I build GenAI systems that go beyond demos. RAG pipelines, multimodal applications, and LLM-integrated products that work under real-world constraints.

5 years of experience across production software engineering and applied AI research. Currently finishing my MSc in Artificial Intelligence at FAU Erlangen. My most recent work was at Carl Zeiss AG, CRT team, where I designed and built a hybrid Bayesian segmentation system for industrial microscopy, combining uncertainty estimation with contrastive learning to separate reliable predictions from edge cases at production scale.

I am most interested in the intersection of generative AI and high-stakes domains: healthcare, industrial AI, and enterprise LLM products where getting it wrong has real consequences.

Carl Zeiss AG RnD, Master's Thesis (unpublished, paper in progress)

In high-throughput industrial microscopy, a segmentation model that confidently misclassifies is worse than one that admits it does not know. The goal of this work was to make uncertainty a first-class output, not an afterthought.

I designed a hybrid Bayesian-deterministic segmentation architecture combining a multi-depth U-Net with contrastive learning and attention-based Bayesian uncertainty estimation. The model produces a unified image-level uncertainty score to separate in-distribution from out-of-distribution samples, enabling reliable decisions at production scale.

Results on 12k high-dimensional microscopic images: Dice 96%, IoU 94%, OOD detection accuracy 95%, false-positive segmentations reduced by 90%. Deployed as an Azure-hosted inference pipeline with live model drift monitoring via W&B.

Work is under NDA. A reproducible public version using open microscopy datasets is in progress.

PyTorch Bayesian Deep Learning U-Net Contrastive Learning Uncertainty Estimation Azure W&B Industrial AI

View Repo

A multimodal pipeline that takes a live camera feed and a voice question, identifies the gym equipment in frame, and retrieves the most relevant instructional video through a RAG system. Designed as a proof-of-concept for XR deployment.

TensorFlow.js object detection runs client-side to avoid streaming raw camera frames to the server. A Whisper STT microservice transcribes the voice query. Phi-3 via Ollama extracts structured entities including equipment name and target muscle groups. Hybrid search on Qdrant then combines metadata filtering with semantic similarity over video transcripts. The full chain runs over WebSockets with under 3 seconds latency end-to-end.

FastAPI Qdrant RAG Whisper Ollama / Phi-3 TensorFlow.js Multimodal WebSockets

View Repo

Tell EasyBuy what you need. It figures out everything else.

Say "I need a skiing kit under 400 EUR" and a multi-agent pipeline decomposes that into every item the kit requires, searches real listings across multiple retail platforms, ranks by price, reviews, and delivery speed, assembles the cart, and delivers a PDF invoice to your inbox. No search bars, no tab switching, no manual filtering.

Five specialized agents each own one stage: intent gathering, kit breakdown with budget allocation, multi-platform product search, transparent ranking with user-controlled filters, and cart assembly with email invoice delivery. Each agent is a dedicated FastAPI endpoint with its own prompt design and structured output schema.

Built in 48 hours at the 4th Hack-Nation Global AI Hackathon, February 2026, VC Track.

FastAPI Google Gemini 2.0 Agents SerpAPI Next.js Docker Agentic AI

View Repo

Industrial sensors produce streams of data. When a fault prediction fires, engineers need to know which sensor actually caused it, not just that something went wrong.

Built an attention-based LSTM where attention weights directly indicate which sensors and timesteps drove each prediction. Compared attention-based explanations against SHAP to validate interpretability quality.

PyTorch LSTM Attention XAI SHAP Industrial AI

MSc Artificial Intelligence, FAU Erlangen-Nürnberg (ongoing)

BTech Information Technology, BVP Pune

🏆 Hackathon Winner, TUM.ai x BKW Engineering Challenge

📄 Published, Attendance monitoring via computer vision, JETIR

AI Engineer or ML Engineer roles in Germany, full-time, hybrid or on-site in the DACH region.

I am most interested in teams building serious GenAI products or applying ML in medical, industrial, or enterprise contexts. I have a valid German work and residence permit and am available immediately.

The kind of work I want: RAG and LLM system design, multimodal pipelines, applied research that ships. Not pure software development. Not tutorial-level AI.

📧 yadavakash1996@outlook.com · LinkedIn