DEV Community

Cover image for EmoryHacks 2026
Maki F
Maki F

Posted on Edited on

EmoryHacks 2026

This is a submission for the MLH x DEV Writing Challenge

What We Built

Cell migration plays an important role in many biological processes, from wound healing and tissue development to cancer metastasis and immune responses. Researchers study these processes using microscopy videos, but interpreting cell movement is still largely a visual task. That creates an accessibility barrier for blind and low-vision scientists.

So we asked a different question: what if you could hear the way cells move?

Cellsound turns microscopy videos into interactive, multimodal analyses that researchers can see, hear, and read. It tracks individual cells, groups them into spatial clusters, and analyzes how those clusters move. Users can listen to the whole population or to individual chunks, explore movement statistics, and read plain-language descriptions of what's happening.

How the sound works

You choose a reference point on the video, which we call the Point of Audio, and Cellsound maps each chunk's movement into sound:

  • 🏃 Faster cells create faster beats
  • ⬆️ Movement toward the Point of Audio raises the pitch; movement away lowers it
  • 🔊 Proximity controls volume
  • 🎧 Horizontal position sets stereo placement

Every step works with a keyboard and screen reader, with adjustable text size, contrast, and motion, and the site uses Atkinson Hyperlegible, a font designed for low-vision readers.

How we built it

  • Detection and tracking: We used two pretrained models. Cellpose detects cell nuclei in each frame, and Trackastra tracks them across video frames.
  • Our own systems (Python): On top of those, we coded our own spatial clustering, ranking, and sonification. K-means initializes the cell groups on the first frame, and our custom clustering algorithm keeps those neighborhoods consistent as the cells move.
  • Custom sound synthesis: Audio is generated in sync with the tracking analysis, so what you hear lines up with what's on screen.
  • Stack: Python and FastAPI on the backend; HTML, CSS, and JavaScript on the frontend.

The hardest problem: keeping groups consistent

Simply running K-means on every frame causes cells to switch between groups, which makes it impossible to follow the same neighborhood. We also had to decide where a cluster should live. If groups were based only on similar velocity, distant cells that happened to drift the same way would be lumped together, which is hard to explain on a map and hard to listen to in stereo.

So we defined chunks as spatial neighborhoods:

  • Position dominates membership.
  • Each group has a maximum radius.
  • Centroids stay far enough apart that regions don't heavily overlap.

That way a chunk means "what's happening over here," not "who shares a velocity vector."

To let neighborhoods adapt without constantly changing identity, our custom algorithm:

  • Follows its members: each centroid moves with the cells that already belong to the group.
  • Weights newcomers lightly: new joiners pull at about one-tenth the weight of continuing members, so a single cell wandering in doesn't yank the whole group.
  • Is sticky: cells stay in their current chunk as long as they're inside its radius, so identities are stable enough to sonify and describe across frames.

We started with very little technical experience

My partner Ryan and I came from biology and public policy backgrounds. Building software was completely unfamiliar territory. Even creating a GitHub repository was part of our learning curve! But we had a genuinely fun time figuring things out, learning new tools, and tackling problems we'd never encountered.

Over one weekend, we built a working tool that turns raw microscopy footage into synchronized video, audio, statistics, and plain-language descriptions, and we developed our own clustering and sonification systems along the way. Most importantly, we built something that puts accessibility at the center of scientific analysis, not at the end of it. In the end, we were extremely proud to have won the Rogue Track (beginners).

Demo

🌐 Live site: cellsound.select

Partner Technologies

GitHub

GitHub was our collaboration home and our frontend hosting:

  • Repo: github.com/flautam/cellsound holds the FastAPI backend and the frontend.
  • GitHub Actions + Pages: a small workflow deploys our frontend folder on every push to main, so the live site updates about a minute after each commit. For beginners, not having to manage a frontend server was huge.
  • Lesson learned: at one point the site broke with a mysterious Cannot read properties of undefined (reading 'file') error. The browser was running an old cached copy of our JavaScript after an update. A hard refresh and a version tag on our script files fixed it.

Free .select domain (GoDaddy Registry)

Through GoDaddy Registry's sponsorship, we claimed cellsound.select for free. We pointed it at GitHub Pages with four A records and a www CNAME, then turned on Enforce HTTPS once the certificate was issued. A real, memorable URL made the project feel real and easy to share with judges and testers.

Coder

We used Coder as our cloud development environment. It gave us a shared workspace without either of us having to set up a local Python environment from scratch, which mattered a lot for two first-timers. Unfortunately, the Coder hardware available to us didn't have the GPU resources cell tracking needs, so our final demo ran locally. Along the way we learned a lot about ports, HTTP vs HTTPS, and CORS while trying to expose the backend publicly.

Also in our toolkit: Claude and Cursor

We used Claude and Cursor to develop the project. Cursor wrote most of the code from plain-English descriptions. Claude helped us design the interface, write clear instructions for Cursor, and debug things like our domain setup. Our job was knowing what to ask for, and checking that it actually worked.

Hackathon Experience

I attended Emory Hacks in person in Atlanta.

The moment I'll remember most came from the team at the table across from us. Even though we were competing, they went out of their way to help. They told us about Claude's capabilities, which completely changed how we worked for the rest of the weekend. They also helped us get our Waffle House dinner order in 🧇 :) That openness, helping a beginner team just because they could, is what made me realize how great this community is.

Working with Ryan taught me something too. Building alongside someone from biology stretched my thinking: I had to understand a field I'd never studied well enough to design for it. It showed me that the skills I use in policy (problem-solving, mapping systems, breaking big questions into steps) carry over directly to building technology.

Bringing this format to policy

I care most about development and welfare, and I'm leaving Emory Hacks with a new way to work on them.

In policy, ideas usually turn into memos and proposals that take months before anyone can try them. At Emory Hacks, we went from "what if you could hear cells move?" to a working, accessible tool in one weekend.

After this project, I feel really inspired to bring that format to my field and try to bring a policy hackathon to life.

If two people from biology and policy can do this in a weekend, policy teams can too.


Massive thank you to MLH, the Emory Hacks organizers and judges, the team across the table, and my partner Ryan! 🔊🧫

Top comments (0)