Project Portfolio
AI platform engineering and applied people analytics: governed decision infrastructure, agent orchestration, and machine learning, alongside the Total Rewards and HRIS work it grew out of. Headline outcomes are documented results; supporting chart values are representative illustrations for portfolio display, and platform diagrams describe systems as built.
DecisionLedger AI, by the build
Measured business outcomes live on the Outcomes page. The figures here describe what was built, not what it earned.
Predictive models and generative AI for HR, alongside the analysis and reporting that puts their output in front of the people deciding. Built on WatsonX Orchestrate, OpenAI, Python NLP pipelines, and Power BI.
Self-service Power BI reporting on headcount, turnover, and time-to-fill, giving leaders a single governed view of workforce health.
A conversational assistant trained on benefits plans, policies, and leave rules, answering employee questions instantly and deflecting routine tickets from the HR team.
A supervised model scoring departure risk from engagement, tenure, and performance signals, surfacing at-risk talent early for targeted retention.
An NLP pipeline classifying open-text survey feedback by theme and sentiment, turning thousands of comments into a leadership-ready signal.
Analyzes benefits utilization trends to model cost-effective plan designs, pinpointing where spend and employee value diverge.
High-cost, low-utilization benefits (upper left) are flagged as redesign candidates.
Drafts role-ready job descriptions and screens the language for bias, flagging exclusionary or gendered phrasing and rewriting it inclusively in seconds.
A conversational assistant that walks new hires through paperwork, provisioning, and training, auto-completing routine steps and nudging the rest.
Maps each employee's skills against role targets and career goals, then recommends the next training modules to close the gap.
Recommended next: Advanced SQL, Leading Teams, Power BI DAX
Continuously scans HRIS records for missing fields, duplicates, and inconsistencies, scoring data health and routing exceptions for cleanup before they reach payroll or compliance.
Slow hiring was being read as a recruiter throughput problem. Instrumenting the funnel stage by stage put the delay somewhere else entirely: requisition approval, upstream of anyone sourcing a candidate. Rebuilding that one step cut time-to-post 35% and accelerated pipeline starts on critical roles.
Three divisions ran three separate performance review processes, so nothing aggregated and nothing compared. Unified them onto a single Culture Amp cycle, then ran the engagement pulse across the whole organization and took the findings to the Board, the executive team, and employees, delivered through Power BI and the intranet rather than a slide read once and filed. An Employee Voice Committee was launched off the results.
Modules of a single governed decision platform, architected and built solo: application, infrastructure, supply chain, and nine client surfaces. Each one ships inside DecisionLedger AI.
A full learning platform that builds a course in minutes from a prompt or from a decision that already happened. When an outcome is recorded against its prediction, the platform drafts the training the organization needs from that gap. Every claim in a course cites the decision that produced it, approval is blocked while any claim is uncited, and each course clears its own approval gates before a single learner is enrolled. The authoring surface is 41 content block types including 11 assessment types with drag-and-drop, hotspot, matching, and ranking. Courses package to SCORM 1.2 and 2004 and emit xAPI into a built-in learning record store, so they run inside an existing corporate LMS instead of replacing it. Ships with a player, gradebook, certificate designer, translations, folders, invites, share links, and revision control.
Browser-based, Excel-style workbooks with a real formula engine (cell and range references, function library, and the standard error set) plus Excel import and export. The governance move is the binding: any cell can be registered as the source of a tracked KPI, so a number living in a spreadsheet becomes a versioned metric with thresholds, alerts, forecast history, and an audit trail instead of an untracked file on someone's desktop.
A visual builder for multi-agent workflows (system prompts, tool access, model selection, execution triggers) sitting on a runtime governance layer. Every agent is registered, policy gates run before execution rather than after, multiple agents vote independently on high-stakes calls, high-risk output is quarantined until an admin releases it, and a kill switch halts any agent instantly. External agents reach the platform through an MCP gateway with per-agent budgets and scoped permissions.
Every AI call routes through one metered gateway that applies savings automatically: identical requests are served from cache, a lower-cost model is tried first and escalated only when the work demands it, and non-interactive jobs run in batch. Telemetry breaks tokens, latency, and cost down by provider, model, agent, user, and project. A reconciliation pass compares what providers billed against what was actually governed to surface shadow AI spend, and project cost centers combine human labor cost with AI spend in a single budget.
The platform runs on infrastructure defined as code, not clicked together in a console. Terraform provisions ECS, RDS, Redis, CloudFront, WAF, CloudTrail, and Cognito, along with the pieces that make the governance claims real: an S3 Object Lock bucket for audit records that cannot be rewritten, and KMS keys backing decision attestation. Deploys authenticate through GitHub OIDC, so no long-lived AWS credentials exist to leak. Four CI workflows guard the path: pull request validation, a branch guard, a gated database migration job, and deploy, with separate staging and production build pipelines for the API and the web tier.
Every decision model ships as a versioned package carrying an Ed25519 signature, a content hash of the package, a hash of each individual file inside it, and a provenance record naming the builder, build host, environment, and timestamp. Verification runs at load, before execution, so a model that was altered after approval does not run at all rather than running and being caught in review. This is the answer to the question every AI governance conversation eventually reaches: how do you know the model that ran is the model that was approved.
The objection HR, Legal, and Compliance raise about AI assistants is not accuracy, it is exposure. This one classifies a conversation by regulatory domain from the first message and binds it to that domain's rules for the rest of its life: legal privilege with SHA-256 attestation and work-product watermarking, healthcare with PHI awareness and six-year retention, financial with a SOX trail and insider-list tracking, compliance with an evidence chain tied to a named framework. Retention, redaction, watermarking, and audit events attach at classification rather than at export. Behind it sits full administrative oversight: search across every conversation in the organization, legal hold that overrides automatic deletion, and structured eDiscovery export.
The layer that makes the platform legible to a board rather than to an engineer. Committees are created and staffed in the system, motions are voted with approve, block, abstain, and conditional outcomes, and the resulting resolution is recorded as a governed decision rather than an attachment to a PDF. Minutes are generated from the session record itself, so what the committee decided and what the minutes say cannot drift apart. Alongside it: a board dashboard with intelligence, document, and compliance views, a resolution register, director independence tracking, and ESG metrics.
A governed model is only useful where the decision actually gets made, which is rarely inside the vendor's web app. So the platform ships nine ways in, each built and versioned in-house: a 364-page web application, a React Native mobile app, a Chrome extension shipped through eleven releases, an Excel add-in that exposes models as worksheet functions, a Google Sheets add-on, a Python SDK with sync and async clients, two MCP servers including a packaged desktop bundle, a command line client, and a documentation site. Every one of them enters through the same gates, budgets, and audit trail, so the governance does not weaken as the entry point gets more convenient.
Enterprise data architecture and AI governance underpinning trustworthy, audit-safe analytics.
Leadership reporting was spreadsheets: exported by hand, reconciled by hand, and out of date by the time anyone read them. Worse, each division carried its own workbook, so headcount meant one thing in Finance and another in HR and no one could say which was right. This replaced that with a Snowflake warehouse and a Power BI semantic layer, so a metric is defined once and every division reads the same number. It became the foundation the Microsoft Fabric medallion architecture was later built on, and together they took leadership insight from weeks to minutes across six divisions.
Built a full medallion lakehouse from the ground up, piping every division's systems (HRIS, payroll, benefits, ERP, EHR) through Bronze, Silver, and Gold layers into certified semantic models, executive Power BI, and AI orchestration, with a patent-pending hallucination-prevention layer keeping insights audit-safe.
An orchestrator agent plans each leadership question, routes it to specialist agents that query the governed Gold layer through MCP tools, then synthesizes a grounded, cited briefing behind a hallucination-prevention guardrail.
Authored and rolled out an enterprise data governance policy that closed every critical gap identified in a formal audit, establishing controls across access, lineage, and model oversight.
Ran a Data Protection Impact Assessment across 27 processing activities, scoring privacy risk by likelihood and impact, then drove mitigations that closed 18 gaps and eliminated every high residual-risk activity.
Compensation, benefits, HCM, and talent programs designed, transitioned, and measured across multi-subsidiary workforces.
Led the carrier transition from Kaiser to Anthem, restructuring plan design to cut employer spend while lowering employee premiums.
Redesigned 401(k) features with education campaigns and auto-enrollment to lift participation across the workforce.
Directed a six-month platform migration for 1,600 employees across discovery, configuration, data migration, parallel testing, and go-live, landing on schedule with automated payroll and benefits integrations.
Benchmarked pay across 129 roles against market data and built grade-based salary ranges, then ran adjusted pay-equity analysis to keep gaps within tolerance across the workforce.
Stood up a Talent Review Committee and facilitated 9-box calibration across the workforce, mapping performance against potential to guide succession and development.