Systems Architect → AI Platform Architect / MLOps
Lewisville, TX · bdebnam@gmail.com · linkedin.com/in/bryanhd
Resume —request a copy · PDF or DOCX,
AI Platform / MLOps or Systems Architect.
Positioning
20+ years in enterprise IT, 3+ at the architecture level — cloud migrations, M&A technology
integration, and identity modernization across four acquisitions, plus building and leading a
60+-control SOX/ITGC compliance program.
At work, more recently: “Script Portal,” an internal platform for interchangeable
agentic-AI backends under one allowlist + output-masking safety model; Playwright RPA agents
for SOX audit evidence; the org’s GitHub Enterprise Copilot rollout.
Independently (personal R&D; foundation: Texas McCombs AI/ML post-grad, 2024–2025): a full
MLOps governance lifecycle (model registry, champion/challenger promotion with rollback,
automated drift detection, data lineage), an LLM security triage tool, and LLM cost analytics —
all on a self-hosted, dual-GPU Incus/k3s fleet.
What’s distinctive: the intersection of SOX/ITGC audit discipline and hands-on AI-platform
engineering — few people have both.
Selected projects
Each project is tagged [Work] (current employer) or [Independent R&D] (personal), and
scoped: Pilot (in limited use on real workloads, ahead of formal rollout),
Reference build (complete and running, single-environment / not scale-tested), POC
(proves the pattern).
Model registry & staged promotion (MLflow) — preflight validation gates, automated
rollback, and a silent-no-op safety guard that blocks unvetted model changes from reaching
production.
Champion/challenger A/B gate — a staged candidate model is evaluated head-to-head against
the live champion before any cutover.
Automated drift detection — weekly scheduled scanning, parameter-order-insensitive.
Caught a live defect in production: an arithmetic self-contradiction in a prompt-scoring agent
that was inflating its own output score.
Data lineage (Marquez / OpenLineage) across the MLflow ↔ Ollama pipeline — every model
version, eval run, and promotion decision is auditable.
LLM-as-judge evaluation harness — scores agents on substance (accuracy, completeness, no
fabrication) rather than surface grammar; registered judge model; testsets logged as MLflow
GenAI datasets. Brought all 7 agents to a validated quality gate.
AI Gateway (FastAPI/uvicorn) — unified model-access boundary with a local detect-quality
model; GenAI UI wired (Prompts registry for 7 agent system prompts, review queues, label
schemas).
Model-version cost/quality comparison — compare cost and quality across registered
model versions to inform promotion decisions.
2. Provider-Agnostic Agentic AI Platform (“Script Portal” / Agent Bridge) — [Work] · Pilot — daily driver
An internal accelerator — teams get interchangeable agentic-AI backends that can act against
real enterprise systems under one allowlist and masking model, instead of a per-tool approval
for every new agent.
Interchangeable agent backends — Claude Code CLI, Copilot GPT CLI, and a custom VS Code
“Agent Bridge” extension (LLM-broker) — behind a shared allowlist and output-masking layer.
Provider-independent agent contract — spans tool use, memory, reasoning workflows, and
decision frameworks, so the underlying model or vendor can be swapped without changing how
agents are invoked or governed.
Secret-aware execution framework with output masking applied to agent transcripts, so
agents can act against authenticated internal systems under one safety model rather than
per-tool exceptions.
Piloted on identity-lifecycle workflows — account-hygiene checks, expiring-credential
monitoring, and tickets whose summary and detail are model-generated rather than
templated, with credential renewal handled inside the same governed boundary.
Closed loop, not just detection — privileged actions run inside the allowlist and
masking boundary rather than as scripted exceptions around it, so the framework extends to
new use cases under the same governance model.
Scope: single-operator pilot. Broader rollout is designed around a dedicated service identity
with least-privilege delegation.
Two cooperating Playwright RPA agents run the audit-evidence cycle end to end — the first
interprets incoming evidence requests and resolves what each control activity requires; the
second collects the required records from enterprise systems and assembles timestamped,
masked evidence packages filed under the correct control.
Auditor-ready output — evidence arrives already scoped to the control activity rather than
as raw exports someone has to reconcile by hand.
Built and led the SOX/ITGC compliance program — 60+ control activities, PowerShell-
automated evidence generation; primary liaison between Infrastructure, Security, and Audit.
Manage the org’s GitHub Enterprise Copilot deployment; built an n8n workflow monitoring
seat utilization — reclaimed 25 inactive seats (~$475–725/month).
Led M&A technology integration for four acquisitions — M365 tenant consolidation, AD/Entra
ID identity integration for ~1,800 users, 400+ mailbox/OneDrive migrations, multi-forest AD
trust across 6 forests.
A personally-owned 9-account AWS Organization built as a governance reference
architecture — the multi-account control plane, not a workload environment.
Account topology separated by blast radius rather than by team: Security (CloudTrail
administration, Config aggregation), Shared Services (Central Backup, Backup Administrator,
Network), and Workloads.
OU hierarchy — Security, Shared Services (nested Network OU), and a Workloads OU
subdivided into Dev, Test, QA, Prod, Prod-Sim, ML, and IoT, so service control policies
scope per environment tier.
Federated identity — IAM Identity Center fronting all member accounts, federated to an
external Entra ID tenant; SSO device-code flow with per-account permission sets, so no
account carries long-lived IAM access keys.
Centralized security telemetry — dedicated CloudTrail administration and Config
aggregator accounts, keeping audit evidence where workload OUs cannot write to it.
Backup separation of duties — distinct Central Backup and Backup Administrator accounts,
isolating backup data from the identities that administer it.
Scope: control-plane and governance build. The workload OUs are provisioned but intentionally
unpopulated — the artifact is the account structure, delegation model, and identity federation,
not running services.
An Azure-hosted internal portal that turns scattered vendor licensing data into a single
view of software spend and seat waste.
Multi-source aggregation — pulls licensing and usage data from vendor REST APIs and
normalizes seat models that differ from vendor to vendor into one comparable schema.
Waste detection — surfaces duplicate license assignments, unused seats, and per-vendor
spend, producing an actionable reclaim list rather than a static dashboard.
Shared component library — reuses helpers from the agent framework (project 2): Active
Directory and Graph email helpers, plus the duplicate-account detection logic generalized into
duplicate license assignment detection — the same instinct applied to identity and to spend.
Delivery — built through agentic AI-assisted development (Claude Code, GitHub Copilot)
and shipped via Azure DevOps build and release pipelines with Azure service connections.
The CI/CD, security-scanning and recovery layer behind project 3: every change to the platform
repo goes through a multi-scanner security gate, and every finding lands in one tracked place.
Security gate (Azure DevOps, self-hosted agent) — five scanners, hard-gating on the
serious tiers: gitleaks (secrets, full history), checkov + trivy config (IaC),
semgrep (SAST), trivy (container-image CVEs), plus a custom “no secrets on command
lines” check (below). SonarQube runs on every main build for code quality and tech-debt
tracking — informational, not a gate; semgrep is the SAST gate.
One findings dashboard — all scanners feed DefectDojo as one product: SARIF reimports
into long-lived tests with automatic closure of fixed findings, a stale-run guard so an older
build can’t reopen what a newer one fixed, and SonarQube pulled in through its API import.
Case study — credentials in process arguments. Found Vault AppRole secrets, tokens and
database passwords being passed to curl as command-line arguments (readable by any process
via /proc/<pid>/cmdline) across 28 scripts and pipelines. Fixed all of them (stdin and
process substitution instead of argv; CI secret variables mapped through env: instead of
inlined), verified against live Vault, and turned the rule into a CI gate so it can’t
regress.
Pipeline credential model — HashiCorp Vault AppRole per pipeline with single-path,
read-only policies; secret IDs bound to the agent’s address; deploy access through
forced-command SSH keys (one key can run exactly one allow-listed installer script).
App deploy pipelines — deploy/teardown for an EHR demo stack, FHIR/DICOM services, a ledger
API service and sports-analytics apps, each logging in to Vault with its own narrow role.
Backup & recovery — nightly fleet backups via ZFS snapshot + incremental send with a
bounded window and a carryover queue; daily logical dumps of every PostgreSQL and MariaDB
database (read-only dump accounts); everything age-encrypted to a hardware-key recipient,
shipped off-box to a NAS and checksum-verified; a status dashboard over the job history.
Microsoft(active through 2027): Azure Solutions Architect Expert · Azure
Administrator Associate · Azure Fundamentals (2019) · AI-102 Azure AI Engineer
Associate — in progress
Post Graduate Program in AI & Machine Learning: Business Applications — McCombs School
of Business, The University of Texas at Austin (conferred July 2025)
A.A.S., Information Technology — CIT: Network Administration & Support — Dallas County
Community College District (now Dallas College)
U.S. patent (2014 Microsoft hackathon project, placed 25th worldwide)
Honest scope statement
Two separate evidence bases, never blurred:
Independent R&D is personal work, not a production system with a user base — it
demonstrates architecture judgment, integration breadth, and knowledge of failure modes. The
sandbox is a single-host reference platform (projects 1, 3, 4); the AWS landing zone is a
control-plane build with intentionally unpopulated workload OUs (project 6).
“At scale / in production” evidence comes from the work at my current employer: four
M&A integrations, identity modernization for ~1,800 users, a 60+-control SOX program, and an
org-wide Copilot rollout. The agent-platform projects built on top of that experience —
Script Portal, the SOX evidence agents, and the spend portal — are pilots ahead of formal
rollout. Projects 2, 5, 7.
Verification
Issuer-hosted records, verifiable without an account:
Microsoft Certified: Azure Solutions Architect Expert — verify on Microsoft
Learn
· earned Nov 18, 2019, expires Nov 19, 2027 · certification number ABA961-E2X9C0
Microsoft Certified: Azure Administrator Associate — verify on Microsoft
Learn
· earned Apr 24, 2019, expires Oct 25, 2027 · certification number 628584-CT21D1
AWS certifications — Credly badge
wall · Solutions Architect – Associate
(earned Jan 4, 2026, expires Jan 4, 2029), AI Practitioner (earned Nov 2, 2025, expires Nov 2,
2028), Cloud Practitioner (earned Aug 15, 2024, expires Jan 4, 2029)
Post Graduate Program in AI & Machine Learning: Business Applications, McCombs School of
Business, UT Austin — verify · conferred July
2025, 9.5 CEUs