I am an M.Sc. Computer Science student at the University of Freiburg, focused on reinforcement learning, efficient model inference, and retrieval systems. I like projects with a clear question and results that other people can reproduce and inspect.
Outside research, I love photography, especially nature, animals, and landscapes. Photography makes me slow down and pay attention, which is part of why the field-note style of this profile feels like me.
One detector exported to TensorRT, ONNX Runtime, and WebGPU. FP16 reached 536 FPS at 9.3 FPS/W with a 0.06% mAP drop; FP8 was fastest at 560 FPS. The same nominal INT8 precision produced an 8x accuracy-loss gap across runtimes. |
A PyTorch reproduction with three seeds per method on Hopper, Walker2d, and HalfCheetah. RLPD reached 88.0, 89.6, and 88.6% of the Minari v5 expert return, with lower seed variance than IQL on all three tasks. Repository · 3 methods · 3 seeds · 245k steps |
| Project | What I built or measured | State |
|---|---|---|
| English to Hindi Transformer | A 43M parameter, 6-layer Transformer built directly in PyTorch and trained on 500k Samanantar pairs. Beam search reached 16.93 SacreBLEU and 41.58 chrF++ on 500 frozen test pairs, but cost 9.3x the latency. | Complete |
| Mamba and attention hybrid LM | A roughly 50M parameter ratio study at matched tokens seen. The reduced-scale 1:7 preview leads at 102.4 validation perplexity; inference speed and KV-cache measurements are still open. | In progress |
| UNETR 3D abdomen segmentation | A MONAI pipeline for 14-class CT segmentation with resampling, balanced 128³ crops, and sliding-window inference. Validation Dice was 0.8027 on one split. | Complete |
| Arise | An offline-first fitness RPG built with React 19, TypeScript, Dexie, and optional Supabase sync. Training history drives XP, quests, levels, and progression. | v1 complete |
| RL foundations | NumPy implementations of TD control, value iteration, and the 10-armed testbed. | Complete |
I am finishing the full Mamba ratio study and checking whether its perplexity result survives the measurements that matter at inference: tokens per second and KV-cache growth with context. Alongside that, my Freiburg coursework currently covers Bayesian hyperparameter optimization and multi-criteria optimization.
| When | Work |
|---|---|
| 2025 to present | M.Sc. Computer Science, Artificial Intelligence at the University of Freiburg. Current interests: reinforcement learning, efficient inference, and robotics. |
| 2023 to 2024 | Machine Learning Intern at WiZdom Ed. Built and evaluated a RAG study-path system over 5,000+ documents with LangChain and ChromaDB. A company-provided 100-batch evaluation reported 71.7% Recall@5, 93.4% groundedness, and 89.1% refusal accuracy. |
| 2020 to 2024 | B.E. Computer Science at N.M.A.M. Institute of Technology, GPA 9.33/10. |
My repositories keep the result next to the conditions that produced it: configs, seeds, evaluation artifacts, hardware, and known limitations. Preview runs stay labeled as previews. Single-split validation stays labeled as validation.
I am based in Freiburg, where the Black Forest is never far away, and I am open to research collaborations, working-student roles, and internships in ML research or engineering. The easiest way to reach me is email. More context and project write-ups are on my portfolio.
