Automotive and industrial AI systems, built and shipped as working software rather than notebooks — fault diagnosis, predictive maintenance, anomaly detection, and control.
Every project below verifies its own results, and publishes the ones that aren't flattering. DocuBot's eval harness was rewritten mid-project after file-level scoring was found to hide a total retrieval failure; pdm-lite states plainly that its 100% test accuracy is expected on a clean benchmark and the pipeline is the actual achievement; rag-enterprise-assistant reports a RAGAS faithfulness score of 0.675 rather than a rounder, better-looking number. A hiring engineer can't fake-check a demo, but they can check a claim — so the claims here are measured, not asserted.
- DocuBot — German-language RAG agent over automotive/technical documents
- pdm-lite — bearing-fault predictive-maintenance API (Swagger UI)
Both are free-tier deployments and cold-start slowly on the first request — that's normal, not broken.
| Project | What it does | Stack | Proof |
|---|---|---|---|
| DocuBot | German-language RAG agent over automotive/technical documents | FastAPI · ChromaDB · ONNX int8 · Groq · LangChain | Live demo · image cut 934→251 MB, cold start 17.5→5.0s · 68 tests, CI green |
| pdm-lite | Predictive-maintenance API classifying bearing faults from raw vibration signals | 1D-CNN · ONNX Runtime · FastAPI · Docker | Live demo · CWRU dataset, 923 held-out windows · CI → GHCR |
| patch-cae-anomaly-detection | Patch-based autoencoder for unsupervised industrial defect detection — Master's thesis, Audi AG Production Lab | PyTorch · computer vision | ROC-AUC 0.9745 / 0.9962 vs. DFR baseline 0.801 / 0.940 and PatchCore 0.753 / 0.660 (MVTec avg 0.9787) |
| rag-enterprise-assistant | Document Q&A with hybrid search and grounded citations | FastAPI · pgvector · React/TypeScript · RAGAS | RAGAS: faithfulness 0.675, relevancy 0.875, context precision 0.933, recall 1.000 |
| fleetpulse | Real-time vehicle telemetry pipeline | Kafka · Spark Structured Streaming · PostgreSQL · Grafana | CI green · dead-letter path proven by deliberately corrupting 2% of events |
| chartersense | Natural-language analytics over shipping data via a versioned semantic layer | FastAPI · PostgreSQL · LLM-to-SQL | LLM's SQL answer diffed against independently computed ground truth · read-only DB role |
Ich entwickle KI-Systeme für die Automobil- und Fertigungsindustrie – von der Fehlerdiagnose über prädiktive Wartung bis zur Anomalieerkennung. Dabei ist mir wichtig, dass jedes System seine eigenen Ergebnisse überprüft, statt sie nur zu behaupten: DocuBot etwa ist ein deutschsprachiger RAG-Agent für technische Dokumente mit einem eigens entwickelten Auswertungsverfahren und einer echten Live-Demo. Ich freue mich über den Austausch mit Teams im DACH-Raum, auf Deutsch oder Englisch.
- Email: prennoy.babu@gmail.com
- LinkedIn: linkedin.com/in/prennoybabu