Bird Brain
Local AI research repository
Role
Designer & builder
Type
Local AI (RAG) tool
Client
Merlin Bird ID
TL;DR
After my capstone with Merlin ended, I volunteered to keep helping. Merlin’s team is small, and they don’t always have time to read long research reports, but they still need their designs to be user-centered. So I designed and built Bird Brain, a local research repository built on the research my team did. They can ask it questions, browse and filter studies, and add new research of their own. Feedback from their team turned my first designs, a sidebar that linked out to reports, into full-page summaries that are quick to skim.
01 · Ask
Ask questions in plain language
Ask something like “what problems do blind birders have with Sound ID?” and get a short summary of what the research found, with links to every study behind it.
02 · Browse
Browse and filter every study
Filter by Merlin feature, like Sound ID or Life List, and by research topic. Every study opens as a full page written like an article, so it’s quick to skim.
03 · Grow
Room to grow
Add new research by uploading a spreadsheet or entering a study by hand, so Bird Brain keeps growing with Merlin’s team long after the capstone.
Background
After my capstone ended, I stayed on to help
Bird Brain grew out of my MHCI+D capstone with Merlin, where my team and I researched how blind, low-vision and novice birders use the app. When the capstone ended, I volunteered to keep working with Merlin.
This project is built on the research my team did, but it’s built for Merlin’s team to grow.
The research we did
The research Bird Brain started with
Over the capstone, my team ran five kinds of research. Each study became a page in Bird Brain, and Merlin can keep adding more.
Semi-structured interviews
6 one-on-one interviews with blind, low-vision and novice birders about how they bird and use Merlin.
Field observations
2 field observations, birding alongside participants to see how they use the app outdoors.
App review analysis
About 6,000 App Store and Google Play reviews, coded by theme to check our findings at scale.
Co-design sessions
2 mixed-visibility co-design sessions, designing an audio-first birding experience together.
Birding sharing data
A look at how birders already share their sightings on social media.
The problem
A small team with big research reports and no time to read them
Merlin’s team is small. They don’t always have the resources to dig through long research reports, but they still need to make sure their designs stay user-centered.
I could hand off any one of our reports fine. Handing off all of it as something the team could actually use later was a different problem, and one I’d had in the back of my mind since we started: what would happen to this research after I leave?
Why local
Privacy wasn’t a nice-to-have. It was the reason to build it at all.
I told our participants their information was protected, and I wasn’t going to make that a lie because a hosted model would have been faster to set up. So Bird Brain runs entirely on infrastructure Merlin controls. Nothing about a participant gets sent to a third-party API to end up in a training set somewhere.
Where it runs
On infrastructure Merlin controls
The app, the database and the search all live with Merlin, not on someone else’s servers.
The model
A local AI model
Qwen3 runs through Ollama on the same machine, so questions and answers never leave it.
Participant data
Never sent to third-party AI
No transcripts or quotes go to a hosted API, so nothing ends up in someone’s training set.
Working with Merlin
My first designs didn’t cut it
My goal was to give Merlin the easiest possible access to their research. Working with their team, I realized my first designs fell short of that. Studies opened in a sidebar that was hard to read, and answers just linked out to the reports.
If all Bird Brain did was point people somewhere else, they could just use a Google Drive. What they needed was a quick way to skim.
Before
Links out, in a sidebar
Answers listed the studies that matched, and each one opened in a narrow sidebar that was hard to read.
After
A summary first, then the sources
Answers lead with a short summary of what the research found, with links to every study behind it.
Easy to skim
From a cramped sidebar to a full-page article
The sidebar was hard to read, so every study now opens as its own full page.
I wanted Bird Brain to take as little brain space as possible, so I had AI format each study like an article: the key findings up front, the details below, all in plain language.
Before
A sidebar pop-up
Studies opened in a narrow panel beside the list, with little room to read.
After
A full page, written like an article
Every study opens as its own page, formatted by AI so it’s quick to skim between meetings instead of a report you block off an afternoon for.
Room to grow
Built on our research, made for Merlin to grow
Bird Brain started with the research my team did, but I didn’t want it frozen at the moment I handed it off. So I added an Add research feature that gives Merlin room to grow.
There are two ways to add research, so Merlin can use whichever fits the work in front of them.
Option 1 · Spreadsheet
Add a batch at once
Drop in a spreadsheet of studies and Bird Brain formats and indexes every row, so a whole round of research is searchable in one step.
Option 2 · By hand
Add one study in a form
For a single new study, fill out a short form. It shows up in the repository and in chat answers right away.
Staying critical
Built to keep people checking the research
AI can miss the point of a question or grab the wrong study, and it sounds just as confident when it’s wrong. Unless someone checks the source, they’d never know.
I wanted Bird Brain to make the research easier to reach, not quietly replace it. So I built two nudges into the design that keep the real studies in view.
In every chat
A reminder under the chat box
Right under where you type, Bird Brain reminds you that AI isn’t always right and to check findings against the source study.
In every answer
Always linking to the source
Every summary lists the studies and quotes it drew from, each one a click away, so you can check the claim yourself instead of taking the AI’s word for it.
How it works
A full-stack RAG app
In technical terms, Bird Brain is a retrieval-augmented generation (RAG) app: a Next.js front end, a FastAPI back end, a local Ollama instance running Qwen3, and LanceDB and SQLite for retrieval and storage.
1 · Import
Studies come in from a spreadsheet or by hand, with links to transcripts and reports.
2 · Index
Text is pulled from each document and stored for search by meaning.
3 · Retrieve
A question finds the most relevant studies and passages.
4 · Answer
A local model writes the answer, with links back to its sources.
What I learned
Stay critical of what AI makes easier
I came to this with a background in informatics, so I already knew how to do this work by hand. That gave me a baseline to judge AI against, instead of trusting it by default.
AI made it possible to build something this big on my own. But the friction, critical thought and intention it removes are often where the value is. The more I use these tools, the more I hold myself accountable for privacy, intention and infrastructure in everything I make with them.
