🎓 B.Tech CSE (AI & Robotics), VIT Chennai — Graduating July 2027  |
🤿 Member, Dreadnought Robotics — Underwater AUV Team

I like taking ML ideas from research to something that actually runs in production — including all the infra pain in between.

• 🏎️ F1Net-V2 — Neural learning-to-rank F1 race prediction system with a full MLOps stack (FastAPI, DVC, MLflow, DagsHub, React). 4 containerized Docker services (ingest/predict/finetune/frontend), live deployed demo. Solved real production issues along the way — CUDA pickle errors in CPU containers, MLflow hanging in Docker, CORS, and DVC build-time access, working around the last one by pushing pre-built images straight to Docker Hub.
• 🌊 DIVE-GAN — GAN-based underwater image enhancement, combining a WaterNet-style prior-based generator with a PatchGAN discriminator. Resubmitting to IEEE Access after reviewer feedback — added 4-run convergence plots, 10-run mean±std tables, and cross-dataset evaluation (UIEB, EUVP). SSIM 0.82 / PSNR 20.01 dB on UIEB, competitive with WaterNet baselines.
• 🧭 VLA Maze Navigation Agent — Vision-Language-Action RL agent built from scratch: custom Gymnasium maze env with A* pathfinding, CNN encoder, frozen CLIP text encoder, cross/self-attention transformer policy, PPO + GAE. Currently isolating training instability (entropy dominance, PPO buffer/value-loss issues) via behavior-cloning diagnostics.
• 🤖 Language-Conditioned Navigation Robot (in planning) — 3D-printed differential-drive robot streaming camera feed over WiFi; pipeline planned around SAM2-tiny + TensorRT, CLIP language grounding, and an RL policy trained in Gazebo.
• 🧠 BCI/EEG Decoding (explored) — Non-invasive BCI decoding EEG into multilingual text via CLIP-space alignment, built on EEG-ImageNet with a from-scratch MAE encoder and InfoNCE contrastive loss. Hit ~6-8x random baseline retrieval after fixing a subject-split bug; discontinued after evaluation.

PyTorch FastAPI Docker MLflow DVC DagsHub GitHub Actions Airflow Blender (synthetic data) — training locally on an RTX 3060 6GB, with A100 cloud access (Brev.dev) for heavier runs.

Embodied AI, vision-language models, RL theory (PPO/SAC, Sutton & Barto), generative vision, dynamic programming — I like understanding things from first principles rather than just calling a library.

📫 Reach me here on GitHub.