Anthony A. Robinson
Summary
I get dropped into somebody else's undocumented system, work out how it actually behaves, build the thing that fixes it, and leave the team able to run it without me. Twenty years of that inside Microsoft and enterprise IT — reverse-engineering deployment systems, writing the tooling, then writing the training plan for whoever inherits it. For the last three years, the same work with LLMs in the loop: production Anthropic and OpenAI API pipelines, multi-model routing, evaluation and cost discipline, and hard-won judgment about which parts of a system a model should own and which parts it must never touch.
What I Bring to a Forward-Deployed Role
Land in an unfamiliar system and ship
Inherited a set of Azure deployments at SanMar that nobody could explain — no IaC, no documentation, pure tribal knowledge. Reverse-engineered them into a custom PowerShell/Azure CLI deployment toolset, then produced the development plan, test plan, and training plan so the team could own it after I moved on. Later extended it to PowerShell 7 and Linux so the same toolset ran on Linux build agents and developer desktops.
Put LLMs into production — and know when to take them back out
Built an AI SRE service (FastAPI + Grafana Loki + Claude) that reads logs across five hosts and reports real
incidents. Version 1 kept an LLM-maintained issue ledger; the same read-modify-write defect recurred four times in
two weeks. I root-caused it as an architecture problem rather than a prompt problem and redesigned v2 so
deterministic code owns detection and state — declarative detectors into a SQLite findings table, severity as an
outcome lookup, issue retirement as a WHERE clause — with the model permitted only to narrate a diff
the code had already computed. P1 alerting became a deterministic cron of nine checks with stamp dedup and no model
in the path at all.
Make the context budget do real work
The same service's daily reports were useless because a single 2,000-line query budget was being consumed by public bot-scan 404 floods from one host. Rebuilt it as a per-host balanced budget with a server-side error filter that drops 4xx scan noise, and the reports started leading with genuine P1s (a JWT crash-loop) instead of scanner noise. Routine practice: route high-volume classification to Haiku and deep analysis to Sonnet, and forecast token burn before batch runs.
Enterprise integration and the security conversation that comes with it
Migrated public ingress for 15 vhosts to a Cloudflare Tunnel — zero inbound ports, zero A records, all 16 certificates moved to DNS-01 issuance. Deployed fail2ban with a Cloudflare-edge ban action so attackers are stopped before they reach the network, and found during that work an API token scoped with Account API Tokens: Write — a credential sitting on a host that could mint arbitrary tokens and compromise the entire account. Removed it, verified the policy fail2ban actually depends on was untouched, and documented the failure mode.
Leave behind documentation that does not rot
Built and maintain a hub-and-spoke infrastructure documentation repo: 20 host blueprints written to a hard-death-rebuild standard, 9 service documents, 9 operational runbooks, and 9 architecture decision records. Collector scripts emit deterministic generated output — no timestamps or PIDs, credentials masked, secret material fails loudly — so the same run on Windows and on the hypervisor is byte-identical, and a weekly cron commits the drift. The standing rule is that no infrastructure change lands without the matching document in the same sitting.
Translate between rooms
Chose the toolchain and trained the dev and test organization on it for a new Azure-hosted product at Garden City Group. Owned the OEM-to-product-group technical relationship at Microsoft under the Compatibility Labs program. Build the reporting executives actually read — including AI-generated weekly narrative summaries assembled from repository commit and PR activity.
Experience
Senior DevOps & AI Systems Engineer · SanMar
Deployment and release engineering for enterprise SAP Commerce, WMS, and Dynamics 365 systems, alongside building the AI-augmented tooling the team runs on.
AI systems in production
- Built an end-to-end meeting intelligence pipeline: GPU-accelerated Whisper transcription → Claude API classification across a 13-category taxonomy → structured summarization → action item extraction reconciled against Azure DevOps work items. Dual-model architecture with retry/backoff and state persistence.
- Built AI-powered weekly activity reporting that scans ADO repositories for commits and PRs and generates executive narrative summaries via the Claude API, replacing hours of manual status writing.
- Authored production prompt specifications (230+ lines) with anti-drift safeguards, explicit escalation tiers, output-format contracts, and meeting-type-adaptive structure — plus boundary-detection rules for transcripts containing multiple meetings.
- Built an ADO API resilience framework: progressive retry-backoff (5 attempts, 2–10s), per-item isolation to stop cascade failures, and pipeline-native error logging.
- Built a PowerShell security audit framework covering 6 risk categories — credential exposure, exfiltration, dynamic code execution, Base64 obfuscation, temp-file races, and network MITM.
- Maintain a 233-function composable PowerShell library used as a reusable context layer for AI-assisted development, with a search-before-create protocol enforced through agent instruction files, and blast-radius-aware tool permissioning with human approval gates on deployment operations.
Deployment and infrastructure
- Built the "What's Deployed Where" system: real-time deployment state across every environment, with an automated dashboard mailer and CAB weekly emails driven by nested ADO work item queries. Became the reference both engineering and management use.
- Reverse-engineered undocumented Azure deployments into a custom IaC toolset (PowerShell/Azure CLI) for both desktop and pipeline use, shipped with a development plan, test plan, and training plan; later extended to PowerShell 7 and Linux.
- Redesigned serial ADO release pipelines for parallel execution — 48% deployment time reduction. Re-architected Azure deployments — 97% deployment time reduction.
- Designed and built the custom tooling around
jira-azuredevops-migratorfor an enterprise-wide Jira-to-Azure DevOps migration. - Automated environment and asset data retrieval from Microsoft Dynamics Lifecycle Services; built the Prod Date Calendar and D365/LCS environment tracking systems; parallelized Linux repository updates; automated orphaned release-branch cleanup.
- Stood up a local ADO-based PowerShell module repository as a DR path for when PowerShellGallery is unavailable.
DevOps Engineer · Microsoft
- Designed, monitored, and triaged CI/CD pipelines for the Windows Photos App; analyzed workflows and implemented performance improvements.
- Designed an automated test-failed-retry pipeline that separated flaky infrastructure from genuine test failures, so developers stopped triaging noise.
- Implemented a Docker-based ELK stack collecting build, release, and test telemetry from ADO for bottleneck analysis.
- Built ADO activity dashboards and reporting on PowerShell and the ADO REST API.
DevOps Lead · Garden City Group
- Researched, selected, and implemented the Git-based development and release toolchain for a new Azure-hosted product, then trained the existing dev and test organization on ADO, Azure hosting, Git, GitFlow, and NuGet packaging.
- Built CI/CD processes, pull request policies, and pipelines; implemented dotCover unit/component coverage and Karma/Istanbul UI coverage with Elasticsearch and Kibana coverage dashboards.
- Improved the legacy TFS2013 enterprise application's builds and branch maintenance; wrote and implemented feature-branch and automated-merge procedures for bi-weekly deployments.
- Managed flow of work across Epics, Stories, and Work Items in ADO.
Senior Build Engineer · Intentional Software
- Built auto-deployment scripts that self-registered OpenStack/BlueBox cloud VMs with Jenkins on provision, in a mixed Jenkins 1.6/2.0, SVN/Git, and TFS environment.
Senior Build Engineer · FTI Consulting
- Managed and administered the division's Visual Studio TFS 2015 instance and product builds.
- Automated code coverage metric calculation and publishing using JetBrains dotCover and SQLCC.
- Co-designed a division-wide dashboard reporting build status, infrastructure, and coverage metrics on Elasticsearch and Kibana.
- Automated build monitoring, queuing, and team alerting — including SMS on broken or overdue builds — via PowerShell, MSBuild, and the TFS API.
Build Lead / Build Manager · Microsoft
- Ten years running enterprise-scale build and release infrastructure across Windows Media, Windows Mobile, Windows Live and Live Search, Developer Division (Visual Studio), and console/services.
- Hired, grew, and managed a team of 5 build engineers (3 FTE, 2 contract) delivering 24/7 build and source management support.
- Redesigned the manual Service Pack creation process into a multi-language automated pipeline, scaling from 4 to 26 languages.
- Managed 100+ Windows Mobile product builds; built an isolated build network that cut build times and corporate network load; delivered a web-based automated build status site.
- Built automated VM provisioning on Windows Server 2012R2 Hyper-V, a local NuGet Gallery, and TFS Build Controller/Agent environments for customer builds.
- Implemented the Compatibility Labs program — later "Windows Certified" — testing OEM hardware compatibility against Microsoft products, and owned technical issues between OEMs and Microsoft product groups.
Systems Built and Operated
homelab-ai-sre — FastAPI service pairing Claude with Grafana Loki for cross-host log analysis over Docker, Proxmox, OPNsense, and nginx. Ships as a Docker Compose stack (Loki + Promtail + API). Ingests from five hosts, including journald-only Proxmox and RFC5424 syslog from the router. v2 moved detection and state into deterministic code with the model confined to narration. Detail →
homelab-traffic-report — Dependency-free Python that parses nginx JSON access logs, classifies hostile versus legitimate traffic, enriches source IPs with country and ASN from an offline dataset, and renders a self-contained HTML report. A 21-day window showed ~76% of external traffic hostile and 64% originating from cloud providers. Its exposure check — probes for secrets and config that returned 2xx — found a real leak, a docker-compose.yml being served publicly, which I fixed at the reverse proxy across every public vhost.
TLDW (Too Long; Didn't Watch) — AI content intelligence pipeline tracking 51 YouTube channels and 783+ videos with automated transcript retrieval, summarization, and analysis. Runs daily. Detail →
Infrastructure documentation system — Hub-and-spoke repo that serves as single point of truth for a 17-host environment: 20 host blueprints, 9 service docs, 9 runbooks, 9 architecture decision records, plus collector scripts producing deterministic, credential-masked generated output and a weekly drift-commit cron. Detail →
Kubernetes GitOps cluster — Talos Linux on Proxmox with two-loop reconciliation: OpenTofu for infrastructure, Flux CD for applications. MetalLB, Traefik, NFS storage, SOPS. Migration scope deliberately limited to stateless services, with a written finding on exactly what blocks each remaining candidate.
Self-hosted platform — 50+ Docker Compose service stacks including AnythingLLM, n8n workflow automation with scheduled AI integrations, Immich, and Prometheus/Grafana/Loki observability. Ansible playbooks for patching, reboot orchestration, and hardware inventory across the fleet.
Technical Skills
| LLM & AI Engineering | Anthropic Claude API (Haiku/Sonnet/Opus), OpenAI API, Claude Code, MCP, agent workflows, prompt specification, evaluation harnesses, multi-model routing, token economics, Whisper / faster-whisper (CUDA) |
|---|---|
| Languages | Python, PowerShell / pwsh, Bash, YAML, JSON, SQL |
| Cloud & Platform | Azure, Azure DevOps, Azure CLI, Azure Key Vault, Cloudflare (Tunnel, DNS, API), Dynamics 365 / LCS |
| Containers & Orchestration | Docker, Docker Compose, Kubernetes, Flux CD, Helm, Talos Linux, Proxmox, Hyper-V |
| IaC & Configuration | Terraform / OpenTofu, Ansible, SOPS |
| CI/CD | Azure DevOps Pipelines, Jenkins, TFS, GitHub, Git / GitFlow, NuGet |
| Observability | Grafana, Loki, Promtail, Prometheus, Elasticsearch, Kibana, ELK |
| Security & Networking | nginx reverse proxy, fail2ban, TLS / DNS-01 automation, OPNsense, Tailscale, secrets management |
| Enterprise | SAP Commerce / Hybris, WMS, release and change management, CAB process |