🎮 flavor on the surface · 🔬 details in the foldouts · open the ▸ panels as you go
A self-reported model card, with the measurable parts listed below.
| field | value |
|---|---|
| architecture | curiosity-driven · chai-cooled · stubbornly empirical |
| pretraining | B.E. Computer Science (9.33/10) → production ML internship |
| fine-tuning | M.Sc. Computer Science (AI) · University of Freiburg 🇩🇪 |
| alignment | to measured baselines; vibes are not an eval |
| known limitations | will re-run your experiment with 3 seeds before agreeing with it |
| intended use | research collaborations · working-student roles · hard problems |
🔬 full spec sheet · the verifiable part
| M.Sc. Computer Science (AI) | Albert-Ludwigs-Universität Freiburg, Apr 2025 → present. Deep learning, probabilistic graphical models, statistical pattern recognition, robot mechanics. |
| B.E. Computer Science | N.M.A.M. Institute of Technology, 2020 → 2024. GPA 9.33/10 (German equivalent 1,3). |
| ML Intern | WiZdom Ed, Oct 2023 → Oct 2024. Production RAG over 5,000+ documents (LangChain + ChromaDB); ingestion −40% via recursive splitting; cosine-similarity feedback loop → 90% answer accuracy. |
| Certifications | MLOps Specialization, Duke · ML Specialization, Stanford/DeepLearning.AI |
| Languages | English C2 · Hindi native · German A2 → B1 |
| Base of operations | Freiburg im Breisgau, DE · CET |
the training soundtrack · live
Two active repositories right now.
🟢 mamba-hybrid-lm · a ~50M Mamba-2 × attention hybrid LM comparing three attention:SSM ratios. 1:7 currently has the lowest validation perplexity in the reduced-scale run
🔵 edge-yolo26-deployment · live demo ▸ · one detector across TensorRT, ONNX Runtime, and WebGPU. On the tested RTX 5070, FP16 has the best latency per watt and FP8 is fastest. Detection runs in the browser tab.
🔬 run configs · what is inside
mamba-hybrid-lm · in progress
- Interleaves Mamba-2 selective-SSM blocks with causal attention using the Jamba pattern · d_model 768 · bf16 · SwiGLU · RoPE · OpenWebText · one RTX 5070 12GB
- Compares 1:3 / 1:7 / 1:15 attention:SSM ratios at matched tokens-seen. In the reduced-scale preview, 1:7 has the lowest val PPL (102.4) and 1:3 trains fastest.
- KV-cache use at 8K context, inference throughput, and the live token-streaming demo are still pending.
edge-yolo26-deployment · shipped · live WebGPU demo
- NMS-free YOLO26 fine-tune (SKU-110K dense shelves, mAP@50-95 0.572) shipped as one ONNX graph → TensorRT (RTX 5070), ONNX Runtime (Ryzen 7700) and WebGPU in-browser
- MLPerf-style p50/p95 latency + NVML power. FP8 reaches 560 FPS, FP16 reaches 9.3 FPS/W with almost no accuracy loss, and INT8 is slower and draws more power than both on this Blackwell GPU.
- The two INT8 paths differ by roughly 8× in accuracy loss (TensorRT −5.65% vs ONNX Runtime −0.72%). The CPU path uses per-channel quantization and keeps the detection head in FP32.
- Detection runs 100% client-side; the frame never leaves the browser
Version history of the author. Semantic-ish.
 |
release | notes | |
|---|---|---|---|
 | v2026.07 | feat: reproduced RLPD and extended it to Humanoid (online-only beat the 50/50 mix) | |
 | v2026.06 | feat: one detector on GPU, CPU, and the browser · FP8 at 560 FPS · live via WebGPU | |
 | v2026.05 | fix: rebuilt and re-evaluated EN→HI translation · frozen test set · beam search · chrF++ 41.6 | |
 | v2025.04 | major: relocated to Freiburg — M.Sc. CS (AI), Albert-Ludwigs-Universität | |
 | v2024.05 | feat: the foundations arc — from-scratch RL · UNETR 3D segmentation | |
 | v2023.10 | feat: production RAG @ WiZdom Ed — 5k docs, 90% answer accuracy | |
 | v2020.09 | init: B.E. Computer Science, first gradient descended |
The current research queue. XP bars track progress.
🔬 quest briefings · papers behind each bar
| quest | the plan |
|---|---|
| World-model RL | DreamerV3 (arXiv 2301.04104) on Crafter at 1M steps; ablate imagination horizon (H = 5/15/30) and categorical vs Gaussian latents; render dream-vs-reality rollouts |
| GRPO / RLVR | verifiable-reward post-training on math (DeepSeekMath, arXiv 2402.03300); measure accuracy vs samples-at-inference |
| Efficient inference | GPTQ/AWQ × speculative decoding × KV-cache compression; a serving-throughput Pareto on one GPU |
| Diffusion LM | masked-diffusion (arXiv 2406.07524) vs a compute-matched autoregressive twin |
| Robotics VLA | SmolVLA/OpenVLA behaviour cloning on LIBERO; discrete-token vs flow-matching action heads |
| Agentic capstone | n8n supervisor + RAG + tool-use pipeline with pass^k reliability evals |
A snake is released into my contribution graph every night at 04:00. It has never once been full.
assembled in freiburg · fueled by chai