Work/04
Kineticly
AI-powered research and publishing platform.
Kineticly is an AI-powered research and publishing platform that I built in Python. Specialized AI agents research new science and engineering work and draft articles about it, and every article goes through automated checks and human review before it is published. It's live at kineticly.io.
- Status
- Live since October 2026
- Site
- kineticly.io
- Stack
- Python, CrewAI, SQLite
- Role
- I designed and built the platform, and I review what it publishes.

01Overview
Making AI-generated research more reliable
Language models can write clear, confident text that turns out to be wrong. For a site that explains real research, that is the main problem to solve, so most of my work on Kineticly went into the system around the writing.
I built the workflow on the assumption that any single model output could be wrong, incomplete or badly formatted, and added checks at each stage to catch those cases.
02Workflow
How the research workflow works
Each agent handles one task. Regular application code controls how work moves from one stage to the next and what happens when a stage fails.
Discover
Find new science and engineering research to write about.
Plan
A planning agent decides what needs to be established and where to look.
Investigate
A research agent reads the underlying sources and records what they support.
Draft
A writing agent drafts the article from the recorded research.
Validate
Structured data validation and citation checks. Failures go back for revision.
Edit and approve
An editing agent reviews the draft, and then I make the final decision.
◇ Review gateA draft that fails a check is sent back to an earlier stage or stopped.
03Validation
Validation and error handling
Structured outputs
Agent outputs have to match a defined structure, so malformed or partial responses are caught before they move downstream.
Evidence kept separately
Research findings and sources are stored in SQLite separately from the article text, so a claim can be traced back to what was collected.
Citation checks
Citations are checked in code against the stored sources.
Error handling
Stages can fail for ordinary reasons, like a source that will not load or a response that gets cut off. The workflow handles those cases so they do not end up in an article.
Human review
The automated checks reduce what reaches me, and I review and approve every article before it is published.
These checks help catch problems, but they do not guarantee that every article is accurate. That is why the human review step is there.
04Results
Results
Kineticly launched in October 2026 and publishes articles on topics including physics, robotics and computing.
Building it gave me hands-on experience with software architecture, automation, validation and error handling. A lot of it felt familiar from mechanical design: define the interfaces, check the work at each handoff, and plan for parts to fail.