Services
Generative AI consulting services for B2B operators who need shipped systems, not slide decks. We score use cases against revenue impact, pick the right LLM per task on an actual eval harness, build RAG and prompt pipelines that survive contact with real users, and ship MLOps that your team can own. Senior engineers who have run production LLM workloads at scale — not prompt-engineers reading Twitter. Discovery sprints from 12,000 EUR, working pilots from 45,000 EUR, production retainers from 18,000 EUR per month.
Most generative AI projects fail for the same three reasons. The wrong use case — a chatbot replacing a search bar nobody used. The wrong eval — "looks good on five examples" until production users find the sixth. The wrong architecture — a 14-step LangChain agent where two function calls would have shipped. We start with a written ROI model and a 200-item eval set before a single line of orchestration code. We pick LLM providers on measured latency, quality, and cost per 1k requests at your traffic mix — not on the demo that went viral last week. By week 10 you have a working system, a regression harness, observability, and a runbook your team owns.
We interview product, ops, and support, then score 8 to 15 candidate use cases on revenue impact, build cost, feasibility, and risk. You get a ranked shortlist, a written ROI model per top-three, and a clear "do not build this" list with reasoning.
Eval-driven selection across OpenAI, Anthropic, Bedrock, and Vertex. We measure quality, p50/p95 latency, and cost on a 150 to 400-item task-specific eval set. Output is an ADR with the chosen model, a fallback model, and the trigger to re-evaluate.
Corpus ingestion, chunking strategy calibrated to your document distribution, embedding model selection, vector store sizing, hybrid retrieval. We size for your actual corpus growth rate, not a default 100k-vector demo.
Versioned prompts in git, regression eval sets that run in CI, Ragas and DeepEval rubrics, LLM-as-judge with human spot-check. We refuse to merge prompt changes that regress a tier-1 metric — even our own changes.
PII stripping at ingress (Presidio or fine-tuned classifier), zero-retention provider contracts, EU endpoints for EU data, egress scans for hallucinated PII. DPAs and sub-processor lists aligned to your customer contracts.
Observability through LangSmith, Langfuse, or Helicone. Per-request logging of prompt version, model, tokens, latency, cost. Cost alerts, latency SLOs, automated A/B prompts, and a written runbook for model upgrades and provider outages.
Weeks 1–3: stakeholder interviews, use-case scoring, ROI model, eval set design. Output is a ranked shortlist plus an architecture proposal that the founder and the board can read.
Weeks 4–5: build the eval harness, run candidate models on the task-specific set, write the ADR with chosen model, fallback model, and re-evaluation trigger. Prompts checked into git.
Weeks 6–10: end-to-end system, RAG or agent orchestration, observability, PII handling, customer-zero deployment behind a feature flag. Regression evals running in CI before any prompt merge.
Weeks 11+: expand the eval set, add fallbacks, train your team on the runbook, set cost and latency SLOs, run the first quarterly model-upgrade review. We step out when your team is operating it.
Three weeks. Use-case scoring, ROI model, provider eval design, architecture proposal, written ADRs. Best for teams who do not yet know which GenAI bet is worth making. 12,000 EUR fixed.
8 to 10 weeks. Working end-to-end system, eval harness in CI, observability, PII pipeline, and customer-zero deployment. Knowledge transfer to your engineers built into the timeline. 45,000 EUR fixed.
Monthly. Prompt iteration, model upgrades, eval expansion, cost optimisation, on-call for LLM-specific incidents. Best after pilot ships and you need ongoing senior coverage. From 18,000 EUR/month.
All engagements start with a mutual NDA, IP assignment, and a DPA. Three-month minimum on the production retainer, month-to-month thereafter with 30 days notice.
Single-page Tilda landing with Telegram-bot lead capture for an ad agency — shipped in two weeks, US & EU ready.
A high-throughput loan decision engine on Laravel — automated scoring, credit-bureau integration, and 10x faster decisions for US & EU lenders.
GDPR-aligned · ISO 27001 ready · SOC 2 Type II in progress · HIPAA-capable · CCPA-acknowledged
Every engagement starts with a 150 to 400-item eval set before architecture lock-in. We refuse to ship prompts that have never seen a regression run. Demos are not evidence.
Our LLM leads have shipped production ML before transformers were cool — ranking, classification, search relevance. They argue about latency budgets and Postgres query plans, not Twitter threads.
GDPR, SOC 2, HIPAA, CCPA — we have negotiated zero-retention contracts with OpenAI and Anthropic, written DPAs that hold up in customer reviews, and walked auditors through LLM scope.
We treat LLM provider choice as a quarterly decision, not a religion. When the frontier moves, your evals will tell us — not a vendor's quarterly earnings call.
Provider selection is an eval problem, not a marketing problem. We build a task-specific eval set of 150 to 400 representative inputs with graded golden outputs, then score candidate models on quality (LLM-as-judge plus human spot-check on a 50-item subset), latency at the p50/p95/p99 we actually need, and cost per 1k requests at our token mix. Typical lineup: GPT-4o or Claude 3.7 Sonnet for reasoning, GPT-4o-mini or Claude 3.5 Haiku for classification, gpt-4o-realtime for voice, an open-weights model on vLLM for cost-sensitive batch. We re-run the eval every quarter because the frontier moves.
RAG is right when answers live in a corpus that updates faster than you can fine-tune (policies, product docs, tickets, contracts). It is the wrong answer when the model already knows the domain (general code, general knowledge), when latency is below 500ms p95, or when you need deterministic outputs that fine-tuning produces more reliably. We frequently combine: RAG for grounding in customer data, a small fine-tuned model for output structure or domain vocabulary, and prompt engineering for orchestration. Pure RAG is rare in production. Pure fine-tuning is rarer.
Three layers. At the ingress, a PII detector (Microsoft Presidio or a fine-tuned classifier) strips or tokenises emails, names, phone numbers, and account IDs before the prompt leaves our infrastructure. At the provider, we contract zero data retention with OpenAI, Anthropic, and Bedrock (signed BAAs where applicable) and prefer EU-hosted endpoints for EU data. At egress, output is scanned for hallucinated PII before being shown to the user. DPAs are signed before kickoff, and we maintain a sub-processor list aligned with your customer-facing contracts.
Prompts are code. They live in version control, they are reviewed in pull requests, and they are tagged with semantic versions that get logged on every inference. Each prompt change runs against a regression eval set in CI (Ragas for RAG quality, DeepEval or a custom rubric for task-specific metrics), and we will not merge if any tier-1 metric regresses by more than 2 percent. In production we log prompt version, model, latency, and token cost per request through LangSmith, Helicone, or Langfuse so you can A/B prompts the same way you A/B features.
Yes. Most engagements integrate into an existing backend (Node, Python, Go, Java) and existing infra (AWS, GCP, Azure) rather than running on a separate platform. We do not push you to a vendor-specific orchestrator if you do not need one. We pair with your engineers, run code reviews together, and write ADRs for the architectural calls so the choices outlive our engagement. Knowledge transfer is contractual: by end of the pilot, your team owns the codebase, the evals, and the runbook.
Three tiers. Discovery sprint is 12,000 EUR over three weeks: use-case scoring, ROI model, provider eval, and a written architecture proposal. GenAI pilot is 45,000 EUR over 8 to 10 weeks: working end-to-end system, eval harness, observability, and a customer-zero deployment. Production rollout retainer starts at 18,000 EUR per month and covers prompt iteration, model upgrades, eval expansion, and on-call. Typical path from kickoff to revenue-impacting production is 12 to 16 weeks.
