Personal · Active
Logcat Intelligence Engine
A self-improving AI diagnostic system for Android OS logs. A ReAct agent diagnoses root
causes from logcat/dmesg/bugreport artifacts using specialized tools and FAISS-backed RAG
retrieval. Every investigation is logged as a structured trace, traces become SFT + DPO
training data, and a local Qwen2.5-7B model is fine-tuned with QLoRA — gated by an eval
harness and served via vLLM in an airgapped Docker container. Baseline: 90% diagnostic
accuracy across 10 labeled scenarios.
Python
Qwen2.5-7B
QLoRA
FAISS RAG
Docker
vLLM
View Live Demo →
GreyOrange · Internal
GSB Automation Suite
Entity-agnostic, data-driven test engine for a PixiJS warehouse map editor. 147 scenarios
defined in JSON; the engine auto-discovers and executes them with topological ordering.
Assertions via Redux store injection instead of pixel matching, plus 33 shared helpers and
38 action handlers used across the entire QA team.
Playwright
TypeScript
Redux
Google Cloud Build
Confidential — enterprise project
GreyOrange · Internal
0-to-Hero Test Generation Suite
AI-powered E2E test generation and execution platform for three enterprise warehouse
apps. Write a test scenario in plain English — the platform assembles semantic RAG
context (Qdrant + Voyage-4 embeddings, 378 indexed UI elements), generates a
Playwright spec via Claude Sonnet with TypeScript validation and auto-retry, then
immediately executes it through a ReAct loop (Claude Haiku) that observes live page
state, auto-corrects action types (fill vs select), and handles canvas grid-to-pixel
barcode operations. Passing specs are pushed to GitHub with PR creation and triggered
on GCP Cloud Build CI, with HTML reports streamed back to the dashboard.
Next.js
Playwright
TypeScript
Claude (Sonnet + Haiku)
Qdrant
Redis
GCP
Confidential — enterprise project
GreyOrange · Internal
GSB Simulator
Full-stack platform for physics-based warehouse automation simulations run before real
deployment. 30-worker Celery + RabbitMQ distributed backend, PostgreSQL, WebSockets for
live updates, Three.js 3D visualization, scikit-learn/LightGBM models, and BigQuery
analytics — with custom Jira ticket automation on simulation failures.
Django
Celery
RabbitMQ
Next.js
Three.js
GCP
Confidential — enterprise project
Personal
Baklola E-Commerce
An end-to-end e-commerce application built entirely through an AI-assisted "vibe coding"
workflow using Cursor & Claude — a rapid-prototyping exercise in going from idea to
working product with AI as a co-pilot.
React
Cursor
Claude AI
View on GitHub →
Personal · Active
OmniMind AI Platform
An open-source design handbook and reference implementation for a unified AI platform
fusing ChatGPT, Cursor, Claude, Perplexity, Copilot, and more into one system. Built one
roadmap phase at a time — async FastAPI backend, Postgres-backed identity and conversation
bounded contexts, JWT auth, SSE-streamed chat over a locally-served Ollama model, and a
Next.js frontend — with every phase verified end-to-end before moving on.
FastAPI
PostgreSQL
Redis
Next.js
Ollama
Docker
View Live Demo →