I engineer self-modifying models, computer vision pipelines, and multi-agent infrastructure from training through deployment.

Final-year CS at IIIT Pune, Class of 2027. Open to junior AI/ML roles and internships.

Projects

Selected Work

Selected Work01 / 03
Neuro-Genesis Engine title card - self-expanding Mixture-of-Experts, AMD Developer Hackathon Act II
Neuro-Genesis Engine title card - self-expanding Mixture-of-Experts, AMD Developer Hackathon Act II

Neuro-Genesis Engine

Self-Modifying Mixture-of-Experts

300-step self-expanding MoE · 4 → 11 experts · ROCm 7.2

Built a self-modifying Mixture-of-Experts system that detects training plateaus, asks a local Gemma model to generate new PyTorch expert modules, and hot-swaps validated code into the live network. An AST screen, sandboxed execution, smoke tests, and automatic rollback protect the training run while the model expands from four to eleven experts on AMD ROCm hardware.

  • Detected loss plateaus and expanded from 4 to 11 experts
  • Generated new PyTorch modules with local Gemma-2-2b-it
  • Validated code through AST screening, sandboxing, and rollback
  • Agentic AI
  • Mixture-of-Experts
  • AMD/ROCm
View SourceRun Visualizer
High-precision medical image segmentation, Hybrid Xception-VGG DoubleUNet
High-precision medical image segmentation, Hybrid Xception-VGG DoubleUNet

Modified Double U-Net

Medical Image Segmentation

0.85 F1 (BUSI, 3-class) · ~25% faster training (AMP)

Built a dual-stacked U-Net for three-class breast ultrasound segmentation, combining VGG-19, DenseNet-121, and Xception encoders with attention-gated feature fusion. Mixed-precision training and a combined cross-entropy and Dice objective improved training efficiency while handling class imbalance, reaching a validation F1 of roughly 0.85 on the BUSI dataset.

  • Fused three pretrained encoders through attention gates
  • Handled background, benign, and malignant classes
  • Cut training time by roughly 25% with AMP
  • Deep Learning
  • Medical Imaging
  • Computer Vision
View Source
BandWidth cover
BandWidth cover

BandWidth

Autonomous Multi-Agent CI/CD Pipeline

~2-3 min autonomous review cycle · 30+ autonomous PR cycles, zero dropped handoffs

Built a five-agent, cross-model pipeline that autonomously reviews, tests, fixes, and documents GitHub pull requests across six containerized services. A Flask webhook engine routes PR events through sandboxed pytest validation and the GitHub API, while deterministic handoffs keep the workflow recoverable. The system completed more than 30 autonomous review cycles on Google Cloud without dropping an agent handoff.

  • Orchestrated 5 agents across 6 containerized services
  • Executed sandboxed tests and pushed autonomous fix commits
  • Completed 30+ review cycles with zero dropped handoffs
  • Agentic AI
  • Multi-Agent
  • CI/CD
View SourceLive Demo
More work

Additional systems

Autonomous AI triage and synthetic data engine. Engineered a procedural 3D Blender pipeline to automate bounding-box annotations across occluded disaster scenes. Trained a custom YOLOv8 model on ~6,115 images to achieve 96.7% precision, and deployed a low-latency model inference endpoint (FastAPI REST, Docker, GCP) integrating Gemini AI to translate live drone telemetry into actionable triage reports within ~4.5 seconds.

PyTorch, YOLOv8, Blender Python API, Next.js, FastAPI, Google Cloud

View SourceBackendLive Demo

A hybrid recommendation engine combining content-based filtering and collaborative filtering to improve NDCG by ~12-18% over standalone baselines. Integrates diversity-aware re-ranking to increase catalog coverage by ~20% (reducing popularity bias) and handles cold-start user scenarios using metadata-driven fallback logic, evaluated on large-scale Steam interaction data.

Python, Scikit-learn, Implicit ALS, Steam API

View SourceBackendLive Demo

There's more on GitHub.

02 / In Progresslive pipelines

Currently building

Activenode_00

Agent Observability & RAG Tooling

[RESEARCH][PROTOTYPE][EVALUATE][REFINE][DEPLOY]

Instrumenting BandWidth's multi-agent runs - trace capture, handoff timelines, and RAG pipeline tooling.

Agentic AIRAGLLMOps
In Progressnode_01

Open Source Contributions (Roboflow)

[EXPLORE][IMPLEMENT][TEST][PR][MERGE]

Working toward first contributions to Roboflow's supervision, an open-source computer vision library. Focus on reusable CV utilities and annotation tooling.

PythonComputer VisionOpen Source
03 / Skillsdrag to interact

The stack

Technologies and frameworks I use to engineer robust, scalable systems.

04 / Profile

About me

I'm a final-year Computer Science student at IIIT Pune who builds real tools, not just coursework. My focus areas are AI/ML, computer vision, and systems design, and I gravitate toward projects that solve practical, tangible problems.

Whether it's training a dual-stacked U-Net for medical image segmentation, generating synthetic disaster scene data in Blender, or engineering a deterministic document converter, I focus on software that works in the real world.

Final-year CS @ IIIT Pune (Class of 2027) · open to Junior AI/ML roles & internships.

guest@jonathan: ~
➡︎~
Get In Touch

I'm actively looking for internships and opportunities to build impactful systems. Whether you have a question, a project idea, or just want to connect, my inbox is open.