My role
Co-Founder, Design Lead, Design Engineer
- Interaction Design
Noteweave: AI scientist for academic researchers
I designed a human-in-the-loop system for AI that can automate scientific research.
Premise
The core challenge was keeping researchers in control of an autonomous loop
Scientific research is becoming increasingly computational. AI can now support an entire research loop, but researchers still need to understand and control what happens at each stage.
- Problem defining
- Literature review
- Hypothesis generation
- Experimentation
- Drafting
Product
I placed Noteweave inside the IDE to reduce context switching
I decided Noteweave will sit as a plugin inside coding IDEs like Cursor because AI/ML researchers work in IDEs, this helped in reducing context switching and leveraging familiar workflows.
Constraints
Technical constraints shaped the interaction model
Limited UI space inside an IDE
I used progressive disclosure to surface only relevant information, improving readability and reducing cognitive load.
1 of 3
High compute costs across the research loop
I paired a compute-conscious architecture with BYOK, reducing unnecessary usage while giving users greater cost visibility and control.
2 of 3
Fragmented authentication across browser and IDE
I reduced onboarding steps and preserved login state, making setup faster with fewer interruptions and context switches.
3 of 3
Discovery
I mapped the work AI performs behind the screen
Formal user studies were not feasible in a startup setup, so I started by mapping inputs and outputs for each transition stage.
Understand space
- Trigger
- User input
- Output
- Weave chat output
- Failure path / Weave role
- Incomplete data ingestion, weave intervenes in chat
Create hypothesis
- Trigger
- User approval
- Output
- Hypothesis
- Failure path / Weave role
- Approval denied or no hypothesis formed, weave intervenes
Perform experiments
- Trigger
- User approval
- Output
- Experiment data
- Failure path / Weave role
- Code/systemic/environment errors, experiments restart, Weave intervenes
Analyse results
- Trigger
- Successful or failed experiment
- Output
- Findings
- Failure path / Weave role
- No hard failure, weave intervenes contextually
Export findings
- Trigger
- User intent to share/export
- Output
- Research paper artifact
- Failure path / Weave role
- Null data in paper, error message shown
Autonomy map
I matched AI autonomy to risk and reversibility
Each stage receives autonomy based on risk and reversibility.
Experimentation
Hypothesis generation
Problem defining
Drafting
Literature review
The agent shifts from review to execution as risk increases.
| Execute | Explore | Scope |
|---|---|---|
| Researcher approves | Researcher reviews | AI acts |
| Handles Experimentation | Handles Problem defining Hypothesis generation Drafting | Handles Literature review |
Interactions
I used rapid prototyping and implemented a feedback funnel to make the hardest interactions.
1 of 10
I added a context step before autonomous reasoning
Old flow
A broad prompt immediately triggered autonomous reasoning.
New flow
A context agent asks a few targeted, optional questions to bound the research space before work begins.
2 of 10
I moved complex work beyond chat
| Old flow | New flow |
|---|---|
| Long-form research outputs accumulated inside the conversation. | Chat kept short summaries while detailed outputs moved into visible markdown files. |
3 of 10
I made AI state visible without adding noise
| Old flow | New flow |
|---|---|
The first version used bright colors to create a distinct identity.  | The interface moved from bright distinction to a calmer IDE-native presence.  |
4 of 10
A state machine kept long-running work understandable
| Old flow | New flow |
|---|---|
| The interface treated research as one continuous chat state. | We modeled scoping, exploring, approval, execution, analysis, failure, and export as distinct states.  |
5 of 10
Inspectable evidence made AI outputs easier to trust
| Old flow | New flow |
|---|---|
| Consequential actions were not consistently reviewable. | We made consequential actions reviewable and required explicit approval before the AI could run experiments or turn hypotheses into production plans.  |
6 of 10
Failure states explained what happened and what to do next
| Old flow | New flow |
|---|---|
| Failures behaved like dead ends with no clear recovery path. | Failures were treated as product states: Noteweave surfaced constraints, preserved the research trail, and made the next recovery action explicit.   |
7 of 10
The chat input made research controls visible before work began
| Old flow | New flow |
|---|---|
| The input captured a prompt, but agent behavior and execution mode stayed implicit. | The redesigned composer brought the research goal, agent choice, progress, and Plan mode into one place.  |
8 of 10
Files preserved context across the research loop
| Old flow | New flow |
|---|---|
| Failures were framed as technical dead ends. | Assistive, warm, BUT pragmatic  |
9 of 10
Opening Noteweave felt native to the IDE
| Old flow | New flow |
|---|---|
| Starting research required moving away from the active workspace. | Noteweave opens automatically where the rest of the agent window lives.  |
10 of 10
Modes balanced guidance with researcher control
| Old flow | New flow |
|---|---|
| The agent's degree of autonomy was implicit. | The modes translated autonomy into an explicit user choice: Auto lets the agent self-plan, Permission asks before every tool call, and Plan writes the plan before execution.  |
Result
The final design outcome
Outcome
The final prototype connected the complete research loop
Within two months, Noteweave reached 150+ users across AI/ML, research institutes, universities, and scientific R&D; it collaborated with Wavelength, Diagnosis, and Audria, with support from OpenAI, Microsoft, Anthropic, AWS, and Founders Inc.
Learnings
What I took forward
Autonomy is a UX decision
Match AI control to risk, not model capability.
Stay close to users
Real workflows expose edge cases early.
Question internal biases
Use questions to challenge assumptions.
Prioritize usability
A clear workflow matters more than backend complexity.
Make uncertainty visible
State what the AI knows, omits, and needs next.
Paridhi was my Co-founder in 2 startups, which we scaled to users from 30+ countries. She has phenomenal grit and thinking. You are at a loss if you do not have her in your team!
