graph LR
    subgraph See["👁 Perception"]
        A["Multi-camera RGB-D<br/>detection · segmentation<br/>pose · depth"] --> B["VLA / scene<br/>understanding"]
    end
    subgraph Move["🦿 Policy"]
        C["Task & reward design<br/>curricula · randomisation"] --> D["PPO in MuJoCo Warp<br/>(mjlab · RSL-RL)"]
    end
    subgraph Ship["🔩 Sim → Sim → Hardware"]
        E["Second-simulator<br/>validation"] --> F["ONNX export<br/>C++ runner · HAL"] --> G["Robot"]
    end
    B --> C
    D --> E
    B -. "INT8 / INT4 on edge" .-> G

I build the path a humanoid policy takes from a training simulator to the robot:

• The robot model is checked, not trusted — URDF → MJCF conversion with an automated verifier that compares the simulated model against the CAD before any training starts.
• Training — velocity-tracking and motion-tracking tasks with PPO in MuJoCo Warp / mjlab, with mass randomisation and datasheet torque limits.
• Then the reality gap, on purpose — an actuation-latency model, a second-simulator pass, and ONNX export into a C++ state-machine runner behind a hardware-abstraction layer.
• One measured lesson: a policy with 0 % falls in an idealised simulator fell 100 % of the time once a 45 ms actuation delay was modelled. Retrained with the delay in the loop, it went back to 0 %. That is the gap that breaks robots on day one, and it was caught before any hardware was at risk.

I also build the perception layer for multi-camera robots (detection, instance segmentation, pose, depth, quantised FP16 → INT4) and perception for a medical screening robot. Robot, customer and architecture specifics stay with my employer; everything below is work I own outright.

Trained from scratch, with no pretrained policy and no imitation data, in MuJoCo Warp with RSL-RL PPO on free-tier T4s:

• Running that is actually running — a speed-adaptive gait clock shortens the cycle from 0.6 s to 0.4 s with commanded speed and pushes stance below 0.5, opening a true flight phase. A fixed clock turns "running" into a fast walk.
• Take hits from any direction — velocity impulses mixed with sustained horizontal drags from random headings, so balance recovery is learned as a skill rather than memorised against one shove.
• Reverse locomotion is a trained command (lin_vel_x ∈ [−1.2, 2.0]), on a three-stage command curriculum.
• Anti-cheat reward shaping — standing-still and feet-air terms stop the classic marching-in-place reward farm.
• Built to survive session limits — every run is checkpointed and resumable across 9–12 h GPU session kills.

Base stacks I build on: unitree_rl_mjlab (fork) · τ0-VLA (fork, tracked for VLA work).

• Tiangong 2 Pro in MuJoCo with a head-mounted RGB-D camera (real intrinsics, metric depth), and a 17-joint upper-body action contract resolved by joint name, so a model can't silently drive the wrong joint.
• A safety filter (±0.30 rad envelope plus rate limiting) between the model and the actuators, and SmolVLA in a 50 Hz control loop.
• Every number is measured on a laptop: SmolVLA-base, fp32 on an i5-10210U, takes p50 9.6 s / p95 11.5 s per 50-action chunk at 6.6 GB peak RSS. That gap to real time is exactly what TernaVLA attacks. This milestone is plumbing; the motion is not trained yet, and the README says so.

• TernaVLA benchmarks ternary VLA backbones on a laptop CPU (real layer shapes): the 0.5B tier runs 0.44 s per action chunk at 8 threads, which is real time, against 1.48 s for 1B and 2.42 s for 2B. bitnet.cpp runs the real 2B model 21 % faster than ONNX Runtime.
• ternary-vlm: a SigLIP2 vision encoder on BitNet b1.58 2B, with LoRA that is quantised jointly with the ternary weights and merges bit-exactly (tested) for llama.cpp export. Research in progress; paper in preparation.

pip install easyannotate

Boxes, polygons, oriented boxes and keypoints, with SAM 2.1 and YOLO-World assist, a fast review loop for model predictions, dataset operations, and export to YOLO / COCO / VOC / LabelMe / CSV. Local-first, with AI tools sized to your machine. Measured: the editor opens a 20k-image project with the first image in 1.6 s and page loads of 5–7 ms; SAM takes 0.044 s per click on an RTX 2050 and 0.17 s on CPU. CI, tests, desktop releases.

pip install vision-data-doctor

A delivered production train/val split contained 81 groups of byte-identical images spanning both sides, so 40 % of validation was leaked. The model's 0.75 recall did not survive an honest split, and nothing in the training stack warned. data-doctor makes that a 10-second CI check:

• Hashes nominate; only pixels convict. SHA-256 catches exact copies, and two perceptual hash families across all 8 dihedral orientations nominate near-duplicates. Every candidate is then verified on decoded pixels, so a reported pair reads 99.4% similar, rot90, never a hash coincidence.
• It also checks for train/val leakage, COCO structural defects and corrupt files. It exits 1 on failure, and Pillow is its only dependency.

Hobby project: conveyor-belt rock segmentation and sizing (Mask R-CNN, GigE camera, operator dashboard).

• KrishiDisha: Revolutionizing Agriculture with Intelligent Recommendations, Disease Detection, and Yield Prediction — first author · IEEE ICoEIT 2025, pp. 1028–1040 · doi:10.1109/ICoEIT63558.2025.11211713
• Web-BCD: A Machine and Deep Learning Approach for Breast Cancer Detection — IBM Technical Report, Dec 2024 · ROC-AUC 0.96, deployed for clinician use
• A Review on Understanding the Patterns of Student Dropout — first author · Studies in Smart Technologies, Springer, 2025 · doi:10.1007/978-981-97-9006-7_20

ML Engineer · Kody Technolab — humanoid perception & RL (Nov 2025 – present; intern Apr–Oct 2025) · DL Research Mentee · IBM India (2024) B.Tech CSE (AI/ML) · Medi-Caps University · CGPA 8.67