An ongoing program training whole-body locomotion for the Unitree G1 (23-DoF) with no pretrained policy, no imitation data — reward design up. Task registry, gait mechanics, perturbation curricula, and the training infrastructure are all custom:

graph LR
    A["Task & reward design<br/>(custom registry)"] --> B["Massively parallel sim<br/>MuJoCo Warp · mjlab"]
    B --> C["PPO · RSL-RL<br/>multi-GPU T4"]
    C --> D["Checkpoint / resume<br/>across session limits"]
    D --> E["Failure diagnosis<br/>dedicated analysis runs"]
    E -->|curriculum update| A

• Running that is actually running — a speed-adaptive gait clock shortens the cycle with commanded velocity and pushes stance fraction below 0.5, opening a true flight phase. Stock fixed clocks make "running" an unlearnable fast walk; mine makes it a gait.
• Take hits from any direction — perturbation curriculum mixing instantaneous velocity impulses with sustained horizontal drags from uniformly random headings. The policy doesn't memorize one shove; it learns balance recovery as a skill.
• Reverse locomotion as a first-class command (lin_vel_x ∈ [−1.2, 2.0]) — walking backwards is trained, not hoped for.
• Anti-cheat reward shaping — standing-still and feet-air terms that stop the classic failure mode of a robot marching in place to farm gait rewards.
• Debugging like an engineer, not a gambler — when a policy failed backward pushes, it got its own isolation run, diagnosis, and targeted curriculum fix before retraining. Every experiment checkpointed and resumable across free-tier GPU session limits.
• Foundation-model track: custom design work on τ0-VLA — world-model-guided test-time computation for robot control. Base stack: my fork of Unitree's RL suite on mjlab/MuJoCo Warp.

At Kody Technolab (full-stack robotics company) I build the vision layer for multi-camera robots operating as fleets — security, advertisement, and data-gathering platforms plus the Mahindra Assistant:

graph LR
    A["Multi-camera ingest<br/>sync & calibration"] --> B["Multi-task perception<br/>detect · segment · pose · depth"]
    B --> C["Quantize<br/>FP16 → INT4"]
    C --> D["Edge deploy<br/>Jetson · Android robots"]
    D --> E["Fleet in production"]
    E -->|"data flywheel: field footage → retraining"| A

• Multi-task heads sharing one latency budget: detection, instance segmentation, pose and depth estimation running together on embedded compute.
• The whole quantization ladder — FP16 / INT8 / UINT8 / UINT4 via TensorRT, ONNX Runtime, TFLite, OpenVINO — chosen per platform, per model, per latency budget.
• Pipelines that have processed 500K+ images and 100+ hours of robot video; preprocessing time cut 40%, production throughput up ~30%.
• Model strategy per constraint: YOLO family, SAM-1/2/3, D-FINE, RF-DETR, MobileOne×ArcFace, custom architectures — the right tool for the hardware in the room, with classical projective geometry as the fallback where deep learning runs out of data.

Building perception for medical robotics targeting rare anomalies — conditions with no existing CV solution and almost no training data when development starts. That's the point: low-data strategies, geometry-first fallbacks where deep learning runs out of examples, and rigorous evaluation on curated clinical splits.

When real medical data runs out, I manufacture it — and measure whether it actually helps, not whether it looks pretty:

• Pluggable generative backends: SD 1.5 + per-class LoRA, and a from-scratch DDPM with zero natural-image prior for sensitive domains.
• Every synthetic image is provenance-tracked and screened for memorization of real patient images before it may train anything — privacy treated as a hard gate, not a footnote.
• Honest paired multi-seed A/B on HAM10000: rare-class augmentation moved vascular-lesion F1 +0.050 ± 0.013 and melanoma recall +0.116 ± 0.048 (full measured results).
• The entire train→evaluate loop executes remotely on free Kaggle GPUs through a git-pinned execution runner built for autonomous agent iteration — every run reproducible to the commit.

Public lab notebook → Kaggle: 50+ notebooks, 19 datasets — D-FINE fine-tuning with ONNX/OpenVINO export, SlowFast action recognition, MoveNet+LSTM pose pipelines, face embedding, and the G1 locomotion runs above, all reproducible.

• KrishiDisha: Revolutionizing Agriculture with Intelligent Recommendations using Computer Vision — IEEE ICoEIT, Jul 2025. Multi-task CV platform (F1 0.99 / precision 0.96 / R² 0.98), field-validated by 150+ farmers.
• Web-BCD: A Machine and Deep Learning Approach for Breast Cancer Detection — IBM Technical Report, Dec 2024. 89%→94% accuracy (ROC-AUC 0.96), deployed live for clinician use.
• Understanding the Patterns of Student Dropout: A Review — Springer, Smart Technology, Jun 2024 (chapter).

ML Engineer · Kody Technolab (Nov 2025 – present; intern Apr–Oct 2025) · DL Research Mentee · IBM India (2024) B.Tech CSE (AI/ML) · Medi-Caps University · CGPA 8.67 · Head of Research & Astronomy, Science Club