Applied AI Engineer · Lausanne, Switzerland

Applied AI Engineer building production LLM systems that are observable, testable, and cost-aware.

Six years taking ambiguous AI problems from discovery and architecture through deployment and adoption.

Swiss Permit B · No sponsorship required

View selected work LinkedIn

6 yearsshipping applied AI

10+ clientsfrom discovery to adoption

70%lower latency

50%lower cost

Hours → minutesfor production debugging

Selected systems

From uncertain brief to reliable production behavior.

Selected work is anonymized because it was delivered for private clients and employers.

01 / Industrial AI

Production Agentic AI for Industrial Operations

Turned an open-ended automation brief into a production agent architecture with structured tool use, measurable quality, and operational visibility.

System notes

Problem: High-value operational workflows required contextual reasoning across tools and data, with little tolerance for opaque failures.
Contribution: Led architecture and delivery: tool contracts, evaluation strategy, tracing, reliability controls, and adoption feedback loops.
System: Agent runtime with structured tools, automated evaluations, end-to-end traces, fallbacks, and token-level cost observability.
Result: Reduced latency by 70%, cost by 50%, and production debugging from hours to minutes.
Tech: Python, LLM APIs, structured tool calling, evaluation pipelines, distributed tracing, containerized services.

02 / LLM engineering

LLM Evaluation & Prompt Optimization Platform

Built the machinery to compare prompts, models, and optimization strategies as repeatable engineering experiments rather than one-off demos.

System notes

Problem: Prompt changes were difficult to compare fairly across datasets, evaluators, remote compute, and token budgets.
Contribution: Designed evaluator contracts, job provisioning, the mutation pipeline, reproducible datasets, and optimization workflows.
System: Asynchronous evaluation jobs, modular evaluators, prompt mutation, remote workers, state-machine optimization, and leaderboard datasets.
Result: Made quality, latency, and token cost visible in one reproducible comparison loop.
Tech: Python, LLM APIs, asynchronous workers, structured evaluation, experiment tracking, containerized compute.

03 / Agent systems

Multi-Agent Knowledge & State Platform

Shaped a coordinated agent system in which context, knowledge, and state remain explicit as workflows become more complex.

System notes

Problem: Multiple agents needed relevant context without uncontrolled retrieval, duplicated observations, or hidden state transitions.
Contribution: Implemented coordinator and context-builder flows, source budgets, graph-knowledge contracts, shared prompt configuration, and deduplication.
System: Coordinator-led agents, RAG sources, graph knowledge base, state machine, source-level context budgets, and observation deduplication.
Result: Created clearer ownership of state and more predictable context assembly across agent workflows.
Tech: Python, agent orchestration, RAG, graph data, state machines, prompt configuration, service APIs.

04 / Product analytics

Natural-Language Analytics for User Activity

Converted noisy event streams into explainable user actions and made the resulting behavioral data accessible through natural language.

System notes

Problem: Raw activity logs obscured meaningful behavior, while analysts still needed reliable access to detailed event data.
Contribution: Developed cleaning and action classification, aggregation, motifs, clustering, change-point analysis, LLM annotation, and Text-to-SQL experiments.
System: Event processing, behavioral representations, similarity analysis, tool-calling agents, schema linking, SQL generation, and error correction.
Result: Produced interpretable activity patterns and a practical path from natural-language questions to structured analytics.
Tech: Python, SQL, clustering, time-series methods, embeddings, tool calling, Text-to-SQL.

05 / Knowledge systems

Knowledge-Graph Agent for Expert Documents

Built an end-to-end document-to-graph workflow so domain experts could inspect, version, and export machine-extracted knowledge.

System notes

Problem: Expert documents held valuable relationships, but extraction needed to be inspectable and usable beyond an LLM response.
Contribution: Implemented document and URL ingestion, graph extraction, versioning, visualization, export, and pre-run cost estimates.
System: Multi-source ingestion, structured graph extraction, persistent versions, interactive graph review, downloads, and token-cost estimation.
Result: Delivered a traceable workflow from source document to reviewable, reusable knowledge artifact.
Tech: Python, LLM APIs, graph data, document parsing, interactive data apps, Docker, persistent storage.

Additional work

The supporting systems matter too.

Multichannel orchestration

Early conversational service architecture with asynchronous model calls, message history, summarization, and containerized service communication.

Enterprise GenAI adoption

Proof-of-concept work that connected technical feasibility with stakeholder discovery, workflow fit, and a credible route to production.

Secure research infrastructure

Reliable environments for applied AI experimentation, with reproducible services, controlled data access, and operational handover in mind.

What I bring

Architecture, evaluation, and delivery for AI that has to work outside the demo.

I move between product discovery and implementation: decomposing uncertain workflows, building agent and retrieval systems, defining evaluation loops, and making quality, latency, cost, and failure modes visible.

Connect on LinkedIn