🤖 AI Applications

Applied AI in production

I apply AI as an engineer — integration, evaluation, cost/latency tradeoffs, and shipping systems that organisations actually depend on. Not demos.

Approach

How I think about AI

The hardest part of production AI is not picking a model — it's knowing when the answer is wrong. My approach starts with evaluation before architecture: define the failure modes first, then design the system to catch them.

I hold IBM's RAG & Agentic AI and Generative AI specialisations, and I've applied these methods on real industrial data at WEISS GmbH — not on benchmark datasets.

RAG before fine-tuning

For most retrieval problems, RAG over a well-chunked index outperforms fine-tuning and is far cheaper to update when data changes. Fine-tune only when latency or domain vocabulary make it necessary.

Eval before architecture

Define the failure modes, ground-truth examples, and acceptance criteria before writing pipeline code. Retrofitting evals is three times the work.

Cost and latency are features

A system that costs €4,000/month in inference or takes 8 seconds per query is not production-ready. I model total cost of ownership from day one.

Guardrails and safety

Every LLM integration I ship has output validation, fallback paths, and logging. Hallucinations in industrial contexts are not an acceptable UX failure.

When NOT to use AI

A deterministic rule or a SQL query is better than an LLM call for structured lookups. I recommend AI where it genuinely outperforms the simpler option — not as a default.

Own the full stack

Embedding, retrieval, prompt, inference, output parsing, API exposure — I build and own the complete pipeline, not just the "AI part".

Projects

AI case studies

Each project used the Problem → Role → Approach → Stack → Outcome structure.

Case Study · RAG Pipeline

Retrieval-Augmented Generation for industrial document Q&A

Problem: Engineers spent hours searching unstructured internal manuals and spec sheets for answers that should take seconds.

Role: Sole engineer — architecture, implementation, evaluation, deployment.

Approach: Chunked and embedded the document corpus, built a retrieval layer with semantic + keyword hybrid search, wired it to an LLM inference endpoint, added citation tracking so every answer is traceable to a source document.

Outcome: Query resolution time cut significantly; engineers can now self-serve answers without interrupting SMEs.

PythonLangChainRAGFastAPIPostgreSQLOpenAI

Case Study · LLM Fine-tuning

Domain-adapted model for product classification and description generation

Problem: A rule-based system for classifying industrial components and generating product descriptions was brittle, required constant manual updates, and produced inconsistent output.

Role: Lead engineer — dataset curation, fine-tuning pipeline, eval framework, integration.

Approach: Curated a training dataset from existing product data, fine-tuned a multimodal model for both text and image inputs, built an eval suite comparing output quality against the old system and human reviewers.

Outcome: Replaced the rule-based system; manual review load reduced substantially; description quality rated higher by product managers.

PythonLLM Fine-tuningMultimodalLangChainIBM watsonx

Case Study · Agentic AI

LLM agent for automated data extraction and workflow routing

Problem: Incoming customer requests arrived in free-text form and required manual reading, classification, and routing to the correct internal team — a time-consuming and error-prone process.

Role: Architect and implementer — agent design, tool definitions, guardrail layer.

Approach: Built an agentic pipeline using LangChain that reads incoming requests, extracts structured fields, classifies intent, and routes to downstream systems — with a human-in-the-loop escalation path for low-confidence cases.

Outcome: Automated the majority of routine routing; escalation rate and misrouting both declined.

PythonLangChainAgentsFastAPIAWS

Credentials

AI certifications

🏅

IBM RAG & Agentic AI Specialisation

Retrieval-Augmented Generation, agentic architectures, tool use, evaluation methodology — applied with IBM watsonx.

🏅

IBM Generative AI Specialisation

Foundation models, prompt engineering, fine-tuning, responsible AI, and production deployment patterns.