- 22Boltz-2
- 23RDKit
- 24Qiskit
- 25PennyLane
Product and backend
APIs, web apps and browser automation.
- 26Python
- 27TypeScript
- 28FastAPI
- 29Next.js
- 30React
- 31Tailwind
- 32Playwright
Delivery
Containers, CI, cloud and infrastructure as code.
- 33Docker
- 34GitHub Actions
- 35AWS
- 36Terraform
- 37Vercel
Logos are trademarks of their owners and appear here only to name the tools we use.
What are you building?
Tell us what you're working on and we'll get back to you within two working days.