Skip to content
View ChaitanyaParate's full-sized avatar
2 followers · 2 following

Highlights

  • Pro

Block or report ChaitanyaParate

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ChaitanyaParate/README.md

Chaitanya Parate

B.Tech Computer Science (AI & Data Science) · MIT-WPU, Pune · 2027
Building at the intersection of robotics, deep learning, and systems engineering.
Open to ML/AI research internships.

Projects

Heterogeneous GNN Morphology Transfer DOI

Zero-shot locomotion policy transfer across robot morphologies using a 31,582-parameter heterogeneous GNN (GATv2Conv + PPO), with YOLOv8 perception and LLM-guided navigation deployed on ROS2 Jazzy + Gazebo.

  • Zero-shot quadruped → hexapod transfer (12-DOF → 18-DOF): 106 ± 25 reward, ~47 steps survival. MLP hard-fails with RuntimeError on any unseen morphology — fixed-input-dim structural limitation
  • 500K-step fine-tuning: 3.8× reward gain (110 → 416 ± 114), survival 47 → 193 steps
  • 85% fewer parameters than MLP baseline (31,582 vs 210,457), while MLP scores higher in-distribution only — GNN trades peak reward for architectural generalization
  • Terrain robustness (zero-shot, no terrain training): 95% success at 5° slope, 0% at 10°
  • Deployed at 200 Hz with yaw-rate PI correction (HAA joint offset) to eliminate circular drift from training bias
  • LLM planning layer: Qwen 2.5 7B via Ollama, natural language → skill → GNN → joint commands, fully on-device

SE-Attention Half-UNet

Polyp segmentation model combining attention gates and SE blocks with BCE-Dice loss. Evaluated on Kvasir-SEG. Dice: 0.9436.

DeskAI

Local-first Linux desktop AI daemon. Persistent daemon with UNIX socket IPC, X11/OCR context tracking, and streaming LLM inference via Ollama with online fallback.

  • Intent router dispatches to typed executors (summarize, explain_error, search)
  • Fault-tolerant streaming pipeline with full daemon-client lifecycle decoupling
  • Internet-aware backend switching without daemon restart

stm32-hyperloop-embedded

STM32 firmware in C for Team Vegapod. Custom Bluetooth bootloader, FreeRTOS BMS/VCU braking, 3-phase inverter PWM, CAN bus. Represented MIT-WPU at European Hyperloop Week 2025.

Other

ResNet-50 from scratch (86.63% on Imagenette, live ESP32-CAM inference) · Brain Tumor MRI Classification (98% accuracy, encoder-only Half-UNet)

Open Source

metatensor/metatrain · Fixed zero-sized validation split edge case in ML training pipelines. PR #1003 merged January 2026.

Stack

Python C C++ PyTorch PyTorch Geometric ROS2 STM32 Linux

LinkedIn · paratechaitanya5@gmail.com

Pinned Loading

  1. hetero-gnn-morphology-transfer hetero-gnn-morphology-transfer Public

    Heterogeneous GNN policies for constrained zero-shot morphology transfer in legged locomotion with ROS2/Gazebo deployment, YOLO-based vision, and LLM-guided navigation.

    Python

  2. Deskai Deskai Public

    Local-first desktop AI daemon that runs fully offline. Tracks active desktop context, exposes a CLI, streams responses from local LLMs via Ollama, and runs as a systemd user service. Built for syst…

    Python 1

  3. SE-Attention-Half-UNet SE-Attention-Half-UNet Public

    Implementation of a compact Attention Half U-Net with Attention Gates and Squeeze-and-Excitation blocks for medical image segmentation. Features a modular PyTorch pipeline, BCE-Dice hybrid loss, mi…

    Python 3

  4. Modified-Half-UNet-for-Medical-Image-Classification Modified-Half-UNet-for-Medical-Image-Classification Public

    A modified Half-UNet model for brain tumor classification using MRI scans. This project repurposes the UNet architecture for classification by integrating dense layers at the bottleneck. Built with…

    Python

  5. ResNet50-From-Scratch-Pytorch-Imagenette ResNet50-From-Scratch-Pytorch-Imagenette Public

    Custom ResNet-50 from scratch in PyTorch, trained on Imagenette. Bottleneck blocks, projection skip connections, label smoothing, LR scheduling, and live inference via ESP32-CAM stream.

    Python

  6. stm32-hyperloop-embedded stm32-hyperloop-embedded Public

    STM32 firmware projects: custom BT bootloader, FreeRTOS BMS/VCU braking system, 3-phase inverter control, multi-channel ADC, and on-device ML inference. Built in C using HAL/LL drivers.

    C 2 1