Open to AI Engineer / LLM roles · Munich, Germany
I build production-grade LLM systems, RAG pipelines, and agentic workflows that ship to real users. Currently shipping DocGPT at CHECK24 — serving 80K+ weekly API requests across multiple verticals — after completing my MSc in Applied Data Science at the University of Göttingen.
01 — About
I'm an AI Engineer with 5+ years of project experience across LLM applications, RAG, fine-tuning, MLOps, and large-scale fraud / propensity modelling. I love the full lifecycle — from data curation and retrieval design to fine-tuning, evaluation, and shipping behind a FastAPI service with proper monitoring.
Right now I'm at CHECK24 building the DocGPT RAG platform — multi-format document support, hybrid retrieval on Weaviate, a DAG-based agent-workflow engine, and OCR pre-processing for the Autoteile vertical. I previously fine-tuned LLaMA-3 and Gemma with LoRA + RAFT for a German legal QA platform at Fraunhofer IAIS, where I also built real-time monitoring dashboards (Grafana, TimescaleDB, PostGIS) for autonomous-driving telemetry.
My completed MSc thesis investigated Recurrent Chain-of-Thought reasoning in Diffusion-based LLMs — studying iterative latent reasoning on tasks like Countdown and GSM8K against autoregressive baselines.
{
"name": "Muhammad Rehan Khalid",
"role": "LLM / Agentic AI Engineer",
"location": "Munich, Germany",
"current": "AI Engineer @ AI Forge, CHECK24",
"education": "M.Sc. Applied Data Science, Göttingen",
"focus": [
"LLM Fine-tuning",
"Retrieval-Augmented Generation",
"Agentic Workflows",
"Diffusion LLMs",
"MLOps"
],
"stack": [
"PyTorch", "Transformers", "LangGraph",
"LlamaIndex", "Weaviate", "FastAPI",
"Docker", "Kubernetes", "AWS"
],
"languages": [
"Urdu (Native)",
"English (Professional Working)",
"German (Limited Working)"
],
"work_permit": "Valid (Germany)",
"notice_period": "30 days",
"status": "open_to_work"
}02 — Experience
DocGPT — AI document extraction & chat platform deployed across CHECK24 verticals.
German Legal AI Assistant & Autonomous Driving Signal Monitoring — grounded legal QA over 70GB / 30K+ documents, plus real-time telemetry observability.
GenAI Chatbot & AI Fraud Detection Platform — production LLM chatbot and real-time Card-Not-Present fraud ML.
Customer Propensity Modelling Platform — customer ranking and experimentation engine for revenue optimisation.
03 — Featured Projects
Recurrent Chain-of-Thought reasoning inside a diffusion-based LLM. Iteratively denoises latent reasoning traces over Countdown, GSM8K, and other reasoning benchmarks, studied against autoregressive CoT baselines.
Hybrid retrieval and ReAct agent over Instagram post data. Fuses ChromaDB semantic search (Ollama bge-m3 embeddings) with BM25 keyword search via Reciprocal Rank Fusion, served behind a FastAPI tool-calling endpoint that returns grounded answers with post IDs, URLs, and tool traces.
Automated YouTube brand-compliance auditing on LangGraph. Downloads a video, extracts speech and on-screen text with Azure Video Indexer, retrieves matching rules from an Azure AI Search RAG index, and generates a structured compliance report with Azure OpenAI GPT-4o.
Multi-vertical document extraction and chat platform across 4 CHECK24 verticals. Multi-format document ingestion, hybrid (keyword + vector) retrieval on Weaviate with optimized chunking, metadata filtering and re-rankers, and a real-time observability dashboard.
Scan-quality agent for the Reifen vertical: dewarping, deblurring, shadow / lighting correction, and content-aware routing. Improves downstream OCR + LLM extraction accuracy on noisy real-world scans.
End-to-end legal QA stack on 70GB / 30K+ documents: LLaMA-3 / Gemma fine-tuned with LoRA on 23,905 QA pairs (36.8 → 56.4% accuracy), RAFT for factuality (+20% correctness), multi-step RAG with re-ranking, and automated LLM-as-a-Judge evaluation.
Geospatial telemetry monitoring on TimescaleDB + PostGIS for autonomous driving. Real-time Grafana dashboards over GPS, RSU status, and network-quality signals — query latency cut by ~40% via optimized TimescaleDB / PostGIS queries.
Gradient-boosted (XGBoost) models for real-time Card-Not-Present transaction fraud — contributing to ~$40M estimated fraud savings, a 20–30% stronger fraud signal, and feature-logging optimised from 5 h → 10 min to unblock daily retraining.
04 — Technical Skills
05 — Education
Georg-August-Universität, Göttingen, Germany
Thesis (completed): Recurrent Chain-of-Thought reasoning in Diffusion-based LLMs — iterative latent reasoning on Countdown, GSM8K, and reasoning benchmarks.
Lahore University of Management Sciences (LUMS), Lahore, Pakistan
Quantitative foundation across mathematics, statistics, and economics — the backbone of my ML work. Graduated with Distinction; Dean's Honor List.
06 — Contact
I'm open to AI Engineer, LLM Engineer, ML Engineer, and Applied Scientist roles in Germany / EU — onsite, hybrid or remote. Notice period: 30 days. Valid German work permit.