🎮 flavor on the surface · 🔬 science inside the folds — click the ▸ panels as you go
Every model ships with a card. This one is self-reported but honestly benchmarked.
| 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 veins. The minecart runs daily.
🟢 mamba-hybrid-lm — a ~50M Mamba-2 × attention hybrid LM, trained three ways to answer one question: how few attention layers can you get away with? 1:7 currently leads
🔵 edge-yolo26-deployment · live demo ▸ — one detector, three runtimes; the latency-per-watt answer turned out to be FP16/FP8, not INT8. Detection runs in your browser tab (webcam mode next)
🔬 run configs — what's actually inside
mamba-hybrid-lm · in progress — the ratio study
- Interleaves Mamba-2 selective-SSM blocks with causal attention (the Jamba pattern) — d_model 768 · bf16 · SwiGLU · RoPE · trained on OpenWebText, one RTX 5070 12GB
- Sweeps the attention:SSM ratio — 1:3 / 1:7 / 1:15 — at matched tokens-seen; reduced-scale preview: 1:7 wins val PPL (102.4), 1:3 trains fastest
- The real payoff is at inference: attention's KV-cache grows with context, Mamba's state doesn't — KV-cache @ 8K and tok/s columns land next, then a live token-streaming demo
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. Verdict: FP8 = 560 FPS, FP16 wins latency-per-watt (9.3 FPS/W, near-lossless), and INT8 is dominated on Blackwell — slower and hungrier than both
- The two "INT8"s disagree ~8× on accuracy loss (TensorRT −5.65% vs ONNX Runtime −0.72%) — closed with per-channel quantization + an FP16 detection head
- Detection runs 100% client-side; the frame never leaves the browser
Version history of the author. Semantic-ish.
 |
release | notes | |
|---|---|---|---|
 | v2026.07 | feat: humanoids learn to walk from offline data (seed 2 remains hostile) | |
 | v2026.06 | feat: one detector → GPU · CPU · browser, benchmarked — FP8 560 FPS, live via WebGPU | |
 | v2025.04 | major: relocated to Freiburg — M.Sc. CS (AI), Albert-Ludwigs-Universität | |
 | v2023.10 | feat: production RAG @ WiZdom Ed — 5k docs, 90% answer accuracy | |
 | v2020.09 | init: B.E. Computer Science, first gradient descended |
The season pass. Progress bars update as runs converge.
[##########..............] world-model RL on Crafter — DreamerV3, imagination ablations
[########................] reasoning via GRPO/RLVR — the test-time-compute curve
[####....................] efficient-inference lab — quant × spec-decode × KV-cache
[##......................] diffusion LM vs a matched AR twin
[........................] robotics VLA fine-tune (LIBERO) · n8n multi-agent capstone
🔬 quest briefings — papers behind each bar
- 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.
no template survived contact with this readme · assembled by hand in freiburg