I build AI that
holds up in the
real world.

I built Clearance because I kept noticing the same gap: benchmark scores and real-world reliability are not the same thing, and teams were finding that out the hard way. It's an evaluation platform that tests whether LLM agents are ready for unsupervised operation, running real API calls with edge-case injection across GPT-4o, GPT-4o mini, Claude Sonnet 4.5, and Gemini 2.5 Flash. The largest gap I found between benchmark and production was 61 points. That number is what the platform is for.

• •Live evaluation pipeline making real per-trial LLM API calls with randomized edge-case injection, tracking per-trial latency and cost for cost-vs-reliability tradeoff analysis.
• •Scoring system separating benchmark accuracy from real-world reliability, surfacing gaps as large as 61 points between the two.
• •Automatic Cleared / Supervised / Not-Ready classification with configurable thresholds, tracking unsafe or irreversible actions as a distinct failure category.
• •Full reporting layer with dashboards, leaderboards, and expected-vs-actual failure breakdowns to make results auditable.

While looking at Everpure's published case studies in healthcare and banking, I noticed there was no clean way to separate regulated data from data safe for AI use. So I designed the Trust Passport, a portable metadata record that existing systems can act on automatically, covering sensitivity, legal basis, retention, AI eligibility, and protection tier. I checked that it holds under both HIPAA and GDPR/DORA. This is the kind of problem I find genuinely interesting.

• •Identified a real gap in Everpure's published healthcare and banking case studies: no automated way to separate regulated data from data safe for AI use.
• •Designed the Trust Passport, a portable metadata record (sensitivity score, legal basis, retention, AI-eligibility, protection tier, carbon budget) that five existing Everpure systems can act on automatically.
• •Proved the same architecture holds under both a US sectoral law (HIPAA) and an EU rights-based law (GDPR/DORA), evidence the pattern generalizes to any regulated industry.
• •Benchmarked against named DSPM competitors (BigID, Varonis, Cyera, Securiti) to show none pair classification with the physical storage layer the way this design does.

I built this to answer one question: does explaining AI usage in plain English, the way a budgeting app explains spending, change how people use these tools? The routing is deterministic, the caching is verified against real requests, and the dashboard is careful to separate actual savings from estimated ones. Getting that last part honest felt important.

• •Deterministic keyword/regex classifier decides routing with no LLM call spent on the decision itself, then routes to the cheapest model tier that fits the task.
• •Exact-match and semantic caching (Redis Stack, similarity-tuned) verified end-to-end against real requests, confirmed via matching completion IDs on repeated queries.
• •Full-stack dashboard (Express, Drizzle, React) clearly separates real dollars saved from estimated opportunity on flat-subscription tools, never conflating the two.
• •Built to test one question: does seeing AI usage explained in plain English, the way a budgeting app explains spending, actually change how people use these tools.