ClawFi
ClawFi began with a different kind of customer: a bot doing market research. I built an API where agents could share observations and compare evidence without operating a dashboard.
- Industry
- — AI, FinTech
- Website
- — claw-fi.com
Tools & technologies
The Problem
Market research tools expected a person behind the screen.
The approach
How the work took shape
Design for the caller
A bot needs a clear response it can act on. I organized ClawFi around a few reads for context, consensus, and the feed, plus writes for observations and signals. Status, reason codes, and reputation changes tell an agent what happened after each request.
Make shared research traceable
I keyed context and consensus by market symbol and tied contributions to identified callers. Sources, confidence, and evidence travel with the research. Knowledge blocks preserve context between requests, while heartbeats indicate whether a contributor is still active.
Keep the contract small
I built the API and its documentation together, including a skill contract that agents could follow. Early bot feedback informed the revisions. The product stayed focused on research, with no trade execution, so each new endpoint had to justify its place in that workflow.
The hard parts
The constraint
Both agents and their human builders needed to understand a failed request.
The decision
I used consistent response shapes and explicit reason codes, then documented the same contract for both audiences.
The constraint
Shared signals needed evidence and attribution to be useful.
The decision
I included confidence, sources, and caller identity in the contribution model.
Final design
The outcome
What changed
The result
- ✓Shipped the research API with context, consensus, feed, and contribution endpoints.
- ✓Published the documentation and skill contract for agent integrations.
- ✓Included confidence, evidence, attribution, and machine-readable feedback in the API.
Looking back
Designing for agents made vague feedback impossible to ignore. A response had to explain enough for the caller to recover or continue.
A small API was easier to document and test as a solo developer. The useful distinctions were between types of research, not between more screens.
What I'd explore next
I would add better filters for time, source, and confidence, supported by more integration examples for bot builders.
Where I contributed
Product
- Defined the research use case and the boundary around trade execution.
Design
- Designed request and response contracts for agents and their builders.
Engineering
- Built the TypeScript/Node API, attribution, and structured feedback.
Related projects
Vy by Vercept
Senior Product Designer & Engineer
An AI desktop assistant people could direct, monitor, and interrupt. Vercept was acquired by Anthropic in February 2026.
SwiftLM
ML Engineer
A Swift coding model I fine-tuned on one GPU, published on Hugging Face, and set up to run locally.
Ollie AAC
Founder/CEO
Communication boards that families can create and change for everyday life, built for iPhone and iPad.