{
"name": "Amey Narwadkar",
"location": "Heidelberg, Germany",
"education": {
"masters": "Scientific Computing @ Universitat Heidelberg",
"bachelors": "Mathematics @ Fergusson College, Pune"
},
"current_role": "Working Student @ NEC Laboratories Europe",
"focus_areas": [
"Multi-Agent Systems",
"RAG & LLM Systems",
"Applied AI Systems",
"Efficient NLP"
],
"status": "Open to AI/ML roles"
}I build ML systems at the intersection of math, research, and engineering. I care about making models robust, efficient, and actually useful in production.
|
Python |
PyTorch |
FastAPI |
Docker |
Kubernetes |
GCP |
Linux |
|
PostgreSQL |
Redis |
React |
Next.js |
TypeScript |
Git |
Bash |
|
OpenCV |
sklearn |
R |
HTML |
CSS |
Supabase |
VS Code |
|
Production-grade, end-to-end RAG system featuring a two-stage retrieval funnel (hybrid search and cross-encoder reranking), conversation history rewriting, semantic caching, and verifier-backed citation constraints.
|
A 7-agent system orchestrated via Google ADK that searches ArXiv, PubMed, and OpenAlex. It clusters papers, rigorously assesses evidence quality, and synthesizes publication-ready literature reviews.
|
|
Implemented early-exit strategies for BERT using entropy, margin, and patience-based halting. Integrated micro self-verification and conducted latency & calibration analysis.
|
End-to-end tennis video analysis pipeline featuring YOLO-based player detection, CNN ball tracking, court keypoint estimation, real-time speed metrics, and a dynamic mini-court visualization.
|
|
Building autonomous agent teams that collaborate to solve complex tasks, from research synthesis to automated workflows. |
Production-ready retrieval pipelines with hybrid search, reranking, structured outputs, and rigorous evaluation. |
|
End-to-end ML applications: computer vision, conversational AI, sentiment analysis, and automation pipelines. |
Making language models faster and cheaper with early-exit inference, calibration, and compute-aware optimization. |
