Tarun Upaday.

Projects

I build small systems to understand larger questions.

Some projects begin with a business problem. Others begin with an unreliable camera, an awkward workflow or a question from a robotics student.

The useful ones usually combine software with real operational constraints. This page is the editorially selected portfolio — the complete archive, including works in progress and less presentable experiments, lives on GitHub.

Applied AI

I am less interested in what a model can demonstrate than in what work it can reliably remove.

My current experiments focus on operational systems: phone calls, email coordination, reconciliation, travel policy, computer vision and exception handling.

The hard questions are usually not about model capability. They are about context, permissions, failure modes, human escalation and whether the system produces a better operational outcome.

Travel and agents

Google Flights Policy Recommender

Prototype

A policy-aware flight-search MCP server.

It searches Google Flights, labels results as in-policy or out-of-policy and ranks them using preferences such as nonstop travel, airline alliance, price, duration and red-eye avoidance.

The hosted version returns ranked choices and directs the traveler to Google Flights to complete the transaction. A separate browser-driven implementation can take the process as far as payment without completing the purchase.

Traveler / Claude MCP server policy + preferences (Cloud Run) Google Flights search results Policy engine in / out of policy Ranked, labeled results → Google Flights to book
View repository →

Computer vision and education

Pillbox

Student project

A Raspberry Pi camera and computer-vision system, built together with a student in the Uprobotics program.

Users can remotely capture full-resolution pillbox images from a browser. A small CNN runs locally and identifies which of 21 pillbox cells — seven days across morning, noon and night — still contain pills.

The project combines hardware, image capture, remote access, model inference and a simple status interface. Training data and model versions live in a companion repository.

Pi camera weekly pillbox photo, on demand CNN model on-device inference 7 days × 3 periods status grid, browser
View repository →    View training data →

Systems that almost work

Home network, cameras and automation

Many modern systems work well until several products, protocols and failure modes meet. I use my home network, cameras and automation setup as a practical laboratory for understanding interoperability, observability and reliability. These notes document what failed, how I isolated the cause and what I changed.

Most of this runs on forks of existing open-source plugins that I adapt and maintain for my own setup, rather than projects I originated:

GitHub

The complete source, works in progress and less presentable experiments are available at github.com/tarun101.