JEANNIE DUARTE

Employee since 2021, and consultant and subject-matter expert from 2014. The company's only technical staff member and, since 2023, the sole remaining holder of the platform's complete technical history. HAPselect is a candidate and employee assessment platform: validated instruments, automated scoring, configurable reporting, client integrations. Owns system architecture, data architecture, technical specification, and the roadmap, and directs a contracted offshore development vendor and an infrastructure vendor against those specifications. The vendor writes the code; the specification says what the code must do, and the sign-off says whether it does.

Sole responsibility for the delivery function since 2022. Scope, specification and design, then handoff with expectations set at handoff. Priority and sequence for the team's work, and the project board with them — every JIRA ticket, feature and bug alike, personally written. Stand-ups twice weekly since the launch into Azure, often three or four times weekly, carrying the migration from score-based to percentile-based ratings. Follow-up through implementation, pre-release QA, and release sign-off.

Designed and specified the platform's scoring layer with the company's psychometricians across two paradigms — ratings applied at the score level with the founding psychologist, then the conversion of all scoring and calibration to percentile ratings with the R&D lead. The Overall and Potential algorithms are of this design, along with an original multi-pass percentile rescoring method in which composite percentiles are re-normed against the benchmark. A distinction held throughout: scores are measured, ratings are the interpreted layer, and the adjustment mechanisms adjust ratings, never scores.

The reporting layer is composable and client-configurable — elements mixed and matched by trait type and rendering model, then stacked, with narrative driven by a prioritized-trait comment mechanism that returns database-configured comments in priority order. In production for more than ten years. Chief API architect for client integrations: Workday, iCIMS, other ATS and HRIS platforms, a bridge integration for a client HRIS that offered no integration surface, and the general API consumed by clients.

Design authority for every user surface — candidate, client, and administrative — across the legacy platform and the redevelopment alike. The fully documented instance is the redevelopment's candidate experience, researched and written here: a primer for the designer on selection regulation and applicant-reactions theory, a sourced research brief, a surface-by-surface design direction covering all eight candidate surfaces, and working HTML mockups rather than pictures of screens. What makes it assessment design rather than web design is that several interaction patterns are locked by measurement validity — the forced-choice direction flip, the timer on a speeded cognitive instrument, the four-point Likert with no midpoint — so the design moves only where moving it cannot change what the instrument measures. WCAG 2.2 AA throughout, with screen-reader testing scheduled inside the phase rather than after it. Co-branding is a constrained theming layer rather than a per-client fork, and no candidate-facing string is hardcoded: locale is a rendering parameter and every label resolves from the content collections. Design authority returned in-house when the vendor's design resource did not engage with the handoff — the build specifications, tokens, component decomposition, flow specifications, the co-brand contract, and accessibility annotations are produced here — and the vendor replaced the resource with a materially stronger one.

Designed the platform-modernization decision for a non-technical owner. The vendor's single estimate came bracketed with a conservative alternative and an aggressive one and a plain-language decision framework — budget ceiling, timeline criticality, risk tolerance, feature sufficiency, coordination capacity — so the owner could exercise real judgment on her own terms. Scope held at the application layer. Azure infrastructure specified as a complete resource bill of materials, with the infrastructure vendor's quote validated against market pricing and the variance published in the analysis rather than buried. Migration sequenced by blast radius — candidates, then clients, then administrators — with staged cutover and rollback. Acceptance testing routed to psychometricians rather than developers, so the people who know whether a score is correct maintain the tests. The first risk named in that proposal was reliance on one person's knowledge; the proposal was hers.

hapbi, the company's internal analytics and operations site — billing generation and export, pricing with change history, client and activity reporting, and a candidate right-to-be-forgotten module built in visible phases — was built solo, human-directed and AI-assisted from concept to implementation. Node/Express on Azure, Entra ID and Key Vault, SQL Server and MongoDB Atlas, CI/CD with staging-slot swaps. One author, 195 commits. Development paused since May 2026 pending a decision on the platform's historical data.

The automated assessment line is being retired in autumn 2026. Data-preservation and retention analysis through the decommission is led here, and this is the named technical contact for clients' integration shutdown and data-export requests.

Two legally separate state-licensed Texas child-placing agencies sharing staff and systems. Established both corporate entities, with technology, operations, quality, security, and board roles since. This is the second half of a working life, not a side project, and it is where the discipline described above runs in a second regulated domain — with no team and no vendor.

Architected and hand-built Radius, the agencies' case-management and document platform, in VB.NET and SQL Server from about 2004, in continuous production since, and touched by no other developer — architecture, data model, and the whole of its interface. A registry-driven form model resolves and instantiates forms at runtime; Microsoft Graph integration carries per-child shared mailboxes; a React-in-WebView2 bridge retired InfoPath. Its front end is now migrating to Pulse on Node/Express and Azure.

What the agencies run on is a layered platform with a versioned source of truth at every level, not a single application. An artifact-definition catalog classifies every document the agencies can hold — 1,610 definitions with a full taxonomy. A data-source registry carries an audience ceiling on every source and verbatim-only guarantees where they matter. A playbook registry binds document types to executable governance rules with real supersession history. A git-versioned policy knowledge base, built from scratch in late 2025, serves as human reference, API payload, and AI-agent context at once. A Model Context Protocol server exposes the platform to AI agents as first-class tools, with transactional seeding: begin, preview, commit, rollback. All of it implemented by directing AI coding agents against written specifications and acceptance checklists.

AI governance is machine-enforced rather than documented. Nothing AI-generated becomes operative without a named human approving it, and that rule lives in the rendering layer: briefs ground in evidence, flag gaps rather than fabricate, redact before the model sees anything, and hard-fail on an unresolved token. Audience ceilings live on the data source itself, so a mis-scoped prompt cannot leak — the source is simply not resolvable for that audience. Every credentialing playbook runs recommendation-only. The policy a person chose was kept separate from an architecture that would foreclose the alternative, and the reversal seam was built deliberately.

Led the agencies' end-to-end readiness for Texas's T3C redesign, the first overhaul of the state's foster-care service-level and reimbursement system since 1988. The 496-page Blueprint was decomposed into implementable service-package specifications with page-level citation traceability; a 120-point caseload weighting index reconciles nine incompatible state-mandated ratios into one governable number; the workforce model, the program-evaluation framework, and the IT security and data-governance policy suite were written alongside it. State credentialing reviewers approved the documentation, and the agencies hold interim-active credentials across the requested service packages. After submission the regulator issued a 143-item enhancement request citing a newer edition of the Blueprint; every item was classified against the review form's own history columns and the documents already in the regulator's possession, and none of the 143 constituted a new requirement. Approval followed. In the next wave that instinct became a governed intake step, and the count of net-new work was zero.

Led the technical and operational elements of achieving Council on Accreditation (COA) accreditation in 2010–2011.

A graphic design, print production and vinyl installation firm with two owners and complementary functions. Operations, finance, tax registration and monthly filings in two states, pricing structure and market validation, contracts and policy, permitting, systems, client-facing pricing documents, and crew-pay reconciliation sit here; creative, fabrication, install and the customer relationship sit with the partner. Earlier, creative direction across the full range — print, signage, apparel, billboards, and digital. The LLC predates the dba and carried general software consulting from 2013 to 2019 alongside the Grant Thornton and Knox Technology years; it took the Moxie dba in 2014 and The Golden Thread dba in 2026.

Pricing is built as a system rather than a number — rates derived from stated rules, the boundary between standard and specialty work defined by a stated test, market-validated and reviewed on a schedule, with one model rendering for three audiences. A loss on a specialty job became two standing rules and a signed-agreement requirement, and the systemic fix was named as such rather than left to memory.

Designed a local-first message-classification architecture for a business that runs largely over text. The obvious fix was to point a cloud model at the message database; that was refused, because cloud models read every message they touch, including personal threads. Instead: an on-device model, raw content never leaving the machine unless already judged confidently business, and classification running before sender identification — so the privacy gate is the model's verdict rather than the contact list, and a message that does not qualify never has its sender resolved at all. Real business messages will sometimes be missed; that price was written down in advance and chosen. The operations app (Next.js, Auth.js, Airtable, Vercel) is built and running locally, not yet deployed; a daily morning digest with auto-reply drafting is deployed and running.