AI Engineering

Use-case and model selection

• Use-case definition: we define the task, the success criteria and the risks.
• Model comparison: we compare hosted and open-weight models on quality, latency, cost, data residency and licensing for *your* use case, using your own evaluation set.

Model deployment

• Hosted APIs: set up with gateways, retries, fallbacks, caching and cost controls.
• Self-hosted and open-weight models: on vLLM, TGI or similar, deployed on GPU infrastructure in AWS, Google Cloud (Vertex AI) or Kubernetes.
• Autoscaling and serving: autoscaling, batching, quantisation where suitable, and observability for latency, throughput and cost per request.

Fine-tuning

• When it's worth it: we tell you when fine-tuning is worth doing and when better prompts or retrieval would do the job.
• Data and training: we prepare datasets, train (LoRA/QLoRA or full), and evaluate against a held-out set.
• Versioning and rollback: every model version is tracked and can be rolled back.

RAG and context engineering

• Ingestion pipelines: parsing, chunking strategies and metadata.
• Embedding and vector store choices: pgvector, OpenSearch, Pinecone, Vertex AI Vector Search.
• Retrieval quality: hybrid search (keyword plus vector), re-ranking and query rewriting.
• Context assembly: what goes into the prompt, in what order, and within what token budget. This includes conversation memory and per-user access filtering, so users only ever retrieve documents they are allowed to see.
• Citations and grounding: answers point back to their sources.

Agents and tool use (LangGraph)

• Agent workflows: stateful agents in LangGraph with explicit graphs, checkpoints and human-in-the-loop steps.
• Tool design: a small, well-described set of tools, structured inputs and outputs, and tool selection and routing that we test.
• Guardrails:
  • tools run with the end user's permissions
  • actions that change state require confirmation
  • step and cost limits
  • prompt injection defences

Evaluations

• Eval datasets built from real or representative queries.
• Measures:
  • answer correctness and groundedness
  • retrieval precision and recall
  • tool-call accuracy
  • latency and cost
• How quality is checked: automated scoring (rule-based and LLM-as-judge, calibrated against human review), plus regression evals in CI so prompt or model changes can't quietly lower quality.
• Production monitoring: tracing, feedback capture and drift checks.

Safety and data protection

• PII handling: minimised and masked before model calls.
• Retention: prompts are logged with redaction, and retention settings are reviewed.
• Testing: against the OWASP Top 10 for LLM Applications. See also Security Testing.

1. Scoping call. We discuss the use case, data, constraints and success criteria.
2. Discovery and eval set. We define "good" before building.
3. Prototype. We compare approaches (prompting, RAG, fine-tuning, agents) against the eval set.
4. Production build. Deployment, monitoring, guardrails and cost controls.
5. Launch and iterate. Staged rollout, with evals and feedback guiding improvements.
6. Handover. Documentation, runbooks and training for your team.

• Solution design: architecture, model choice and the reasoning behind it.
• Evaluation suite and baseline results.
• Production code:
  • RAG pipelines
  • agents (LangGraph)
  • model serving infrastructure as code
• Observability dashboards for quality, latency and cost.
• Security and data-flow review of the AI feature.
• Runbooks and handover sessions.

Aligned with: the OWASP Top 10 for LLM Applications, the NIST AI Risk Management Framework, and provider responsible-use guidelines.

Tools:

• Agents and orchestration: LangGraph, LangChain, LlamaIndex
• Tracing and evals: LangSmith / Langfuse, Ragas, promptfoo
• Model serving and platforms: vLLM, Hugging Face, AWS Bedrock, SageMaker, Google Vertex AI
• Vector search: pgvector, OpenSearch
• Pipelines and models: Python and TypeScript, MLflow

We offer fixed-price packages for businesses with different needs. Scope, timeline and deliverables are agreed in writing after the scoping call.

Book a 30-minute call to talk through your product and what you need.

Book a call