The recordAssessment lens →

Jeannie Duarte

Assessment platform and data architect. Twenty-eight years of building the data, scoring, and reporting layers under validated selection-assessment instruments: the same instruments, through four generations of platform. Alongside that, a second working life co-founding and running the systems of a licensed child-placing agency, and a third as co-owner and operations principal of a graphic design, print production and vinyl installation firm.

This is the complete record, and it is long on purpose. A shorter view for assessment and psychometrics readers is at /assessment.

jeannie@goduarte.com · Remote, US Central

Map

Who I am, and what I am not

I am a data and scoring architect. For one lineage of validated psychometric instruments, I have designed the data layer under every platform they have lived on:

The application layers above came and went. The data and scoring architecture is the continuity, because that is where an instrument's fidelity lives. I am its author. Since 2023 I am also the sole remaining holder of its complete technical history.

I am careful about which position that puts me in.

I am not the originating scientist. I do not design methodology, author instruments, or sign validation studies. Those belong to psychologists, and I have never claimed otherwise.

I am not a pipeline operator who moves data without understanding it.

What I am is fluent in the science, and the author of the engineering. I know why scoring logic is shaped as it is and what norms mean. I know what would break a validated instrument, what validity evidence attaches to, and what a report has to demonstrate. I work at the program level as a peer to the methodologist. I decline individual-level interpretation ("should you hire this person") as belonging to licensed psychologists. That boundary is deliberate. I think it is also why psychologists have trusted me with their instruments for nearly three decades.

The other thing I do is specify. For eight years I have run specification-driven development through an offshore vendor I do not sit with. The specifications have to be complete enough to be implemented correctly at a distance. That is the same discipline AI-assisted development rewards, because a coding agent is an outsourced implementer that needs exactly that kind of specification. My AI-directed operating model, described below, is the continuation of a discipline I already had. It is not a new competency.

One assessment platform, four generations

1998 – 2000 · Batrus Hollweg International, via Campanella, Inc. — Hands-on. I was assigned to design the replacement for the firm's existing assessment system. I researched, designed, and implemented OCR capture of the hand-marked "bubble" answer forms candidates filled out. I was also the project's primary business analyst, database architect, and DBA (SQL Server). This is also where I first learned Visual Basic, which years later became the language of Radius (below).

2002 – 2012 · Batrus Hollweg and 3Peaks, via BrainStorm SDG, Inc. (my own firm) — Hands-on. I built the entire SQL Server data and scoring layer, stored-procedure processing included, for a high-volume assessment platform serving hundreds of client organizations. Billing integrated with Great Plains/Dynamics.

2013 – 2015 · Knox Technology, Inc. — SonicAssess — Specification and architecture. I was data architect from inception for the MongoDB-generation platform that absorbed the Hollweg instruments and assessment services. The implementation layer was ColdFusion and was not mine.

2014 – present · Hollweg Assessment Partners — HAPselect — Specification and architecture. I am the architect of the Azure-generation system, application behavior and data architecture alike. A contracted development vendor implements it to my written specification.

The data and scoring layer was mine in all four generations: hands-on in the first two, specification and architecture in the last two. What I built each time was the part that had to survive.

How the lineage arrived where it is. The instruments' corporate home passed from Batrus Hollweg International to Kenexa and then, with Kenexa, to IBM. The product I carried went a different route. It went from the original firm to an independent development shop: 3Peaks, then Knox Technology, which built SonicAssess. The founder's daughter then formed Hollweg Assessment Partners and brought the founder in as Chief Innovation Officer. With him, SonicAssess rolled in the assessment services and the proprietary instruments. Hollweg Assessment Partners purchased the complete product line. I was the constant across every arrow, consultant and support through each transition, and an employee since 2021.

Hollweg Assessment Partners — HAPselect · 2014 – present

Head of Technology (Senior Technical Consultant). Consultant and subject-matter expert 2014–2020; employee since January 2021. The company's only technical staff member.

HAPselect is a candidate and employee assessment platform: validated instruments, automated scoring, configurable reporting, client integrations. I own system architecture, data architecture, technical specification, and the platform roadmap, and I direct an offshore development vendor and an infrastructure vendor against those specifications. The vendor writes the code. I write what the code must do, and I decide whether it does.

The scoring layer

With the company's psychometricians, I designed and specified the platform's scoring across two paradigms in sequence:

  1. Ratings applied at the score level, working with the founding psychologist.
  2. The conversion of all scoring and calibration to percentile ratings, working with the R&D lead.

I specified the Overall and Potential algorithms. That includes a multi-pass percentile rescoring method of my own design, in which composite percentiles are re-normed against the benchmark, and the logic for interpreting percentiles against rating ranges.

A distinction I hold to: scores are what they are. Scores are measured. Ratings are the interpreted layer. When I describe adjustment mechanisms below, they adjust ratings, never scores.

The reporting layer

I designed the reporting layer as a composable, client-configurable model. Report elements are mixed and matched by trait type and rendering model and stacked sequentially — custom trait labels, custom elements such as problem-solving scenarios, custom components and custom reports, configured per client. Narrative is driven by a prioritized-trait comment mechanism: priority traits are measured against a candidate's scores and the system returns database-configured comments in priority order. Configuration decides both what a report says and how it is built. These mechanisms have been in production for over ten years.

The rating-adjustment layer

I designed a data-driven, composable rating-adjustment mechanism: progressive triggers keyed to percentile ranges and compounding parameters adjust dependent ratings and emit outputs for recruiting decisions. Configured rules — deterministic, auditable, and open to inspection.

The bridge to the science

I worked directly with the psychometric team to define the interface between their statistical analysis platform and production: the Excel-structured export format, the import requirements, and the database import routine.

Client integrations and the API

I have been the chief API architect for client integrations.

Client analytics

When the company's I/O staff were gone, I took on client analytics directly. I extracted score data and shaped it for consumption by Tableau and by AI. I built analytical workbooks. I delivered data interpretation in richly formatted output, tables and charts, and presented it to clients in person. Data interpretation, not psychometric analysis. The instrument stays with the psychologists.

A proof of concept with AI (2023–24)

I built a proof-of-concept AI-assisted consultant-summary system. Prompts assembled the client's job description, the platform report, the client's stated preferences, and the candidate's score data into a draft consultant summary. Every draft was reviewed by the owner before release. It was used on real candidates to cover a consultant's absence, and most drafts went out without change. I extended the framework to executive assessments, with data sheets for the owner's own rules and explanations and samples of her writing for voice and word choice. It was never formalized or adopted. It was a working demonstration that AI drafting under human review could stand in for a consultant on the automated tier.

hapbi — the internal analytics and operations site

Separate from the platform proper, I built hapbi solo: human-directed and AI-assisted from concept to implementation.

Node/Express and Handlebars on Azure App Service. Entra ID authentication and Key Vault. SQL Server with numbered migrations alongside MongoDB Atlas. GitHub Actions with staging-slot swaps to production. One author, 195 commits, roughly nineteen thousand lines of owned source. Development has been paused since May 2026 pending a decision on the platform's historical data.

The modernization decision

When the platform needed re-platforming, the vendor produced one estimate. I did not forward it as it stood.

In my own client-facing proposal I named myself as a project risk: heavy reliance on one person's knowledge.

Now

The automated assessment line is being retired in the autumn of 2026. I am leading data-preservation and retention analysis through the decommission, surfacing what must be kept, for how long, and under what obligations. I am also the named technical contact for clients' integration shutdown and data-export requests.

Refuge House, Inc. and Refuge House San Antonio, Inc. · 2002 – present

Co-Founder · Director of Operational Development. Two legally separate, state-licensed Texas child-placing agencies sharing staff and systems. I co-founded them, and over the years have held technology, operations, quality, security, and board roles. This is not a side project. It is the second half of my working life, and it is where the discipline described on this page runs in a second regulated domain.

Radius

I architected and hand-built Radius, the agencies' case-management and document platform — VB.NET and SQL Server — starting around 2004, and it has been in continuous production since. No other developer has ever touched it. Its architecture:

Its front end is now migrating to Pulse (Node/Express on Azure).

The platform estate

What the agencies run on is not one application but a layered platform with a versioned source of truth at every level:

All of it is implemented by directing AI coding agents against written specifications and acceptance checklists. There is no team and no vendor on this engagement. The specifications are checkable.

AI governance, machine-enforced

The rule that governs AI on this platform: nothing AI-generated becomes operative without a named human approving it. It is not a policy document; it is enforced in the rendering layer. Briefs must ground in evidence, flag gaps rather than fabricate, make no inference fills, 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. Verbatim-only guarantees protect identity fields, staff roles, and document hashes from paraphrase. Every credentialing playbook runs recommendation-only: the AI recommends, a credentialed human decides. And the policy is structured so that thresholds for automatic application could be introduced later, per artifact type, without redesign. I distinguished a policy I chose from an architecture that forecloses the alternative, and I built the reversal seam.

Texas's T3C redesign — regulatory architecture

When Texas overhauled its foster-care service-level and reimbursement system for the first time since 1988, I led the agencies' end-to-end readiness.

State credentialing reviewers approved the documentation. The agencies hold interim-active credentials across the requested service packages, including Mental & Behavioral Health and IDD/Autism.

The centerpiece: after the applications were submitted, the regulator issued a 143-item enhancement request citing a newer edition of the Blueprint. I classified every item against the review form's own history columns, the original application, and the supporting documents already in the regulator's possession. None of the 143 constituted a new requirement. The argument, as I made it: adhering to the most recent Blueprint is the reviewer's obligation, and we accept that. But it does not produce a new requirement our submission has to conform to, because the requirements themselves had not changed. Approval followed. In the next wave I turned that instinct into a governed intake step. Every reviewer ask was triaged before any drafting into transplant, adapt, clarify, or new work. The count of net-new work was zero.

Along the way, three catches:

The training platform

I designed the agencies' training platform and the training policy under it. The design record is the clearest statement of how I work now.

My deliverable is a build brief with an acceptance checklist, not code. It names the repository and branch, the conventions the agent must match, the exact files to create, and checkbox acceptance criteria, including database spot-checks. Governance is by construction, not supervision: lint rules the agent cannot talk past, and a repository instruction file it reads every session. Decisions are ratified and dated inline.

Judgment is encoded as schema constraints:

Nothing multi-tenant ships, but the six seams that are nearly free at design time and expensive to retrofit are built now. I priced the option. The one Refuge-House-specific fork is designed for deletion: a dedicated schema, a flag defaulting off, and every file carrying a literal terminal-fork header so an agent can remove it later. Assessment test-out is competency over hours-in-chair. The exclusions are the judgment: skills that cannot be paper-tested stay off, and a delivery channel mandated externally stays off. A note to a future auditor says the thin folder is by design.

The platform was stood up in production in mid-2026: branded landing, staff authentication, people-matching API, catalog and import tooling, staging-slot pipeline. It joined an existing App Service plan rather than creating a new one; adding an organization's brand is one database row; the first consumer of the data model required no schema changes.

Audit and clinical-automation design

I designed a document-governance and audit playbook layer for the case-management platform. Its finding schema is derived from a co-founder's handwritten home-audit notes. The operational audit sheets are the floor and the handwritten judgment is the ceiling; the two are not merged.

Its governing doctrine, which I titled, is "actionable now, integrity later." The operating test: would this finding trigger follow-up, correction, or clarification, or would it trigger an investigation? The answer gates what surfaces now and what waits. It is a sequencing decision, not a permanent exclusion.

What the record of that work shows:

This layer is in production.

Accreditation

I led technical and operational elements of achieving Council on Accreditation (COA) accreditation, 2010–2011.

Duarte + Duarte LLC, dba Moxie Graphic Productions · 2014 – present

Co-Owner · Operations Principal. A graphic design, print production and vinyl installation firm with two owners and complementary functions. Mine: operations, finance, tax registration and monthly filings in two states, pricing structure and market validation, contracts and policy, permitting, systems, client-facing pricing documents, crew-pay reconciliation, and fleet work. My partner's: creative, fabrication and install, and the customer relationship. Earlier, I was also the creative director — print, signage from channel letters to billboards, apparel, and digital.

The design philosophy

The customer-facing side of the business runs through a partner who is creative, hands-on, and prefers dealing with actual people. The obvious answer was productivity tooling for him. I rejected it outright. What the business needed was leverage for me, not tooling for him. The system's job is to help me help him. Four principles followed:

The litmus test for any future feature: does this make him feel he has a person taking care of things, while making me more capable behind the scenes? If yes, build it. If it adds friction or feels like surveillance, redesign.

Privacy architecture with an explicitly priced tradeoff

Much of the client business runs over text messages, invisible to email tooling. The obvious fix was to point a cloud model at the message database. I refused it, because cloud models read every message they touch, including personal threads. I would rather risk a client interaction than invade my partner's privacy. What I specified instead:

Written down in advance: real business messages will sometimes be missed. That is the price, and I chose it.

Same principle as the Refuge House rendering rules: put the boundary upstream of inference, so the model is never shown what it must not use. Two independent implementations, two domains.

Operations as a system

Pricing is built as a system, not a number. Rates are derived from stated rules. The boundary between standard and specialty work is defined by a stated test. The structure is market-validated and reviewed on a schedule. One model renders for three audiences: an internal locked structure, a client-facing transparency schedule with worked examples, and an internal guide with market comparison.

A "Date Confidence" vocabulary (Proposed, Tentative, Scheduled, Slipped) is defined by the language actually used, with matched customer-facing phrasing and an append-only customer promise log as the accountability spine. I built the data-collection step first and withheld the classifier on purpose until real language existed to train against. I also wrote a "when not to build this" section, with three abort conditions.

A loss on a specialty job became two standing rules and a signed-agreement requirement. The systemic fix was stated as such: the signed-agreement step should not depend on remembering it, so e-signature went onto the roadmap.

Catches that mattered:

The operations app (Next.js, Auth.js with Google OAuth restricted to the workspace domain, Airtable as the data layer, Vercel) is built and running locally. It is not yet deployed. A daily morning digest with auto-reply drafting, reading Airtable directly, is deployed and running.

The Golden Thread · 2026 –

A design and development practice — websites, the systems under them, and the applications on top — built to a published standard, with the method and terms stated before the work starts. I am client zero: my own sites are built with my own protocols so the process is proven before it is sold. It lives at goduarte.com. This page is the record; that site is the practice.

Earlier career · 1995 – 2015

Knox Technology, Inc. · Senior Software Architect & DBA; Client Success Leader · 2013 – 2015. Data architect from inception for SonicAssess (see the four generations above); specification and database architecture across relational and NoSQL stores; retrofitting for vendor integration and upgrade; ETL routines delivering decrypted historical data to clients; direct client and user support for a candidate assessment product; sprint planning and scrum master.

Grant Thornton LLP · Senior Associate — Software Engineer, Technology Auditor · 2006, and 2012 – 2013. Two stints, the second after a decade of independent consulting. Internal-audit and consulting engagements across industries; a security review and analysis tool for Oracle Financials; managed development of an audit-management tool; SSRS and T-SQL analytics; business-process review and documentation.

BrainStorm SDG, Inc. · Owner & Principal Consultant · 2002 – 2012. My own firm. The Batrus Hollweg and 3Peaks assessment-platform work described above; data management, reporting, and BI solutions in .NET, SQL Server, and XML; an asset-tracking database for an oil and gas company; data-structure and reporting review for a social-networking analytics company; practice-management design for a multi-unit physical-therapy company.

DAVACO / estorefixture.com · Information Systems Engineer and DBA · 2000 – 2002. SQL Server design and administration across multiple installations; data design and conversion for the estorefixture.com launch; custom multi-user database applications supporting mission-critical functions, end to end.

Campanella, Inc. · Information Systems Consultant · 1998 – 2000. The founding project of the assessment lineage — the OCR capture system and SQL Server data architecture for Batrus Hollweg International — plus custom database applications across industries, with extensive design and analysis documentation, as primary client representative.

AMRESCO, Inc. · Senior Business Analyst/Programmer · 1996 – 1998. A PeopleSoft HRMS implementation — interfaces to the JD Edwards general ledger and external HR and benefits vendors, an intranet time-and-attendance application, payroll and check-distribution procedures, data work against DB2/AS400 — and application development for the company's business lines: specifications and code for financial-calculation, loan-amortization, cash-flow-projection, and due-diligence applications.

FoxMeyer/McKesson Pharmaceuticals · Database Coordinator · 1996. Administered large Access and Oracle databases of sales and pharmaceutical data; built the applications reporting monthly financial and sales data and collecting vendor and customer rebates; month-end procedures across mainframe and PC.

New Horizons Computer Learning Center · Computer Applications Instructor · 1995 – 1996. Trained classes of five to twenty-five adults, on site at corporate locations; Instructor of the Month.

How I work

Specification first. My deliverable is a specification with an acceptance checklist. Whether the implementer is an offshore vendor or an AI coding agent, the discipline is the same: write it completely enough to be built correctly by someone who is not in the room, then verify against the checklist. Nothing is described as done until it is checked.

AI-directed development, with a written routing rule. Explore in Claude Design; construct in Claude Code; refine in Cursor; decide, operate, and work data in Claude Cowork. That started as a rule of thumb and became a ratified row in a decisions table. Governance is by construction — rules the agent cannot talk past, and an instruction file it reads every session — not by watching over its shoulder.

The boundary goes upstream of inference. If a model must not use something, the model is never shown it. Redact before the model. Classify before identifying. Put the audience ceiling on the data source. This shows up independently in the child-welfare platform and Moxie's message triage, because it is a principle, not a feature.

Formal identity where the population is administered; gated challenge everywhere else. Staff get directory accounts. Candidates, foster parents, youth, and outside professionals get scoped, revocable, time-bound challenges: a link, a token. The reasons are concrete. A candidate takes one assessment. A foster couple shares an email address. A youth's inbox may be monitored or shared. Implemented four times across two engagements.

Status is the achievement. Everything I build is tagged designed, shipped, deployed, adopted, or abandoned, and I do not present unimplemented work in the present tense. A proof of concept is a proof of concept; a paused project is paused. The distinction matters more than the accomplishment.

Sequence honestly. "Actionable now, integrity later" is a doctrine I wrote for deciding what an audit system surfaces first. It is a sequencing decision, stated as one, so that what waits is not mistaken for what is excluded.

Kill sunk cost. A compliance tracker went from HTML to CSV-backed to a full web app before I killed all of it and wrote the transfer prompt naming the failure mode. A misdiagnosis stays in the record alongside why it was wrong.

Name the risk when the risk is me. In my own client-facing modernization proposal, the first risk listed was reliance on one person's knowledge. The proposal was mine.

Skills and stack

Data. SQL Server — design, T-SQL, stored procedures, migrations, schema as versioned source; MongoDB and Atlas; historically MySQL, Oracle, DB2. ETL and BI with HEVO, Tableau, Power BI, SSRS.

Platform. Azure App Service and Azure SQL; Entra ID and MSAL, including passkeys; Key Vault; GitHub Actions with staging-slot swaps; Vercel; Docker; Azure Virtual Desktop.

Application. Node/Express, VB.NET/WinForms (DevExpress, SyncFusion), React and TypeScript, Next.js, Handlebars; token, magic-link, and OIDC authorization patterns; API design; Model Context Protocol server design; Airtable as an application data layer; JIRA administration.

AI-directed development. Claude Code, Cursor, Claude Cowork, Claude Design, under a written tool-routing and human-review policy; specification-first, acceptance-checklist delivery.

Domain. Selection-assessment scoring, normative, and reporting architecture; applicant-tracking and HRIS integration; UGESP data and reporting; AI on regulated data — redaction upstream of inference, verbatim-field guarantees, human approval gates; Texas DFPS child-placing regulation and the T3C Blueprint; HIPAA-adjacent authoring boundaries.

Design and production. Adobe Creative Suite for print and digital; brand development; signage and vehicle-wrap design, including true-scale correction for raked surfaces.

Education

Dallas Baptist University — Bachelor of Arts and Sciences, a multidisciplinary degree in Computer Information Systems and Business Administration.

Brookhaven College — Certificate in Nonprofit Management.

Externally validated

Items an outside party has examined:

Last updated September 2026.