Contact

Menu

Overview

Research platform · B2B

2024–2026

brain.dex — a research platform for neuroscience labs

UC1

y25

An interview. The questions a hiring panel should ask about this project — including the uncomfortable ones — answered honestly. Everything in it is real.

Tel Aviv, Israel
2024 – 2026
brain.space
Senior Product Designer

brain.dex is where neuroscience labs plan, run and label every brain recording — one platform where there used to be three.

When I joined, operators logged into three disconnected systems to run a single study session, and the data that came out had to be cleaned and labelled by hand before anyone could learn from it. Fifteen client organisations depended on that patchwork.

I merged the three systems into one operator flow, co-designed the Project → Flow → Run → Session model with the PM, and gave researchers a single search field that slices across every recording. The interview below is how those decisions were made — including the ones I got wrong.

Year
2024–2026
Company
brain.space — neuro-tech, B2B
Role
Senior Product Designer,
sole designer
Team
Amir Shapira, PM
Guy Perera, PM
Jonathan Hepner, VP Product
Yoni Abeshitz, VP Product
Users
~15 client organisations,
several hundred users

A research platform for neuroscience labs, rebuilt around how labs actually work. I merged three operator systems into one, co-designed the data model beneath it, and turned raw brain recordings into data that models can learn from.

Skills
Information architecture
↳
Project → Flow → Run → Session, with the PM. Every session tagged at source — a researcher trained EEG disease-detection models with no hand-labelling.
Product consolidation
↳
Merged 3 operator systems into one. All ~15 client organisations migrated; 3 logins → 1.
Design system
↳
One vocabulary for stations, devices and sensor wear. Sensor wear visible before a session, not after.
User research
↳
$2.5K budget, 4 large client orgs, 2-hour interviews. Found the gap: labs think in protocols and repeat visits, the software thought in tables.
Prototyping in code
↳
Working prototype on real data with 3 researchers. One shared intent — slicing — and ~18% faster time-to-find.
Interaction design
↳
Operator flow, Build / Explore. One schema serving operator, researcher and clinician.

I took a neuroscience research product from three disconnected tools to one platform that every client moved to — and made every brain recording usable for machine learning.

(Client organisations migrated)
15/15
Merged 3 operator systems into one platform
Consolidation · IA
(Systems an operator logs into)
3→1
One vocabulary for stations, devices and sensor wear
Product · hardware UX
(Sessions hand-labelled for ML)
0
Co-designed Project → Flow → Run → Session; every session tagged at source
Data modelling
(Large client orgs interviewed)
4
Fought for a $2.5K budget; 2-hour interviews
User research

brain.dex — a research platform for neuroscience labs

About

Tags

Saas B2B

Position

Design, Research

Year

2025

The product was meant to be straightforward — a platform for collecting and organizing brain data, structured into a clean hierarchy of projects, flows, runs, and sessions. By the time the project reached its turning point, that hierarchy had quietly stopped being the point.

Brain Data Had Outgrown Its Tools

When the Data Had the Wrong Audience

Let's start at the beginning. What was brain.space, and how did this project come to exist?

brain.space started as a focused idea: build a platform that lets research labs collect and manage neural data. EEG signals, cognitive task results, physiological measurements. The kind of data that previously lived in spreadsheets, on local hard drives, in proprietary lab software that hadn't been updated in a decade. There was a real gap — neuroscience research had gotten sophisticated, but the infrastructure for managing the data hadn't kept up.

By the time I joined, the product had a working structure. Projects contained Flows — experimental protocols. Flows contained Runs, individual executions with a participant. Runs contained Sessions, the granular recordings. A clean hierarchy, a dashboard with completion rates and signal quality. The kind of thing you can explain in three bullet points in a pitch deck.

My closest collaborator was the product manager. We built the information architecture together, though he owned it. I remember arguing with him for days about whether Runs needed to be a separate entity from Flows — I thought it was over-engineered. He was right that they did. That kind of negotiation defined how the product got built.

The complication came from the customers. Labs weren't the only ones interested. Neurologists running experimental treatments saw the platform as a way to track patients over time. And a researcher and a clinician use the same data in completely different ways. That tension is where the design problem really began.

"Management wouldn't choose a user, and I finally understood they were right — telling a small startup to drop half its market is telling it to die a little. So I stopped asking them to choose and asked what I was allowed to give up instead."

Three Systems, One Operator

Before any of the research — what was actually broken when you joined?

The operator. Running one session meant three disconnected systems: one for the experiment, an in-house tool tracking electrode and sensor wear, and a third showing which devices were configured for which station. Nothing shared data, so a worn sensor was discovered after a session failed, not before it started. By our estimate, close to 500 sessions were hit that way.

I spent four months merging the three into one. The hard part wasn’t the screens, it was the vocabulary: what a station is, what a device is, when a sensor counts as worn — so the system could warn before a session instead of after a failure.

Did it work?

For collection centres, immediately — they live in that part of the product every day, and all ~15 client organisations moved to the new platform. Research labs didn’t notice; nothing in their work had changed. Necessary, but not sufficient. That’s what sent me to research.

The unified operator view after the merge

Four Seconds That Broke the Map

The turning point involved user research — and a budget fight. Tell me about that.

I pushed for user research, and I had to fight for it. Twenty-five hundred dollars. That sounds trivial, but in a small startup, asking for budget to validate something everyone assumed they already understood was a hard sell.

The pushback was essentially: we know our users, why are we paying to confirm it.Then I made a decision I think mattered more than the budget itself. I refused to run the research on Figma mockups. Static screens tell you what people think they'll do. They don't tell you what people actually do. So I spent two weeks building a working prototype in Lovable — real data, real interactions, the whole thing clickable and live.

That distinction — between what people say and what people do — turned out to be the entire project. If I'd tested mockups, I'd have shown users a picture of the hierarchy and they'd have nodded and said yes, that makes sense, I'd click here.

And I'd have learned nothing. The prototype was the only thing that could have caught what happened next.I want to be honest that this wasn't pure methodological wisdom. Part of it was instinct, part of it was stubbornness. But it's the decision the whole story hinges on. You don't get the four-second moment if you're showing someone a screenshot.

The researcher sees a catalog. The clinician sees a story. Same data, same system, two completely different modes of time. And critically, the same underlying action — both of them are slicing. Just along different axes.

Two Users, Two Axes

What did you actually see when you compared how the two users worked?

Once I stopped defending the hierarchy and started watching, the difference between the two users became sharp. They were both ignoring the structure, but they were ignoring it in opposite directions.Gil is building a dataset from scratch.

There's no prior research on what he's investigating — he's generating the evidence. When he opens the system, he wants to slice across everything: give me all sessions where the paradigm is Stroop, the age is 25 to 45, the signal quality is acceptable. He doesn't know what he's looking for yet. The slicing is the research.

The clinician in Cyprus, Dr. Mavridou, is tracking individual patients through experimental treatment. Each patient comes in every few weeks. She wants one person's story — is this patient's P300 amplitude improving session over session, is the treatment working for them specifically.

She wants depth on one individual, not breadth across thousands.So I started thinking about it geometrically. Gil takes a horizontal cut — one slice across thousands of sessions. Mavridou takes a vertical cut — one person followed down through time. The researcher sees a catalog. The clinician sees a story. Same data, same system, two completely different modes of time. And critically, the same underlying action — both of them are slicing. Just along different axes.

"As personas, Gil and Mavridou are irreconcilable — pick one. As intents, they're almost identical: both are slicing, just along different axes. That reframe is the entire reason brain.space needed one product instead of two."

When Management Won't Choose

Underneath the design problem was an organizational one — management wouldn't pick a user to optimize for. How did you navigate that?

That was the real constraint, and for a while I treated it as an obstacle. I went to management and said we have to choose — are we a research tool or a clinical tool. They said no. We need every customer, we're a startup, we can't afford to lose either market. I read that as cowardice at first.

Refusing to make a hard call. But the more I sat with it, the more I realized they weren't wrong. They genuinely couldn't afford to lose either customer. Telling a small startup to abandon half its market is telling it to die a little. The constraint was real, not lazy. So the question shifted from "how do I make them choose" to "how do I honor both without compromising either."

That forced me to be precise about what was actually non-negotiable. Two things. The first was user research — that's your eyes, you don't give those up. The second was the information architecture. Once the data layer is built, it's built. You can't ask the dev team to rebuild the database every time the UI shifts, and that schema wasn't even mine to trade away. It was the product manager's.

So I drew a line. I wouldn't compromise on research, and I wouldn't compromise the architecture. But the UI layer — the surface, the navigation, the affordances — that was flexible. That's where I had room to move. Knowing exactly what you're allowed to give up is, I think, half of working inside an organization.

Explore with two conditions applied — 38 sessions, with the caffeine split offered below
The Explore screen annotated — the question as chips, the answer as a count, the readings, and the vocabulary panel
The navigation and facet panel annotated — Build and Explore as peer rooms
One participant read as a timeline — six visits, P300 rising
The search field with its results grouped by source — paradigm, condition, participant
AS

You tried three standard search patterns before the real answer appeared. What did putting them on screen reveal that thinking alone couldn't?

RD

I started where everyone starts — three standard patterns. A command palette like Linear. A persistent search bar leading to a results page like GitHub. A sidebar panel like VS Code. I sat with those three for two days and felt increasingly uneasy, like I was being asked to pick the color of a wall when the real question was where to build the house.What all three had in common was the assumption that search is a component inside a product. Something you add on top of the hierarchy. They disagreed about where to put it, but they agreed it was an accessory. And that didn't match what Gil had shown me. He didn't want search that helped him use the product. For him, search was the product


“He took control of my screen, typed three characters, and in four seconds cut straight across the hierarchy I’d spent months building. I could have called him an edge case. Two users doing the same thing isn’t an edge case — it’s a signal.”

Personas Lie, Intent Doesn't

There's a conceptual shift in here — moving from personas to something you call intent. What changed in your thinking?

The breakthrough was realizing I'd been asking "which persona matters more" when the question should have been "what action are they actually doing." Personas describe people — Gil is 38, has a beard, bikes to work, drinks too much coffee. None of that tells me what he needs when he opens the system at nine in the morning.


What tells me something is the sequence of actions he takes in the first thirty seconds. That sequence is repeatable. It's measurable. And crucially, it's portable — the intent "discover a pattern in unstructured data" applies to Gil, but also to any researcher doing that kind of work. The persona is locked to one person. The intent travels.

This is the difference that decided whether brain.space needed one product or two. As personas, Gil and Mavridou are irreconcilable — different people, different goals, pick one. As intents, they're almost the same. Both are slicing.


They just slice along different axes. That reframe is what made the whole thing designable. It also changed what I could measure. You can't put "is Gil satisfied" into analytics. But you can track whether someone completed a slicing operation, where they got stuck, how often they came back to a saved query. Personas are stories you tell yourself in a workshop. Intents are things you can actually build for and instrument.

"The fifth sketch killed the results screen. The chips became the question, the table became the answer, and they lived on the same page. That's when I knew search wasn't a feature inside the product — it was the product."

The User I Almost Broke

The solution — Direction 5 — almost broke a third user you'd nearly forgotten. How did that resolve?

Right when I thought I had it, I remembered the student. I'd seen her in a video Mavridou sent — the operator who actually runs the equipment in the lab. She's not a researcher and not a clinician. She opens a new run, picks a protocol, enters a participant ID.

Linear, step by step. She doesn't want chips. She wants a wizard. The conditions-first design I'd just fallen in love with would have fixed Gil and Mavridou and completely broken her. And she's not even the paying customer — the labs are — but if she can't operate the equipment, nothing works. So I couldn't just optimize for the two users who'd shown me the insight. I had to hold all three.

The answer came from Slack, of all places — the workspace switcher. Not switching between organizations, but between ways of working inside the same one. So: two modes. Build mode looks almost exactly like the existing system — the hierarchy, linear navigation, creation, execution. Nothing breaks for the student.

Explore mode is the conditions shell — chip strip, saved views, the data table. Inside Explore, a view toggle renders the same conditions three ways: Sessions for Gil's catalog, Patient for Mavridou's single-person story, Cohort for when she wants to see a trend across all her patients.And here's the part that mattered organizationally: none of this touched the data layer. The schema stayed exactly as the product manager built it.

“The best design decisions I’ve made all came after I was wrong about something. The worst ones came when I was sure.”

One Field for Everything

Let’s talk about search itself. Every comparable product has two boxes — one for conditions, one for finding a patient. You shipped one. Why?

Because once you look at the model, a participant isn’t a different kind of thing. It’s just another condition. Type an ID and 2,140 sessions drop to 6. Add a condition and you’re down to 4.

With two fields you can’t do that. And worse, you’re asking a clinician to decide “person or condition?” before she even knows a cohort is something she could ask for.

But one field doesn’t tell you what it can do. Wasn’t that a problem?

That was the price, yes. A single field can’t announce itself, so the empty state has to do it instead. Before you type anything, it lists every group it can find, Participant included.

Two fields explain themselves with their labels. One field needs an empty screen that does the talking.

The search field with its results grouped by source — paradigm, condition, participantFirst run — nothing asked yet, every searchable group listed

Filtering Isn’t Exploring

Splitting the product into Build for the operator and Explore for the researcher sounds like it solved it. Did it?

The structure, yes. But I got something wrong that I only saw later. I’d given the researcher a filter. Filtering narrows toward something you already know. Exploration widens toward something you don’t.

Gil wasn’t searching at all. He was cross-referencing — men, dyslexia, after coffee, under five hours of sleep. Then the same thing without the coffee. That combination is his hypothesis, and every time he changes a condition, it’s a new experiment. He didn’t need results. He needed feedback on the hypothesis, and a way to put two of them next to each other.

So what did you change?

Cross-referencing stopped being retrieval and became comparison: two sets of conditions, two panels, and a row that shows the difference. N is always on screen; p never is.

And there’s one detail I’d defend in any review. A cell that was never measured is grey and dashed, not zero. A condition only exists where a PI defined it. It’s not zero. It’s undefined.

Underneath all of it was a simple idea. The conditions a researcher has to define to run a valid experiment are exactly the conditions someone else will later use to find what they weren’t looking for. Nobody had to curate a taxonomy. Rigor is the index.

Compare and pivot — two condition sets side by side with a difference row

What Shipped, and What It Was Worth

Strip the story away. What actually shipped?

The unified platform, the entity model and the hardware lifecycle. All ~15 client organisations moved to it, and it’s what labs use today. An operator logs into one system instead of three.

The bigger change was what the company thought it was selling. Every session now carries its paradigm, population, demographics and signal quality, so the data is sorted before anyone asks. Prof. Hadani used exactly that to train neural networks that identify neurological disease from EEG alone — his research, not mine; what I’d claim is that the architecture removed months of hand-labelling. The PM estimated search got about 18% faster for collection centres. That’s his estimate, not a measurement, and I’d rather say so.

And Explore?

It reached concept with the PM, then the company restructured and I was let go before it was built. And I never usability-tested it. What I ran was a heuristic walkthrough, which is not the same thing. If I did it again, the prototype round would end with five task-based sessions before any concept review.

What Two Years of Being Wrong Taught Me

Last question. What did this project leave you with?

The honest answer is uncomfortable, because it's really about how little we actually listen. Designers love to say we listen to users. Mostly we don't. We listen to the loudest stakeholder, to the deadline, to our own assumptions about what we already know how to build.

Real listening is rarer than that. Real listening is watching someone use the thing you made and letting yourself be wrong about it. Gil took control of that prototype, went straight to search, and showed me in four seconds that the map I'd been handed was the wrong map. I could have defended it. The entire outcome of the project hinged on the fact that I didn't.

What I'd take from this — and I hold it loosely, because I might be wrong — is that the era of hierarchical software might be quietly ending. Users never navigated the way we designed them to. They always cut across.

And now, with AI agents that receive an intent and return a result, the hierarchy is becoming invisible to the thing using it. The intent is everything.The cleanest way I can say it is this: the best design decisions I've made all came after I was wrong about something. The worst ones came when I was sure. brain.space was two years of being usefully, productively wrong — and it taught me more than any project where I was right from the start.

“The part I’m proudest of is the part that didn’t need me to stay: the model still tags every session.”