Clew Labs | Auditable AI for complex public systems

Civic technology · applied AI · public systems
I’m La Shara Cordero, a court operations leader, civic technologist, and applied AI product builder. I design working systems that make evidence, decisions, authority, and unknowns easier to inspect.
§ 01 · Selected work
These projects are the result of learning by doing—separating what a system observed from what it concluded, while keeping human authority visible.
Flagship · civic AI · readiness intelligence
The AI reads the record. It never assigns the score.
A cited public record readiness surface for California superior courts. A bounded agent locates and extracts evidence. A deterministic registry calculates the result. Every claim can be checked and every unknown remains declared.
GPT 5.6 · Tavily · Next.js · deterministic TypeScript engine · recomputable audit bundle · published methods
Why foundation comes before software when introducing AI to court and government systems.
Flagship · empirical product research · agent observation
Shows developers what retrieval agents experience when they read a website.
Deterministic checks, a published methodology, and a fifty site benchmark make the second audience visible. Built to inspect how autonomous agents parse structure, permissions, and latency across the modern web.
AWS Lambda · Amazon Bedrock · DynamoDB · Vercel · fifty site benchmark
The web has a second audience: autonomous agents. Inspecting what retrieval bots experience.
AI governance · access control
Reproducible least-privilege test of RBAC at an AI gateway. Three scoped roles against three model tiers, nine calls, plus a falsification test and a negative control.
Nine calls, three refusals, and a falsification test by a court access administrator.
Education · Personalized AI learning
A stateless, privacy-focused tool providing personalized learning paths through verified free resources. Formulates custom curricula based on learner goals and style.
Stop following the hype, start directing the search. Multi-agent curriculum synthesis on AWS.
Court service · policy alignment
Role-based AI readiness guidance for court staff. Evaluates specific court workflows against sensitivity guidelines to determine where AI can safely assist and where human authority must remain.
Why California court staff need role-based AI readiness self-checks that preserve human authority.
§ 02 · Operating philosophy
Trust is not a tone of voice layered over a product. It is the result of assigning evidence, calculation, uncertainty, and authority to the right parts of the system.
Consequential claims remain connected to their sources. Retrieval and inference are not treated as the same thing.
Rules, boundaries, failure behavior, and deterministic calculations belong in the architecture before a model is invited in.
AI can retrieve, extract, explain, and prepare. It should not quietly acquire the power to decide.
§ 03 · Selected writing
Documenting the architecture, failure modes, real costs, and engineering decisions behind working systems.
“I approved a prompt on a Tuesday morning and went to make a cup of Irish breakfast tea. By the time I sat back down, the agent had refactored forty-two files... My agents never get tired. I do. When generation cost drops to zero, human verification capacity becomes the bottleneck. At some point you have to know when a system is good enough—and when you’re just satisficing.”
“I approved a prompt on a Tuesday morning and went to make a cup of Irish breakfast tea. By the time I sat back down, the agent had refactored forty-two files... My agents never get tired. I do. When generation cost drops to zero, human verification capacity becomes the bottleneck. At some point you have to know when a system is good enough—and when you’re just satisficing.”
“A few weeks ago I hit ReAct in the coursework and had a record scratch moment. I was building autonomous agent loops on Bedrock, but every line of plumbing was generated. If the agent writes the tools that the agent calls... does the engineer still understand the system? Yes—because authority and architecture belong to the human.”
“A few weeks ago I hit ReAct in the coursework and had a record scratch moment. I was building autonomous agent loops on Bedrock, but every line of plumbing was generated. If the agent writes the tools that the agent calls... does the engineer still understand the system? Yes—because authority and architecture belong to the human.”
“I wasn’t going to enter. 33 apps built in 31 days. One self-contained HTML file each. No build step, no dependencies, no Math.random in any selection logic. Same input, same output, with the derivation shown. The whole month cost $44, and the write-up shows the receipts.”
“I wasn’t going to enter. 33 apps built in 31 days. One self-contained HTML file each. No build step, no dependencies, no Math.random in any selection logic. Same input, same output, with the derivation shown. The whole month cost $44, and the write-up shows the receipts.”
The code is newer. The systems judgment is not.
I spent over a decade inside court operations and government before I began building software. Clew Labs is my independent, woman-led laboratory for testing what responsible AI looks like when the consequences are real. I build to investigate: turning questions about authority, evidence, uncertainty, and public systems into working software that can be inspected.
“Who has authority? What evidence supports the claim? What happens when the data is incomplete? Who carries the consequence when an implementation fails?”
I build through a directed, review-intensive AI workflow: I define the problem, architecture, boundary constraints, acceptance criteria, and design. AI accelerates implementation. I inspect, challenge, and verify the result.