RL on humanoids, detectors running in a browser tab, RAG in production, and agents that watch my training runs while I sleep — research-grade when it needs rigor, product-grade when it needs to ship. The loss curve in the corner is your reading progress. yes, it converges.
I'm an AI researcher-engineer doing my M.Sc. in Computer Science (AI) at the University of Freiburg — after a B.E. in Computer Science finished at GPA 9.33/10.
Hand me a strong paper and I'll rebuild it, then push past it. Hand me a vague problem and I'll scope it, build the pipeline, and ship the unglamorous parts too — data, evals, deployment, automation. The whole stack of making models useful, not just the fun layer.
Previously: ML intern building production RAG systems at WiZdom Ed. Currently: coursework in deep learning, PGMs and robot mechanics, plus the 2026 research roadmap below. English C2 · Hindi native · German A2→B1. Off the clock: over-engineering n8n automations for my own life and defending masala chai against German filter coffee — a study with n=1 and strong priors.
Reproduction and extension of RLPD (Ball et al., ICML 2023) in PyTorch with Minari offline data: symmetric 50/50 sampling, LayerNorm critics, large ensembles at high UTD — reproduced on three MuJoCo suites, then pushed to Humanoid-v5.
Fine-tune an NMS-free YOLO26, ship the same network to TensorRT (RTX 5070), ONNX Runtime (Ryzen 7700) and WebGPU in the browser, then measure every path with MLPerf-style rigor + NVML power. On Blackwell, FP8 hits 560 FPS and FP16 wins latency-per-watt — while INT8, the reflex default, is dominated on accuracy, speed and power.
A ~50M-param hybrid LM interleaving Mamba-2 SSM blocks with causal attention (the Jamba pattern), trained on OpenWebText at matched tokens-seen. Sweeping the attention:SSM ratio — 1:7 leads the reduced-scale preview; KV-cache and inference columns land next.
Training sparse autoencoders over the residual stream of a small open LM to decompose activations into monosemantic features, then circuit-tracing induction behaviour. Early dictionaries hit ~78% auto-interp on probed layers.
Albert-Ludwigs-Universität Freiburg — deep learning, probabilistic graphical models, statistical pattern recognition, robot mechanics.
N.M.A.M. Institute of Technology — GPA 9.33/10 (German equivalent 1,3).
Research collaborations, working-student roles, and problems that are interesting at 2am. Based in Freiburg — I usually reply before the next training run finishes.
ckpt 02 — two shipping, two brewing
ckpt 03 — a menu, not a mandate
ckpt 04 — the receipts
ckpt 05 — daily drivers first