RAG

Retrieval-augmented generation (RAG)

For AI to answer from your company's documents and data instead of making things up, the model needs the right context, at the right time, with the right permissions. We build the whole pipeline, not just the prompt.

Ingestion of PDFs, contracts, spreadsheets and internal databases, with structure-aware chunking and metadata for filtering
Embeddings and vector search — pgvector for teams already on PostgreSQL — combined with lexical (hybrid) search and reranking for precision
Access control at retrieval time: each user only retrieves what they are already allowed to see
Answers with source citations, and retrieval recall evaluated separately from generation quality

AGENTS

AI agents in production

An agent is a model in a loop: it reasons, picks a tool, executes, observes the result and repeats until the task is done. In production, the challenge is granting autonomy without losing control.

Agentic workflow design: single agent or orchestration with specialised sub-agents, task decomposition, state and memory management
Tool use with clear contracts, idempotent actions, timeouts, retries and sandboxed execution where there is risk
Autonomy limits and human-in-the-loop approval points wherever a decision has consequences
Full traceability: every step, every call and every cost logged for audit and debugging

LLMs

Large language model integration

Claude, GPT and open models wired into your back end like any other critical dependency: with a contract, tests, observability and a plan B.

Structured outputs with JSON Schema, so model results flow straight into your system without brittle parsing
Prompts versioned and tested as code; model routing (smaller and cheaper for simple tasks) and fallback across vendors
Prompt caching, batching and streaming to cut cost and latency; budgets and usage limits per feature
An abstraction layer that lets you switch models or vendors without rewriting the application

MCP

MCP servers

The Model Context Protocol is the open standard, created by Anthropic, that defines how an AI model discovers and uses external tools, data and prompts. A well-built MCP server is AI's secure front door into your company.

MCP servers exposing ERP, CRM, databases, internal APIs and documentation as typed tools and resources
Authentication (OAuth), per-user authorisation and an audit trail for every call — the assistant only does what that user could do
Compatible with Claude and any MCP client; one server serves multiple assistants and agents at once
Tools designed with the model in mind: clear descriptions, narrow scope and responses that fit the context window

EVALS

Evaluation and quality

Without measurement, every AI system looks fine in the demo and surprises you in production. We treat evaluation as part of the product from day one.

Evaluation sets built from real business cases, including the hard ones and the ones that should be refused
Metrics per layer: retrieval precision, answer correctness and groundedness, hallucination rate, task success for agents
LLM-as-judge with human calibration, and automated regression tests on every change to prompts, model or data
Production monitoring: response sampling, user feedback and degradation alerts

SECURITY

Security and governance

An agent that reads email, web pages or documents is exposed to malicious instructions hidden in that content. We design with that in mind.

Prompt injection defences: separation of instructions from data, output validation and least privilege for every tool
Personal data handled under GDPR, CCPA and Brazil's LGPD: minimisation, anonymisation where possible and a record of who accessed what
Input and output guardrails, content filters and spend limits per user and per feature
Documented risks and decisions for your security and compliance teams to review before go-live

AI that lives in the system, not on the slide.