M.S. Scientific Computing @ Universität Heidelberg
NLP · Deep Learning · Generative AI · Efficient ML Systems
I build machine learning systems that sit at the intersection of mathematics, research, and engineering.
My current focus is on understanding how modern NLP and deep learning models behave under real-world constraints: noisy inputs, limited compute, inference latency, retrieval quality, and evaluation reliability. I enjoy working on systems where the goal is not only to make a model work, but to understand why it works, when it fails, and how to make it more robust.
- Portfolio: ameynarwadkar.github.io
- LinkedIn: linkedin.com/in/amey-narwadkar-474332231
- GitHub: github.com/ameynarwadkar
I work on methods that make language models faster and more reliable, including early-exit inference, exit-aware verification, calibration, and decision stability analysis.
I’m interested in how neural NLP models behave under noisy or imperfect inputs, especially character-level robustness, typo noise, sentence representation degradation, and evaluation under controlled perturbations.
I build RAG and LLM workflows covering document ingestion, indexing, embeddings, reranking, grounded generation, structured outputs, and evaluation pipelines.
Beyond research prototypes, I like building complete AI applications: computer vision pipelines, generative AI systems, chatbots, and ML-backed automation tools.
- Efficient NLP: early-exit inference, dynamic computation, latency-aware model design
- Robustness & Evaluation: noisy inputs, calibration, failure analysis, adversarial prompts
- Representation Learning: embedding geometry, contrastive learning, sentence representations
- Generative AI: diffusion models, multimodal systems, controlled generation
- RAG & LLM Engineering: retrieval quality, hallucination reduction, tool use, structured outputs
| Project | Area | Description |
|---|---|---|
| Early-Exit Inference for BERT | Efficient NLP | Early-exit strategies for BERT using entropy, margin, and patience-based halting, with latency and calibration analysis |
| Character-Aware Encoder Under Typos | NLP Robustness | Evaluation of sentence representations under synthetic typo noise using CANINE, SBERT, cosine similarity, and Retrieval@1 |
| Tennis Analysis System | Computer Vision | End-to-end tennis video analysis with player detection, ball tracking, court keypoint estimation, speed metrics, and mini-court visualization |
| Text-to-Image Generation | Generative AI | Stable Diffusion-based text-conditioned image generation implementation |
| Food Ordering Chatbot | NLP / Conversational AI | Conversational ordering interface with natural language interaction and user-flow handling |
| Sentiment Trading Bot | NLP / Finance | News sentiment-driven trading pipeline with signal generation, backtesting, and execution logic |
| ML Algorithms from Scratch | ML Foundations | Core machine learning algorithms implemented from first principles using Python and NumPy |
Python · SQL · R · JavaScript · HTML/CSS · Bash
PyTorch · TensorFlow · scikit-learn · Hugging Face Transformers · OpenCV · NumPy · pandas
LangChain · LlamaIndex · FAISS · Pydantic · Ollama · Prompt Engineering · Structured Outputs · Evaluation Pipelines
FastAPI · Docker · Kubernetes · Linux · Git · Streamlit · CUDA · Google Cloud
I am currently pursuing my Master’s in Scientific Computing at Universität Heidelberg, where I focus on advanced AI methods, efficient NLP, applied mathematics, and computer vision.
Before that, I completed my Bachelor’s in Mathematics at Fergusson College, Pune. That mathematical background strongly shapes how I approach machine learning: I care about the assumptions, failure modes, optimization behavior, and evaluation design behind every model.
I’m also working as a Working Student in Software Development at NEC Laboratories Europe, supporting applied AI/ML research through experimentation, data analysis, prototype implementation, and reproducible research workflows.
I like building AI systems that are:
- Grounded: outputs should be traceable, evaluable, and explainable where possible
- Efficient: good models should also respect latency, compute, and deployment constraints
- Robust: systems should be tested beyond clean benchmark settings
- Useful: research ideas should eventually become working, reproducible software
I'm always open to conversations around NLP, RAG systems, generative AI, efficient inference, and applied ML research.
- Portfolio: ameynarwadkar.github.io
- LinkedIn: Amey Narwadkar