I build AI systems that ship: multi-agent workflows, RAG pipelines, and production LLM apps. 7+ years building and shipping ML systems, from edge devices to cloud APIs.
- 🏢 Currently AI Strategy & Enterprise Architect at GETEC, building a group-wide enterprise AI platform on Azure API Management, LiteLLM and Azure AI Foundry
- 🎓 MSc Web & Data Science at the University of Koblenz
- 📍 Koblenz, Germany, open to relocation · 🗣️ English (C1), German (A2)
- 📦 Maintainer of 6 open-source Python libraries, and I implement RAG research papers (HANRAG, KERAG, Agent-UniRAG)
Finds why your LLM prompt cache misses and fixes it. Measures cache hits per call site, points at the exact byte where a prompt stopped repeating, and lays prompts out so they cache. 76–88% lower input cost against a naive app across Anthropic, OpenAI, Gemini and LiteLLM. Zero runtime dependencies. Docs
🐦 Roost 
A Postgres-backed job queue for Python. No Redis, no RabbitMQ, just your existing Postgres: transactional enqueue, SKIP LOCKED, LISTEN/NOTIFY, crash recovery. 15,289 jobs/sec on a single async connection. Live dashboard in roost-web. Site
Trainable LLM agent framework for unified single-hop and multi-hop QA, from the arXiv paper. Five-node LangGraph state machine with ChromaDB, FastAPI streaming and a Next.js UI.
🕸️ KERAG
Knowledge-Enhanced RAG (arXiv:2509.04716): SPARQL knowledge-graph retrieval, LLM relevance scoring and Chain-of-Thought answers for complex multi-hop QA.
📚 HANRAG
Heuristic Accurate Noise-resistant RAG (arXiv:2509.09713) for robust retrieval under noisy, multi-hop queries.
Cost-aware routing and provider switching for LLM calls. Docs
Generate Postman collections straight from FastAPI apps.
Speech recognition on Hugging Face with word-level timestamps and speaker diarization.
| Role | Company | When |
|---|---|---|
| AI Strategy & Enterprise Architect | GETEC (remote, Germany) | Oct 2026 – present |
| Machine Learning Engineer | DASS-X, BASF Group (remote, Germany) | Mar 2026 – Sep 2026 |
| AI Python Engineer | Schöne Neue Kinder (Munich, remote) | Oct 2025 – Mar 2026 |
| ML Research Assistant | FZI Forschungszentrum Informatik (Karlsruhe, remote) | Jul 2024 – Jun 2025 |
| Machine Learning Specialist | GagaMuller (Dublin, remote) | Feb 2022 – Apr 2024 |
| AI Software Engineer | Data Pilot (Lahore) | Sep 2019 – Jan 2022 |
Some of what came out of it:
- GameMind Agent: LangGraph assistant for game developers with Postgres vector memory, +65% idea recall accuracy
- Multi-agent Slack bot: offers, lead analysis and follow-ups across Slack, Gmail and Jira, 65% less manual work, 50% faster responses
- License-plate anonymization for self-driving: fine-tuned SAM on German plates, accuracy 75% → 95%
- Multimodal RAG chat with text-to-image and text-to-video diffusion, 30% lower generation latency
- NoiseVu: Wav2Vec2 sound classification on a Raspberry Pi for crime prevention, accuracy 70% → 92%
- LLM-Dispatcher: the smart way to route your AI requests · Oct 2025
- Building HANRAG: making RAG systems actually work for complex questions · Oct 2025
- LangSAM on Replicate: image segmentation with natural language · Jun 2025
- Transforming text into video: an AI-powered approach · Apr 2024
- WhisperX Space: elevating audio transcription · Sep 2023
- fastapi2postman: automate your FastAPI-to-Postman conversion · May 2023
Looking for someone for Agentic AI, RAG, or production LLM systems? Reach me at ashhadahsan@gmail.com, on LinkedIn, or through ashhadahsan.com.



