Open research
I want people to be able to inspect how I approach a problem, what I built with AI and what happened when I tested it. These releases bring the source and the explanation together.
My role is to connect real operational needs with systems that can be examined and improved.
I shape the context, responsibilities and acceptance criteria, direct AI-assisted implementation and review the result. Each case below identifies the existing mechanism, the publication changes and the evidence available to a reader.
OBSERVER8
Can a workflow distinguish a missing source, conflicting instructions and a genuinely unfinished submission?
I developed bidding workflows around the decisions a bid manager has to make. An answer can sound careful while carrying an old deadline or budget into the next step. I wanted to inspect that gap, not just improve the wording of the answer.
Career operations research
What should a digital worker have to show before work can be marked complete?
I built a career operations system because finding work itself involves research, source evidence, preparation, approvals and follow-through. Using it made the difference between a submitted form and a confirmed result very concrete.
Operational reliability
Should a paid model request proceed when its usage-limit check is unavailable?
My research includes the ordinary controls that keep an AI workflow usable. A helpful draft is not a good result if the system ignores the rule that was meant to control its cost.
Creative tooling research
How can a small visual-production pipeline produce an inspectable result without depending on a live model service?
I explore ideas by making something I can see and test. One part of that work was a Python rendering tool that turns prepared images into a short moving sequence.
Telemetry research
Can a parser turn an incomplete response into a plausible measurement by joining unrelated context?
I explored diagnostic software to understand when a response contains usable information. The important failure was a short frame borrowing bytes from the next response and producing a believable RPM value.
Evidence and expiry research
What must a source record contain before it can support a current operational decision?
In Signal I separated discovering a reference from knowing whether a source can support a current decision. I made permission, freshness and quantity explicit in the software because each answers a different question.
A Python semantic behaviour runtime and modeled hardware constraints for expressive robotic creatures. The source includes packaged profiles and installed-wheel tests. These are software simulations, not measured physical-prototype results.
Inspect ParrotCoreThe website’s static publishing pipeline, attributed data import and Cloudflare backend are also inspectable. Public configuration is separated from the company’s production settings.
Inspect the research platformAn earlier public notebook of source-derived mechanisms, synthetic examples and implementation notes. Its application-specific history remains clearly identified.
Inspect the workbenchPublic source does not require publishing someone’s business records, private conversations or credentials. These packages use reviewed source and synthetic fixtures. Source maps describe the extraction instead of pretending that a cleaned example is an entire deployed system.
Read each package’s provenance, reuse notice and verification instructions. A passing local test is evidence about that test and implementation, not proof of organisational adoption.
Publication and evidence policy