Hi, I’m Waseem.

StudyOS is an AI-assisted academic workspace I built to bring courses, schedules, assignments, exams, materials, and study tools into one place. I started it because the information students need is usually scattered across syllabi, portals, calendars, and files.

The application uses Next.js and TypeScript with Supabase, PostgreSQL, Row Level Security, and Edge Functions. The technically interesting work includes syllabus parsing, schedule-image parsing, and Anthropic-powered study features, along with automated tests, GitHub Actions, and a Vercel deployment workflow.

AgentCheck tests how AI agents actually behave before they reach production. I built it because an agent can return a convincing final answer while doing the wrong thing along the way: skipping a confirmation, repeating an action on retry, or reporting success after a tool call failed.

AgentCheck runs an agent against realistic, controlled scenarios, intercepting or simulating tool calls where appropriate. It records what the agent does, including the tools it calls, the side effects and state changes it causes, and how it handles failures and timeouts. That evidence is checked against each scenario’s expected contract, and every result is reported as PASS, FAIL, or INCONCLUSIVE.

It isn’t a generic benchmark or an LLM eval that grades text. The question it answers is whether the agent’s behavior was correct.

The travel department relied on manual coordination to collect employee requests, research options, prepare reports, and track each trip. I built an end-to-end automation system with Microsoft Forms, Microsoft Copilot Studio, Power Automate, Power Apps, and Microsoft 365 to bring those steps into one workflow.

Requests capture the destination, travel dates, purpose, and flight, hotel, or ground-transportation needs. Specialized AI agents assist with flight, hotel, ground-transportation, and estimated-cost research, Power Automate coordinates standardized report generation, notifications, and status updates, and a Power Apps dashboard gives the team one place to review recommendations, costs, and workflow progress.

The system moved the department from a largely manual process toward centralized, end-to-end workflow automation. It reduced repetitive coordination, saved employee time, and helped lower organizational operating costs.

BankAssist is a hands-on machine-learning project in which I fine-tuned Qwen3-0.6B for intent classification on Banking77. I used Hugging Face, PyTorch, TRL, PEFT, and LoRA to build the training and evaluation workflow.

I worked through the full lifecycle, from dataset preparation and parameter-efficient training to evaluation, adapter publishing, merged-model publishing, and standalone inference checks. It was an exercise in learning the complete fine-tuning workflow, not an attempt to present the model as production-ready.

Permitly helps first-time New York City restaurant owners understand which permits and licenses they need, and in what order. A structured onboarding flow captures details such as borough, cuisine, seating, alcohol plans, and target opening date.

From that information, the product assembles permit sequencing, timeline guidance, cost guidance, and AI-assisted recommendations. We kept the hackathon build focused on making an unfamiliar process easier to navigate.