Case study
LLM Consulting at Artefact
From prototype chatbots to grounded retrieval pipelines — how I design and ship LLM solutions in consulting contexts at Artefact.
Problem
Enterprise clients often want 'AI' without a clear path from unstructured documents, legacy APIs, and compliance constraints to something operators can trust. Hallucinations, weak grounding, and brittle prompt-only demos fail the moment real users ask domain-specific questions.
Approach
I start by mapping the decision the system must support — not the model. Then I design retrieval (chunking, metadata filters, hybrid search), evaluation harnesses (faithfulness, citation coverage, latency), and orchestration with LangChain / LangGraph or Google ADK depending on the stack. On AWS Bedrock and GCP we wire guardrails, logging, and human-in-the-loop review where risk is high. Consulting work means iterating with stakeholders: thin vertical slices, measurable acceptance criteria, and honest trade-offs between cost, latency, and answer quality.
Result
Shipped LLM-powered tools that answer from client corpora with citations, reduced time spent hunting across PDFs and tickets, and established reusable patterns for RAG evaluation and agent workflows that later projects could reuse. The biggest win is operational: teams can see why an answer was produced and when the system should refuse.
At Artefact (formerly Effixis), I work as a senior Data Scientist and LLM Engineer on consulting engagements where language models have to earn trust in messy, real environments.
Most requests arrive as a product wish — a chatbot, an assistant, an “AI layer” on top of internal knowledge. My job is to turn that into a system with retrieval, evaluation, and clear failure modes. I care less about flashy demos and more about whether an analyst, support agent, or compliance reviewer can rely on the output under time pressure.
The work spans retrieval-augmented generation, agent-style orchestration, and the unglamorous plumbing: document ingestion, access control, observability, and cost control. First person, day to day: I write the pipelines, sit in discovery workshops, and keep pushing until grounded answers beat confident guesses.