The interesting question isn't whether we can build something. It's whether it actually works.
Experiments in computer vision, AI evaluation, autonomous research, and systems reliability. Less hype. More evidence.
SECTION 01 / CURRENT WORK
No speculative product claims. A clear view of what's live, what's experimental, and what's still an idea.
A local-first video workbench for surfacing candidate anomalies, inspecting frame-level evidence, and separating automated observations from human judgments.
A read-only code scout. Give it a problem; it searches public repositories, inspects limited evidence, and reports what might be worth testing.
› RESEARCH TASK: BROWSER ROI
[01] SEARCH PUBLIC REPOS OK
[02] CANDIDATES FOUND 04
[03] GITHUB READS 17
[04] EVIDENCE REPORT SAVED
_
Revisiting vibration analysis and signal-processing tools through modern, testable software.
IDEA / NOT YET BUILTMaking sense of network conversations and anomalies on systems we own or are authorized to examine.
IDEA / NOT YET BUILTSECTION 02 / ENGINEERING LOG
Not press releases. What we tried, what failed, what the evidence actually supports.
Our browser smoke test stopped at video loading. Instead of changing the application blindly, we checked decoder support: that headless Chromium build didn't support H.264. Swapping the synthetic fixture to VP8/WebM let us test the actual workflow.
OBSERVED RESULT Both JavaScript regressions and browser smoke checks passed; real-world detector accuracy remains unverified. Detailed CI logs are retained privately.
OPEN LIVE PROTOTYPEThe first overly narrow repository search returned nothing. We widened the queries, kept the scope bounded, and retrieved four candidates in seventeen GitHub reads. A useful start—not proof any candidate is the right implementation.
OBSERVED RESULT Read-only discovery ran successfully. Model-assisted assessment still needs a live API test. The underlying source and CI artifacts are currently private.
SEE SPROCKET WORKBENCHAn object detector can mistake a printed figure or advertisement for someone physically present in a scene. Counting detections isn't the same as understanding the environment. Scene Context review makes it possible to flag and inspect those cases without pretending the model understood them.
LESSON Separate the observation from the conclusion—and keep the original detection evidence.
EXPLORE VIDEO FAILURE LABWhen a phone records a screen, measuring changes across the entire camera frame can dilute what happens on the display. The next experiment is reviewer-selected Region of Interest analysis: measure the relevant pixels instead of everything around them.
NEXT QUESTION Can a simple fixed rectangle improve reproducibility without adding unnecessary complexity?
SEE THE PROJECT ROADMAPSECTION 03 / THE OPERATING METHOD
Every good instrument has a measurement range. Every good evaluation has limitations.
Capture what the system actually did. No conclusions smuggled into measurements.
Reduce a complicated situation to a testable event, region, or failure mode.
Look for false positives, hidden assumptions, and alternative explanations.
Run the experiment again. A passing dashboard isn't a validated result.
Show evidence, uncertainty, and what the tool still cannot tell us.
Don't increase complexity faster than your ability to validate it.
SECTION 04 / THE OPERATOR
Engineering experience has a way of making you skeptical of confident-looking outputs.
My background includes industrial reliability, vibration analysis, troubleshooting, and AI evaluation. The Test Bench is where those instincts meet software: take systems apart, understand the measurement, build something useful, and test the hell out of it.
This is an experimental workshop, not a collection of guaranteed solutions. When something works, I'll say what we tested. When it doesn't, that's a field note too.
VIEW PROFESSIONAL PORTFOLIOTHE DASHBOARD IS
NOT THE GROUND TRUTH.