AI imaging SaaS platform for biotech (Cryo-EM) research
My Role
Founding Product Designer at $4M-funded AI BioTech Startup
As the principle designer on a deep-tech team, I collaborated closely with AI scientists to deliver a browser-based AI imaging SaaS platform for biotech (Cryo-EM) research within 3 months:
Spearheaded the 0-to-1 product lifecycle and drove a rapid agile development process with engineers, launching a browser-based AI imaging SaaS platform for biotech (Cryo-EM) research within 3 months.
Streamlined complex multi-step workflows for 200+ structural biologists, enabling over 2M image processing runs with up to 40% fewer manual steps.
Enabled the company to secure multi-million dollar funding from MiraclePlus.
Type:
B2B, Web Design
Timeline:
Sep 2024 - Dec 2024
Responsbilities:
Define the problem and use cases;
Wireframe;
Build prototype for testing;
Iterate interaction;
Craft visual details.
Deliverables:
Interactive prototype;
High fidelity mocks;
Design files & doc;
Presentation.
200+ active users
2M+ tasks processed
A next-gen AI platform that streamlined
cryo-EM workflows into a collaborative, task-centered system.
I.CONTEXT
Almost ALL Cryo-ET labs worldwide rely on IMOD for image processing. It remains the standard because of its obvious advantages: Accuracy, Maturity, and Stability.
IMOD
, developed in the late 1990s by David Mastronarde’s team at the University of Colorado Boulder
" IMOD is perhaps the best known and most used programs for tomographic reconstruction, with well over two hundred sites using it, and the reason for this popularity is because it offers several big advantages:
(a) it's free ....
(b) it works on any platform ....
(c) it's been around for ages
(d) it's designed specifically with 3D electron tomography in mind ...
(e) it handles the whole electron tomography process....
(f) it's quite large and versatile .... "
——— Andrew Noske
( IMOD tutorial author and plugin developer)
Why IMOD is so popular?
But, it comes with a TRADE-OFF.
“Once IMOD is installed you'll quickly realize it has a VERY STEEP learning curve and a bunch of features you’d probably never find unless someone points them out.”
In practice, researchers find IMOD TIME-CONSUMING.
As Andrew bluntly put it:
II.PROBLEM
Why is it so time-consuming?
At first glance, two surface-level reasons stand out:
The tool is outdated —— its interaction model doesn’t match today’s perceptual expectations, which makes it hard to learn.
In the scientific research context, usability is rarely prioritized. Accuracy and engineering efficiency dominate.
Surface-level Reasons
Why does this matter?
We believe tools should co-evolve with their users. Cryo-ET researchers are pushing the boundaries of structural biology — there’s no reason they should be slowed down by outdated interfaces. And while engineering efficiency is critical, neglecting human efficiency lowers overall productivity.
III. CAUSES BELOW SURFACE
We visited the Cryo-ET lab at iHuman Research Institute for in-depth interviews and shadowing with target users.
Our outsider perspective became valuable. Instead of asking about pain points, I asked researchers to “walk me through your day/workflow.” This helped us objectively surface problem spaces they had normalized.
Interestingly, researchers often blame themselves instead of the tool — they adapt to its quirks rather than questioning the system. So when we asked directly, “ What are your pain points?”, answers were vague.
Challenges
Opportunities
Current Cryo-ET Workflow with IMOD
We mapped IMOD’s role in the workflow and identified 4 key bottlenecks:
Fragmented workflow: reliance on command-line tools and frequent context switching.
Black-box processes: manual fiducial tracking and particle picking, dominated by trial-and-error.
Low automation: large Cryo-ET datasets with long computations, errors discovered only post-run.
Weak collaboration & reproducibility: scattered parameters/results and workflows limited to single users.
Key Findings:
After figuring out key questions to target on, instead of applying a universal design strategies to all problems, I adapt unique design methdologies to each of the question flexibily and nichely:
IV. DESIGN IDEATION
Design Strategy：
Wrapping command-line operations into guided UI interactions
Shifting from window-based to task-centered navigation
Problem 1
Key Decision 1 — Migrating Design Logic for a Fragmented Workflow
Researchers had to juggle
a Fragmented and
Error-prone workflow
Due to:
1.command-line tools
2.multiple windows
3.online tutorials
Insights
Across these references, the key strategies are:
Wireframe Design (Partial)
Approach: Case Studies
Instead of applying generic UX fixes, I examined how modernized scientific and creative platforms streamline complex workflows. Case studies included:
Weights & Biases
CoreWeave
RunwayML
Task-centered dashboards connecting multiple stages of ML workflows
Transforming low-level code operations into intuitive visual actions
“Puts machine learning in the hands of creators.”
“Bridges experiment to insight.”
“Turns complexity into clarity.”
AlphaFold Server
Google DeepMind
Abstraction of technical complexity through pre-configured pipelines
Movie-2
Export
Previous
Next
Parameters/Info
Task Index
Status
Index
Movie-1
Movie-2
Movie-12
Movie-13
Movie-11
Movie-10
Movie-9
Movie-8
Movie-7
Movie-3
Movie-4
ctf_estimation_settings: {defocus_step: 100, max_defocus: 50000, max_res: 5, min_defocus: 5000, min_res: 30,…}
global_settings: {acceleration_kv: 300, amplitude_contrast: 0.07, binning_factor: 2, p_size: 0.41,…}
motion_correction_settings: {eer_fraction: null, eer_sampling: null, patch: 5, save_as_float16: false, software: "MotionCor2"}
eer_fraction: null
eer_sampling: null
patch: 5
save_as_float16: false
software: "MotionCor2"
Movie-5
Movie-6
Quality
Movie-2
Status
Analysis
Message
Task Index
Status
Index
Movie-1
Movie-2
Movie-12
Movie-13
Movie-11
Movie-10
Movie-9
Movie-8
Movie-7
Movie-3
Movie-4
Movie-5
Movie-6
Quality
Task Center
Filter
ID
Task Name
Owner
Created Date
Progress
Status
Options
Date
Owner
Status
Cryo-EM ID
Create New Task
Create New Task
Experiment Management
Global
Motion Correction
CTF Estimation
Back
Continue
Software
Raw Movies Directory
Pixel Size
Total dose per movie
Acceleration Voltage
Spherical Aberration
Binning Factor
Note
2.Encapsulate command-line inputs into intuitive selectors, fill-ins, and buttons.
1.Frame the entire platform around task flows rather than tools.
1.Confidence bars & color-coded overlays showing prediction certainty.
2.Dual-view comparison mode showing human vs. AI-picked results.
3.Inline feedback prompts for quick validation or override of AI results.
Problem 2
Key Decision 2 — Redefining Human–AI Collaboration
Manual particle picking and fiducial tracking led to low automation. Errors were often detected only after long computation runs.
Comparison View Design
Picking threashold adjustment in
AI Preliminary Results View
Comparision between Pre-processed
Results and AI Preliminary Results
DRACO —— Cryo-EM Foundation Models for Automated Pipeline (developed by team Cellverse AI)
Approach: Adopt Emerging AI Tech
Instead of treating automation as a patch, I reframed it as a redesign of human–AI collaboration. I studied our lab’s newly developed automatic picking algorithm, understood its technical constraints and input–output patterns, and translated these into UI logic.
Design Strategy
Dual-view Design：
Goal & Implementation
Design a new AI-integrated interface that:
Communicates AI progress and confidence levels clearly
Allows users to fine-tune thresholds or re-run subsets
Reflects new modes of division between human judgment and machine precision
Design Strategy
1.Defining how the new interface should look and function —— clustered stakeholders’ descriptive needs into clear design to-dos
2.From the seminar, I gathered client-user needs (structural biology researchers) and the development team’s technologies to be integrated, and clustered them into design modules.
Design Translation
Mapped these needs directly to UI modules:
Process transparency panels replacing opaque “progress bars”
Collaborative dashboards for shared experiment views
Parameter snapshots ensuring consistent reproducibility
-Black-box processes made users uncertain about system status
-Collaboration was limited because parameters and results were scattered or siloed.
Problem 3 & 4
Key Decision 3 — Translating User Needs into Transparent & Collaborative
Approach: Semi-structured Interview
I conducted interviews and workflow tracing to identify concrete user needs:
Visibility into process and error sources
Versioning and reproducibility across experiments
Shared parameter sets and visualizations
Affinity Mapping
V. INTEGRATE USER FLOW
From scattered points to create a new workflow narrative
Suggested a new user flow that addresses key user needs by enhancing task efficiency and collaborative management, integrating login, user management, a dashboard task center, and dual Cryo-EM/ET pipeline processing.
User Flow Disgram:
VI. DESIGN & FAST PROTOTYPING
Delivered a next-generation, AI-integrated cryo-EM platform that transforms fragmented workflows into a collaborative, task-centered experience.
Efficient task management that adapts to user habits:
preserved frequently used key parameters, while deprioritizing low-value image details into a collapsible list to reduce cognitive load.
Seamingless AI integration:
In the noise reduction workflow, simplifying front-end parameter tuning into a single bar control. This enhanced user sense of control while enabling direct comparison with the original image.
User Management
Easily bring lab partners into the loop and collaborate.
Status
Enable on-time human intervention.
Options
Simplify traditional complex editing.
VII. IMPACT & FEEDBACK
The platform received strong positive feedback from researchers at the iHuman Institute, who reported substantial gains in cryo-EM processing efficiency. It is now adopted by over 200 researchers, powering 2M+ processing runs, with validation testing ongoing and additional data forthcoming.
VIII. REFLECTION
Immersing in an unfamiliar scientific domain revealed how disciplinary silos constrain innovation — design became the bridge translating complexity into clarity.
Collaboration with AI researchers reframed design from a service layer to a cognitive framework that shapes how scientists think and interact with their tools.
Co-developing shared language between data scientists and end-users showed that innovation emerges when empathy meets rigor — where human insight guides scientific precision.
Teamwork Timeline
Product Innovation @Siemens Healthineers
B2B | Strategy Design | Product Design | Case Study
User Acceptance Testing
@Danone
B2C | UX Research | User Interview | Insights Alignment
AI Product Innovation
@UC Berkeley
Entrepreneurship | VUI | UXD | Emerging AI Tech | Marketing
UXR & Inno-Design
@Siemens Healthineers
Field Visit | Contextual Inquiry | Insights Alignment| Concept Design