Automating a task on its own rarely moves anything. I rebuild the workflow around the system, and your team ends up owning it. Every engagement starts with a process map, not a tool.
No 18-month programme. No rip-and-replace. Built on the stack you already have.
Book a 30-minute discovery callMap the workflow end-to-end with the people who actually do it. Not the org chart. Not the BPMN diagram. The real path. Where decisions happen, where data flows, where time is lost.
Score automation candidates by impact and feasibility, not novelty. The first deployment should pay for the program, and it usually isn't the most exciting workflow.
Before any agent, build the context: knowledge graph, retrieval, tool access, identity, audit. AI inherits the quality of the context layer beneath it. Skip this and pilots fail predictably.
Ship into the workflow that was mapped. Observability, evals, fallbacks, governance. No standalone copilot apps that no one opens twice.
The 30/60/90 days after go-live decide whether the system sticks. Eval regressions, decision-log reviews, capability handover, until your team runs it without me.
Switzerland's first production clinical LLM platform, live in Swiss hospitals, shipped in under 90 days.
Swiss legal AI in production since 2021, including court anonymisation for the judiciary.
What I have shipped to production. Other classes are possible, but here is where I can judge fit and risk quickly.
Multi-step agents with the whole system around the model: scoped tools, cross-model verification, memory, guardrails, trajectory evals. They act in your systems of record and hand back to a human on uncertainty.
Document, contract and policy retrieval with citation grounding, governance and audit. Systems that lawyers, auditors and analysts actually use day-to-day.
Inbound and outbound voice with human hand-off, conversation evaluation and compliance recording.
On-prem or VPC inference for open-weight models, custom fine-tuning, model routing and audit-grade observability. For workflows where data cannot leave the building.
Healthcare, legal, financial and public-sector contexts where eval gates, audit trails and provenance are built into the architecture, not bolted on.
Feature design, eval suites, telemetry and gradual rollout inside a product your users already have. No separate AI tab nobody clicks.
Building agents that survive day two is its own discipline: tools, verification, memory, guardrails, observability. You rent the model and build the system around it. Read the field guide to harness engineering.
Everything above still applies here: RAG, agents, voice, the rest. Where regulation requires it (FADP, GDPR, EU AI Act, Swiss Berufsgeheimnis, public-sector procurement), I deploy the sovereign and private-AI variant: custom fine-tuning where off-the-shelf models cannot see the data, on-prem or VPC inference, audit-grade provenance from day one. The strategy and the implementation come from the same hands. The references below are systems I shipped.
Shipped in productionFine-tuning small NER and language models on your sensitive entities, your terminology, your edge cases. Production accuracy where general-purpose APIs cannot match the domain, and never see the data. Reference: 97%+ accuracy on PII detection in Swiss clinical text, zero data persistence (DocIQ Shield + KISIM AI).
Open-weight models hosted in your data centre, your private cloud, or air-gapped infrastructure. GPU sizing for the actual workload. Model routing across cost/latency/quality tiers. No outbound calls to vendor APIs unless you explicitly choose to.
Every model output traces to inputs, model identifier, prompt version, retrieval sources, and timestamp. Regulator-ready trails for healthcare, legal, financial and public-sector deployments. Designed in from day one, not bolted on.
Continuous regression suites, quality/safety/latency/cost gates, drift detection, observability across both proprietary and fine-tuned model paths. Confidence to ship and confidence to keep shipping.
AI systems fail in operations, not in demos. Build for the Tuesday afternoon, not the board meeting.
The strategy session is the on-ramp, fully creditable against any follow-on engagement. Fixed-fee on audit and system build, retainer on fractional lead.
A focused half- to full-day working session, on-site or remote. One specific question: vendor selection, build-vs-buy, architecture review, "is this the right AI use case." Written notes and a clear next-step recommendation by end of day. Fee fully creditable against any follow-on engagement within 90 days.
2-3 weeks. Map workflows end-to-end, score automation candidates by impact and feasibility, propose 3-5 prioritised initiatives with ROI estimates and a 90-day plan.
Production deployment of one of the system classes (agentic systems, context layers, voice, knowledge platforms, private + on-prem AI). Requirements through go-live with operations and governance handover. Quoted firmly after the audit, not before.
1-2 days/week. Program governance, vendor selection, hands-on architecture review, internal capability building. For organisations not ready to hire a full-time AI lead but unwilling to outsource the strategy.
When engagements fail, it traces to one of these six missing. They are not polish; they are the architecture.
No engagement starts without a named sponsor with budget and decision authority. AI projects without sponsorship turn into demo theatre.
Every system carries an eval suite from day one. No production deployment without measurable success criteria the system has demonstrably met.
Every AI output traces back to its inputs, model, prompt, and version. Audit-grade by default, including in non-regulated environments.
Automation depth matched to consequence. Reversible decisions can be fully automated. High-consequence decisions get review gates by design.
Engagement ends when the internal team owns operations. I leave behind documentation, runbooks, and capability, not dependency.
Architecture preserves the ability to swap models, retrieval engines, or tooling. Lock-in is sometimes the right trade, but it is always a deliberate choice.
One call to agree whether a process audit, a system build or a fractional AI lead fits where you are.