Timeline
Jan 2026 – June 2026
Noteweave
- Product Design
- Brand
- Marketing
- 2026
Long story short
I led design for Noteweave: AI co-scientist as an IDE extension for business that can take the context of a business problem and use latest research to propose product specific improvements.
Designed and led frontend dev for Noteweave extension (in Cursor, VSCode and Windsurf), Noteweave user dashboard webapp and a full UI revamp.
Impact & Outcome:
We built Noteweave with Founders Inc in their Canopy program, where:
- It grew to 150+ users in 2 months with users from Huggingface, IIT's, BITs,
- Won $50k grants from OpenAI, Microsoft, Anthropic and Grok for startups and AWS activate to build Noteweave
- Collaborated with Wavelength, Diagnosis and Audria for accelerating their research efforts using Noteweave.
Why rigorous RnD needs a new tool:
Manual RnD becomes increasingly difficult as AI continues to push more research than business teams can use.
Over 70% of 1500 researchers failed to reproduce another scientist's experiments, and over 50% could not reproduce their own.
This makes scientific rigour erode under pressure to ship outputs.
Discovery: the scientific method is the same everywhere:
We talked to AI/ML, bioinformatics, physics, and pharma teams at research institutes and businesses.
And found out the scientific loop looks the same across cross-domain teams:
Hypothesize → experiment → analyze → export artifacts with a verifiable trail
Hypothesis: Build agent systems that improve in performance as frontier models scale in reasoning capabilities.
These architectures will have inbuilt skepticism as a feature that eliminates unreliable research and experiments upfront.
Automate the idea to experimentation loop with a human in the loop and verifiable records.
And we called this Noteweave :)
Process: Designing the core research loop
Every session follows the same scientific rhythm inside the IDE. We mapped triggers, outputs, and Noteweave's role at each stage so scientists shared one mental model.
| Stage | Trigger | Output | Weave's role |
|---|---|---|---|
| Understand space | User input | Weave chat output | Intervenes when data ingestion is incomplete |
| Create hypothesis | User approval | Structured hypothesis | Surfaces candidates; user edits or rejects |
| Perform experiments | User approval | Experiment data | Live progress, early failure detection, full logging |
| Analyse results | Run completes | Findings | Summarises results and proposes next steps |
| Export findings | User intent | Research paper + IP log | Blocks export on null data with clear errors |
Design Decision: Human first thinking
Human-in-the-loop at every expensive step:
Noteweave can draft hypotheses and run literature analysis autonomously, but experiments require explicit approval.
Scientists told us they would not trust a tool that ran GPU jobs or published claims without a checkpoint. Autonomy only where it is safe.
Iteration 1: Context aware search
What changed
A major challenge was hallucination. In Auto Mode, the LLM often drifted into unrelated use cases, wasting tokens and generating irrelevant outputs.
Outcome
We introduced a context agent that asks a few targeted questions around the user's query to tightly bound the context. These questions remain optional and open-ended, giving users full authority over the research direction.
Iteration 2: Surfacing .md files for better visibility
What changed
Representing large amounts of data from literature analysis inside the chatbox became cognitively overwhelming. We shifted detailed outputs to `.md` files while keeping summaries in the chatbox. However, users often missed the generated `.md` files within the IDE's file structure.
Outcome
`.md` files now appear as code files that open automatically when created, reducing friction and making them easier to notice. The scientific loop continues running in the background while users engage with the generated `.md` files and retrieved research papers.
Iteration 3: UI Revamp
What changed
The bright colours gave the extension a differentiated identity, but since research involves hours of reading inside an IDE, we shifted to a calmer colour palette.
Outcome
Revamped the UI with softer colours while retaining hints of the primary palette to keep Noteweave distinctive. Added the logo to every answer to improve recognition within the cluttered IDE interface.
.jpg&w=3840&q=75)
Where Noteweave fits in
Noteweave stands as one of the early prototypes of the AI co-scientist deep techs and AI labs [ core automation, periodic labs, Sakana AI, general intuition] are working towards automating scientific discovery.
When we stopped iterating
We worked in a tight feedback loop with Wavelength on one of their research use cases, iterating closely until the case study was successfully completed.
What I'd do differently:
- Narrow to one ICP and one domain before designing for every R&D vertical.
- Run structured interviews before the full UI revamp
- Focus on Instrumenting the core loops so that habits that drive return visits.
- Prove retention in one IDE before splitting focus across VS Code, Cursor, and Windsurf.
Learnings: What I took forward
- Stay close to your users: Weekly calls with early adopters helped in catching edge cases sooner
- Breaking internal biases is important: The biggest blind spots often come from our own assumptions. Constant user feedback helps challenge and correct them.
- Users don't care about the architecture: Even the most technical people tend to value usability more than the backend architecture/novelty.
- Designing for the AI black box: In research, reliability comes from being transparent about what has been omitted upfront, so users don't worry about missing important information.
