AI Platform & Cloud Infrastructure Architect · 15+ years in cloud architecture · 2x founder before architect
I design and ship AI-enabled cloud platforms — agentic infrastructure, Model Context Protocol (MCP) servers, context/graph infrastructure, and internal developer platforms — for safety-critical and regulated industries. My research asks what it takes for AI agents to operate production infrastructure reliably: what the tool-call plane actually delivers, when autonomous remediation loops go unstable, and how much authority an agent should hold.
Every repository here is runnable, tested and tied to a paper, a book chapter or a long-form post.
→ cloudandsre.com · LinkedIn · YouTube → ORCID 0009-0007-6190-9516 · Google Scholar · Semantic Scholar · ResearchGate
Code and data behind my papers. Each reproduces its paper's results from a clean checkout.
| Repository | Companion to | What it shows |
|---|---|---|
mcp-schema-drift |
Pinning the Version Is Not Pinning the Tools: Deployment-Induced Schema Loss in MCP Servers | Of 116 working release-pairs of the official MCP servers, 18 advertise different tool schemas depending on dependency resolution — 17 with every schema emptied. Ships mcp-surface, a CLI and GitHub Action that catches it |
remediation-stability-sim |
Stable by Design: A Control-Theoretic Account of AI-Driven Self-Healing Remediation Loops | Dead-time alone flips a correct-by-construction remediation loop into oscillation; the settle/diverge frontier follows a delay–gain product law. DOI 10.5281/zenodo.22786822 |
autonomy-budget |
Error Budgets for Autonomy: SLO-Driven Authority Management for Autonomous AI Operators | Error-budget semantics for agent authority: a blast-radius-weighted budget and a fast-down/slow-up authority ladder |
mcp-chaos |
The Tool-Call Plane: An Operational Reliability Model for MCP-Based Agent Infrastructure | Fault injection and reliability SLIs for agent tool use |
| Repository | What it is |
|---|---|
catalog-context-graph |
Backstage software catalog → operational context graph for agents over MCP: blast radius, ownership, change risk, catalog drift. Neo4j optional. Measured against flat retrieval |
mcp-task-reliability |
The reliability layer the MCP Tasks extension leaves unspecified: durable store, retry classes, orphan reaping, adaptive polling |
agent-spend-guard |
Runtime spend circuit breaker for agent trajectories, with OpenTelemetry GenAI cost attribution |
agent-incident-corpus |
Public AI-agent production failures as a replayable regression corpus with a CI gate |
inference-slo |
SLO-aware admission control for LLM inference serving — what to shed, when, against which SLO |
sre-harness |
Minimal reference harness for SRE agents: tools, verification, memory, guardrails, observability |
context-assembler |
Context engineering reference: typed segments, enforced token budget, reranking, compaction |
cloudandsre-skills |
Agent Skills for SRE work with output contracts, guardrails-as-code, golden evals and an MCP server |
| Repository | What it is |
|---|---|
alert-explainer |
Alertmanager → AI triage brief service: dedup, enrichment cache, PagerDuty/Opsgenie/Teams/Slack, Prometheus metrics, Helm chart |
incident-scribe |
Slack thread → blameless incident report with deterministic timelines, local secret redaction and a published JSON Schema |
sre-ai-toolkit |
CLI tools: alert explainer, log anomalies, Terraform plan review with a CI risk gate, Kubernetes event triage |
runbook-rag-starter |
The smallest useful RAG, in 40 lines — a Markdown folder, sqlite-vec, and a CLI |
cloudandsre-python |
Shared utilities: async circuit breaker, retry, throttle, a token-cost meter and OpenTelemetry GenAI attributes |
agent-adam |
Agentic personal assistant: least-privilege Google OAuth, keychain token storage, queued destructive actions |
They share one engineering baseline: CI with tests, CodeQL, OpenSSF Scorecard, SHA-pinned actions, Dependabot, release provenance, and a citation file.
In progress: mcp-gateway — identity and scope-token exchange, Cedar authorization, schema
firewall, quarantine queue for destructive calls, provenance ledger
(design).
Self-Healing Infrastructure: Building Autonomous Cloud Systems with AI —
free on Leanpub · DOI 10.5281/zenodo.22245770 · companion labs in
self-healing-infra-labs
AI platform: Azure OpenAI / Azure AI Foundry · Anthropic Claude · MCP servers · GraphRAG Cloud & infra: AWS (ECS/Fargate) · Azure · Kubernetes · Terraform · Databricks Graph & context: Neo4j · Cypher · knowledge-graph-backed retrieval Platform engineering: Backstage · SwaggerHub · OpenTelemetry · Prometheus Languages: Python (primary) · Go · Bash · SQL / Cypher
Every post ships with working code: cloudandsre.com.
- 15+ years architecting and operating cloud systems at scale
- Currently Staff Software Engineer at a global aviation-SaaS platform (a Fortune-500 aviation technology company); previously Computing Architect there for 8 years
- 2x founder before architect — Veeble SoftTech (scaled 3→20, $500→$500K ARR, 110+ countries) and ExTravelMoney Technosol (CEO)
- MS Management Information Systems, University of Arizona · MS Artificial Intelligence, Indiana Wesleyan University (in progress)
- Azure AI Engineer Associate · AWS Solutions Architect Associate · Azure Administrator Associate · Neo4j Certified Professional · CNSS 4011/4012/4013
- IEEE member for 15+ years · Senior Member application in progress