Tanya Nair

2025 CanSat Competition

CanSat Competition — one of 20 US teams to advance to launch

A 1 kg satellite, designed and built by the team. I worked on the ground control system, the sensor subsystem, and communication and data handling. Flight board runs two ESP32s with a LoRa radio, GPS, an SD card, and a BMI088 IMU, BMP390 barometer, and ADXL375 accelerometer, with 5 V and 8 V buck converters, two pyro channels, and servo and solenoid outputs.

Photos

Motor Driver Pi Hat

Johns Hopkins Mars Rover Team — Electrical Subteam, also on communication and power system design for the University Rover Challenge

A Raspberry Pi 4 HAT that gives the rover its own motor control board for the science mission. The Pi header breaks out to two Pololu DRV8825 stepper drivers and one external microstep driver, with JST connectors for each joint. Designed in KiCad; still iterating ahead of the University Rover Challenge.

Photos

Board layout

Assembled board

Iterations

1. Sep 2026 Rev A schematic

   Pin budget settled: two DRV8825s on GPIO, the microstep driver on its own ENA/DIR/PUL trio. Six JST breakouts for the joints.

2. Next Board layout & fab

   Route the two-layer HAT, check motor-supply clearances, send Rev A out.

Retrieval-Augmented Generation for Pathology Reports

Dartmouth Health, Levy Lab — Summer 2024 internship; returned in 2025 as an EDIT AI Mentor

Built with a team of three other high school interns, bringing a shared interest in computational linguistics to the Levy Lab's healthcare work. The tool helps clinicians understand a patient's cancer case by summarizing key details of similar cases pulled from a large database of pathology reports. It runs on an efficient model, so it needs no heavy compute or internet access.

Retrieval fuses BM25 and FAISS search; generation runs on Microsoft phi-2. As an EDIT AI Mentor the next year, I guided interns across clinical ML projects and led preprocessing and experiment design: WSI tiling, stain normalization, class-imbalance mitigation, and multimodal alignment.

Poster

Download the research poster (PDF) →

Talk

Watch on YouTube →

Automated Accessibility Enhancer

University of Washington — Advancing Accessibility for Engineering Education, with Prof. Anant and Dr. Lalitha Subramanian

An open-source tool that fixes common accessibility problems in PowerPoint files. Screen readers read shapes in the order they were added, not top to bottom, so the tool moves slide titles to the front of the reading order, writes descriptions for images that have no alt text, and produces a slide-by-slide report of every change.

How it works

Written in Python on python-pptx. Decks come in through a Flask upload page, a folder script, or an S3 bucket on AWS Lambda (deployed with Terraform), and all go through one pipeline: older .ppt files are converted with LibreOffice, shapes named “Title” are moved first, and each image without alt text is described by Claude on Amazon Bedrock. Out come a fixed .pptx and a report. 123 tests.

Iterations

1. Nov 2025 EDUCAUSE 2025 Online poster

   Presented Building Institutional Community of Practice around IT Accessibility with Lalitha Subramanian.

2. Aug 2026 Web version

   Upload files in the browser, track progress, read each report, download the fixed deck.

3. Sep 2026 Bedrock alt text, AWS deployment, bug fixes

   Descriptions moved from a local model to Claude on Bedrock, so no GPU is needed. The pipeline runs on Lambda from S3. Four bugs fixed, including human-written alt text being overwritten when it matched the shape name.

4. Next EDUCAUSE 2026, Denver

   Attending in person as an EDUCAUSE Pipeline scholar. Then: title detection by position and formatting, PDF support, university sign-in, and moving the web app to AWS.

Limitations

Titles are found by shape name only. Generated alt text needs human review, and technical diagrams are still hard to describe well. There are no user accounts yet, and files are processed one at a time.

Project site →  ·  Source code on GitHub →

sine

HopHacks Fall 2025 — Winner, Data Driven Track: Best Data Visualization by Marshall Wace. With Kinnera Tirumala, Ujjwal Kaur, and Naavya Jain. Now on the HopHacks organizing team.

A personal reflection engine for students who don't have time to journal. One check-in by text, voice, or video becomes a short summary, advice, and prompts from an LLM, and sine charts positivity, stress, goals, and mindset over time. Built in 36 hours with a team that had mostly met that week.

How it works

Voice and video entries are transcribed with whisper.cpp, tuned to pick out the closest speaker so it works on a subway or in a crowded room. The transcript goes to Gemini, which returns mood and stress scores plus goals and tasks it finds in the entry. Flask backend, HTML/CSS/JavaScript front end.

Screenshots

Demo

Devpost →  ·  Source code on GitHub →