Work you can open and inspect
Five AI systems built end to end, from requirements to deployment. Most have a live demo, and all have public code.
Autonomous QA
Autonomous QA Agent
The problem
Writing UI test scripts by hand is slow, and they break whenever a locator changes.
What I built
An agentic system that turns plain-English requirements or Jira stories into Gherkin, then into Playwright tests in TypeScript. It runs them autonomously, inspects the live browser through Playwright MCP when a test fails, heals broken locators, re-runs, and produces execution reports and diagnostics.
Built with
TypeScript, Node.js, React, Playwright, Playwright MCP, LangChain, Groq LLM, Jira REST API, DeepEval
Conversational AI
Conversational AI Sales Platform
The problem
Chatbots that rely only on the LLM repeat questions, lose context, misread yes and no, and invent booking details.
What I built
A multi-turn text and voice assistant for real estate sales that takes a customer from requirements to a confirmed site visit, including over the phone. The LLM extracts intent and details, while a deterministic state machine decides what happens next and which answers are valid.
How quality is kept
118 passing regression tests cover state transitions, context, extraction and full booking flows, with turn-level logging to reproduce failures.
Built with
Python, FastAPI, Next.js, TypeScript, MongoDB Atlas, Twilio Voice, OpenAI-compatible LLMs
RAG
AI Test Case Generation with RAG
The problem
QA teams sit on thousands of past test cases and user stories but struggle to find the relevant ones or write new ones consistently.
What I built
A RAG platform that retrieves related tests and requirements, then generates new test cases grounded in them. It combines BM25 and semantic search with score fusion, expands queries with domain terms, and reranks results with an LLM before generation.
Built with
Groq, Mistral AI, MongoDB Atlas Vector Search, Node.js, Express, React
AI application · Healthcare
VitalTriage
The problem
In clinical triage, an LLM alone can’t be trusted to flag a patient in danger.
What I built
An AI-assisted triage and monitoring app where deterministic rules sit beside the LLM. Dangerous vitals, such as SpO2 below 85% or systolic blood pressure below 90, mark a patient critical immediately, whatever the model says. The LLM explains conditions and suggests actions, with an auditable risk score. Deployed on an Azure VM with Nginx, PM2 and HTTPS.
Built with
Python, FastAPI, Pydantic, OpenAI APIs, MongoDB Atlas, React, TypeScript, Azure
AI application · Travel
Orbit World Travels
The problem
Travel agencies juggle visas, flights, hotels, insurance and invoices across separate tools.
What I built
An operations platform covering bookings, invoicing with margin tracking, and analytics, with an AI assistant that answers questions about operational data using hybrid vector and BM25 search, plus LLM-powered document processing. Role-based access and deployment on Azure.
Built with
Next.js, TypeScript, Node.js, Express, MongoDB Atlas, Groq (LLaMA 3), Docker, Azure
Want something like this for your product?
Tell me what you’re building and I’ll explain how I’d approach it.
Start a conversation