Engineering leader for government digital services. Director of Engineering for CiviForm at Exygy.

I've built and led engineering for benefit applications, a statewide career platform, Medicaid and CHIP reporting to CMS, and California's energy-data filings. I learned the work inside California government, and I still write code.

chelseakr.com · Resume (PDF) · LinkedIn · Davis, California

What each program is, who it serves, my role, and what the public record shows. Results belong to the teams that built them; my part is stated as scope.

• CiviForm (Exygy, 2026 to present). Open-source software that lets a government offer one application for many benefit programs, reusing what an applicant has already entered. Four city and state governments run it in production, including Seattle and Arkansas. I have been Director of Engineering since August 2026, leading engineering for the platform. The four-person team ships on a biweekly release train.
• MyCareer.NJ.gov (Coforma, 2022 to 2026). New Jersey's statewide career and training platform, in English and Spanish: career exploration, training programs with their outcomes, and job search. I was engineering lead and principal engineer from an early prototype to a statewide service, and primary engineer across its three production codebases. It has recorded 1.8 million active users since December 2023. I did the 2023 data modeling that began the state's move to Credential Engine's CTDL standard, and its training programs are live in the Registry today.
• Medicaid and CHIP Data Collection Tools (MDCT) (Coforma, 2025 to 2026). CMS's open-source suite of seven applications that states and territories use to report Medicaid and CHIP program data. I owned engineering for the suite within Coforma's healthcare portfolio: direction, standards, and the conditions for delivery. Feature leads did the hands-on build. The team took a new CMS Rural Health Transformation application from zero to first production in about eight weeks, and all seven applications shipped production releases in 2026 while I led the portfolio.
• Medicaid Drug Programs (MDP) (Coforma, 2025 to 2026). The CMS system behind the Medicaid Drug Rebate Program. Drug manufacturers report product and pricing data through it, and CMS uses that data to calculate the rebates owed to state Medicaid agencies. About 780 manufacturers participate. My role was oversight within the same healthcare portfolio; a partner's technical leads ran implementation.
• Data Submission Portal (DSP) (California Energy Commission, 2019 to 2022). The Energy Commission's secure, cloud-based portal for filing regulatory energy data, starting with petroleum and quarterly fuel-and-energy reports. As Lead Software Engineer I architected and shipped it on AWS and helped lead the commission's cloud modernization. The commission later expanded the platform across the agency and into residential solar-permit reporting.

I lead managers and senior engineers, and I stay close enough to the architecture to answer for it.

At Coforma I was a Senior Director of Engineering, one of three in a 50-person engineering organization, and I owned the company-wide healthcare engineering portfolio. My reporting structure was 22 people, with five direct reports, three of them Directors. Nine engineers were promoted within it, four to Director or Principal, and three people left voluntarily in three years.

The practices I put in place cover planning, hiring, delivery, and growth. On MyCareer.NJ.gov, accessibility, security, and bilingual parity became part of the release path, and automated test coverage went from zero to 93 to 96 percent while security vulnerabilities fell 94 percent. At CiviForm I supervise two Google.org fellowship workstreams alongside the engineering team.

Independent work on my own time, built in the open on public data. The status beside each one says how far along it is. Most of my public repositories are measurement tools: validators, scorecards, and evaluation harnesses pointed at real transit feeds, hospital price files, health-plan APIs, credential records, college disclosures, and California energy filings. These are personal projects with no proprietary or client material in them. They are not CiviForm or Exygy work, and nothing in them speaks for either. Most are pre-1.0, and I would rather you take a repository's own word for its maturity than mine. State of the portfolio is the honest map, re-measured on 2026-09-18: what is installable, what is live, what is a synthetic demonstration, and what is waiting on a person rather than on code.

• Afterward (source), beta. 3,266 California training programs joined to the state's own job projections, in English and Spanish, with no account.
• Homeroom (source), live. California public school data for every active school, 10,534 of them, readable by the families it describes, in English and Spanish.
• Trout Truck (source), beta. California's trout planting schedule, with a stocking history for each water, rebuilt daily from the Department of Fish and Wildlife's weekly schedule.
• Next Home Game (source), live. Calendar feeds for women's pro and college sports, by league and team, updated nightly.
• Family Greenhouse (source), live. A shared plant-care journal for households: watering schedules, care tasks, and reminders. Free accounts and paid plans.
• Transit Delivery Atlas (source), early release. A source-linked crosswalk of California's transit executive order, N-7-26, directive by directive.

• GTFS Scorecard (source), live beta. Daily plain-language quality grades for more than 2,100 public transit feed records, with a GitHub Action and a read API.
• fhir-scorecard (source), beta. Daily grades for public FHIR health-data endpoints, with every finding cited to the spec.
• TODS Validate (source), beta. A validator for the Transit Operational Data Standard, as a command-line tool, a GitHub Action, and a browser playground.
• ctdl-validate (source), beta. Checks CTDL credential records against the published rules before they go to the Credential Registry.

Every project, with its evidence, is on chelseakr.com/work.

Every public repository here was built since June 2026, and AI agents are part of how I work. I choose the architecture, write the acceptance criteria, review the output, and decide what is ready to release. That is why this much exists in this little time.

It is also why so much of it is gates. The failure mode of working this way is a check that cannot fail and a green build that means nothing, so each project is built so that it can fail, and when it does the failure is written down where you reach it before you reach the claim.

These are independent personal projects with no proprietary or client material in them. They are not CiviForm or Exygy work, and nothing in them speaks for either.

These projects have very few stars. What they have instead is a record of being corrected in public.

A project's scope. Jannis (derhuerst), a longtime open transit-data maintainer, argued that GTFS Scorecard's scoring belonged inside MobilityData's canonical validator rather than in one more dashboard. He was substantially right. I named in the thread the tools I had duplicated, declined to push subjective letter grades into an official project where they would read as guidance, and narrowed this one to the handoff nobody else covers.

A scoring rule. The person who produces the MRC de Joliette feed pushed back on a recommendation to populate trip_headsign on loop routes. They were right, and the rule now credits that case instead of flagging it.

Upstream, MobilityData merged a specification example fix to the Transit Operational Data Standard and an awesome-transit listing. A conformance-language clarification, a second awesome-transit listing, and a Transitland feed-archival PR are open. Five small contributions, two of them merged. That is the whole claim.

There are more repositories here than anyone wants to browse. They fall into six groups, and each group has one worth reading first.

• Transit data. Start with gtfs-scorecard. Then tods-validate, fare-policy-assistant, transit-delivery-atlas.
• California filings and public records. Start with qfer-preflight. Then power-content-check, ca-tariff-parse, ceqa-preflight, permit-bearings, perimeter, wildfire-service-territory-overlap.
• Standards conformance. Start with ctdl-validate. Then oscal-validate, fhir-scorecard, mrf-honest, ctdl-validate-jvm.
• Education and workforce. Start with afterward. Then disclosed, homeroom, chalkline.
• Evaluation gates and receipts. Start with plumbline. Then gauntlet, cairn, sprout, outcome-receipts, obligation-receipts, exitdrill.
• Community and personal tools. Start with habitable. Then ledger, swelter, nearmiss, id-churn-sentinel, davis-bike-hazard-map, queer-the-stacks, encore, family-greenhouse, olive-bark-logger, lavender-rotation.

Most are pre-1.0, and I would rather you take a repository's own word for its maturity than mine. State of the portfolio is the honest map, re-measured on 2026-09-18: what is installable, what is live, what is a synthetic demonstration, and what is waiting on a person rather than on code.

• GTFS Scorecard (live at gtfsscorecard.org) grades more than 2,100 curated transit feed records daily in plain language. Correctness findings come from MobilityData's canonical validator, not a competing one. Because it names public agencies it publishes a listing policy, and its remediation handoff is badged Pilot because that pilot has not recruited a participant yet.
• tods-validate checks Transit Operational Data Standard feeds, the crew runs and vehicle assignments GTFS does not cover, against TODS v2.1.0, with rule IDs that are never renumbered and a GitHub Action so an agency gates a bad feed before publishing. Reading that spec closely enough to write the rules is what produced the two upstream TODS pull requests above.
• ctdl-validate structurally validates CTDL JSON-LD, the national credential-data standard, citing the published schema behind every finding, as a CLI, a GitHub Action, and a browser playground that uploads nothing. Run against 120 documents sampled from the public Credential Registry it returned forty ERROR findings, every one tracing to an inconsistency inside CTDL's own schema encoding rather than to a publisher's mistake. That is a finding about the tool as much as about the corpus, and the write-up leads with it.
• afterward joins 3,266 California training programs reported under WIOA to the state's own ten-year projection for the occupation each leads to, in English and Spanish, with no account and no tracking. A suppressed or unreported outcome never renders as zero, and how much the join actually covers is a published output rather than a footnote.
• fare-policy-assistant answers rider questions about reduced-fare policy over a corpus covering eighteen California transit agencies. The assistant exists so that the public evaluation harness has something to evaluate: 385 cases, versioned prompts, a committed regression baseline, merge-blocking refusal and grounding gates, and a standing record of what it has not measured, including that none of its 28 Spanish answers has been rated.
• habitable makes habitability evidence for tenant unions tamper-evident with content hashes and RFC 3161 timestamps, then syncs peer to peer under end-to-end encryption so there is no central holder to subpoena. Its own README says not to rely on it for real legal matters yet, and I agree with it.

A tool that grades other people's work has no business shipping a check that cannot fail. Hunting those down in my own repositories is a standing part of how I build.

plumbline is a fail-closed audit harness. A target that returned 174 empty responses scored 1.0000 on groundedness, privacy, representational harms, fairness and cross-language, and the gate exited PASS on that alone, because each of those checks is phrased as the absence of a bad thing and silence satisfies all of them. The fix splits the two kinds of check apart, and a test now fails the build if a silent target ever passes again.

That same shape turned up again and again elsewhere, every instance green at the time. ceqa-preflight printed four PASS lines over a package whose PDFs had all timed out, one of them affirmatively clearing a document it had never opened. A weekly monitor in id-churn-sentinel went green four weeks running while zero of its 152 sources were eligible to check, because observing nothing emitted the same signals as observing no change.

I have written the whole class up, with nine instances and the fix commits attached, in The confident zero. It includes the part that is least flattering and most useful: I built a static detector for this shape, it missed more of them than it caught, and its highest-recall rule is the one I cannot ship.

Before that I was engineering lead and principal engineer for MyCareer.NJ.gov, New Jersey's statewide workforce platform, where I set architecture across three production codebases, ran a zero-downtime GCP to AWS migration, shipped full English and Spanish parity, and did the 2023 CTDL data modeling that began New Jersey's migration of its training-program registry into the national Credential Engine ecosystem. Earlier I built public systems for the California Energy Commission, the California Public Utilities Commission, the California Department of Social Services, and UC Berkeley's Graduate School of Education. I am a registered member of Credential Engine's CTDL Advisory Group.

The domains I keep returning to are public health, workforce and social services, energy and utilities, state and local digital services, and responsible AI: work that helps the people public systems routinely fail.

• No weapons, warfare, policing, mass surveillance, or technology that profits from incarceration.
• No AI that decides whether a person gets a job, a benefit, a service, or an opportunity. It can support a human decision, as long as that human can inspect the evidence, correct what is wrong, and make the final call.
• Accessibility, privacy, security, operability, and multilingual delivery are engineering requirements, not a later phase.

CiviForm, and the repositories above. I am not consulting and I am not looking for a role. Reach me through chelseakr.com or LinkedIn.