LLM & Generative AI Solutions
LLM features that cite their sources and know when to stop.
We build retrieval-augmented assistants, agents and copilots that answer from your documents, systems and policies. Every project includes an evaluation set from day one, so "it seems better" becomes a number, and so does the cost per answer.
Schedule 1
What is included
- B-2.1
Retrieval-augmented generation (RAG)
Search over your documents, tickets and databases, with answers that cite the passages they used.
- B-2.2
Agents and tool use
Assistants that act through your APIs (create the ticket, draft the quote, update the CRM) with permission checks and human approval where it matters.
- B-2.3
Evaluation and guardrails
Test sets, automated grading, red-teaming for prompt injection and data leakage, and regression checks on every prompt change.
- B-2.4
Model choice and fine-tuning
Hosted frontier models or open-weight models, chosen per task on quality, latency, cost and privacy. Fine-tuned when it pays off.
- B-2.5
Cost and latency control
Caching, routing and token budgets, so the bill grows with value rather than with traffic.
Schedule 2
Typical projects
- Support assistants that resolve tickets from your knowledge base
- Internal copilots over policies, contracts and wikis
- Document extraction: invoices, forms, claims, CVs
- Sales and proposal drafting from your past work
- Natural-language search and analytics over company data
Materials
- Claude
- GPT
- Llama
- Mistral
- Qwen
- LangGraph
- LlamaIndex
- pgvector
- Qdrant
- vLLM
Schedule 3
How it runs
- 01
Discovery · 1–2 weeks
We map the problem, the users, the data and the constraints, and agree what success will be measured by.
- 02
Blueprint · end of discovery
Architecture, scope, milestones and a fixed price for the first build phase, all in writing. You can stop here and keep the blueprint.
- 03
Build · weekly demos
Short iterations with a working demo each week, automated tests from day one and a shared board you can always see.
- 04
Launch & handover · planned cut-over
Staged rollout, monitoring switched on, documentation and a recorded handover session with your team.
- 05
Run & improve · optional
A monthly support plan, or nothing: the handover is written so that you never need us to keep the lights on.
Schedule 4
Common questions
Will our data be used to train someone else’s model?
No. We use enterprise API terms that exclude training on your data, or run open-weight models on infrastructure you control. Which one is a decision we make with you, in writing.
How do you stop the assistant making things up?
By grounding answers in retrieved sources, requiring citations, refusing when the sources do not answer the question, and measuring the hallucination rate on an evaluation set before and after every change.
Can it run without sending data to an external API?
Yes. See our on-premise and local AI service: the same applications, running entirely on your hardware.