Kelly Michels

The shared foundation under the products above: authentication, per-tenant data isolation enforced at the row level, and an API designed as the platform's core rather than bolted onto a finished app. A common SDK means a new product ships in days instead of months. Proving out now against evo.provensheet and evo.docketmail, and being prepared for release to the open-source community.

Environmental, health & safety compliance for multi-plant industrial manufacturers. Onboards a facility from a company name alone — auto-resolving the EPA ID and importing registered facilities, SIC codes and permits from EPA and Texas TCEQ. Runs the permit lifecycle, OSHA-aligned incident management with CAPA tracking, training matrix, audit binders, plant health dashboards, Stripe billing and PDF reporting.

Proof the platform generalizes past its first domain. Answers against the Texas statutes, every response cited to a specific section — hierarchy, source passage, deep link to official text, snapshot date. A citation gate rejects any answer that cannot be traced back to retrieved statutory text. Runs entirely on a local model.

Turns the spreadsheet a business actually runs on into a shared web application — and proves the numbers survived. Drag in an Excel workbook and it is a working cloud app seconds later: formulas, not screenshots. Every formula is then recomputed and compared against the result Excel itself cached in the file, so the maths is verified cell by cell before anyone relies on it — the report is the product, where everything else in this space asks you to trust the conversion. The built-in sample is an 18-sheet industrial estimating workbook refined over 30 years of in-field use: 5,121 formula cells, all matching. Local-first, so edits survive a dropped connection.

A documented defect loop against text-to-SQL, retrieval and guardrail failure modes — every defect with a reproduction, root-cause analysis, regression test and verification against ground truth. Guardrails are only real if they are tested: a regression suite of off-topic refusals, prompt-injection attempts and SQL escape attempts (stacked statements, schema escapes, write keywords) that must all be rejected; a retrieval eval harness holding golden-set hit rate with the guardrails on; and an LLM-as-judge suite that gains a case per closed defect, so the guardrail surface is regression-locked as it grows. Every gate run ties back to the commit that produced it. This is the part most AI work skips, and why the platform is trusted in production. evo.zscripts and evo.locate (both open source): one-word deploy tooling for a multi-project AWS stack, and a self-hosted geolocation API with no key and no per-query cost. Everything runs on one nginx edge with self-hosted mail, Git and metrics — a single AWS box, no managed PaaS.