Senior Software Engineer · Tri-Cities, Tennessee
William “Billy” Bratz
Senior software engineer with 11 years building distributed systems and, since late 2025, production LLM agent systems. I founded the MCP server and orchestrator that became my organization's LLM platform, and I turn difficult operational lessons into durable systems, standards, and writing.
Experience
Carvana
Senior Software Engineer II
LLM Platform: Conversational AI
Jun 2026–Present · Python, FastAPI, FastMCP, Kubernetes, Datadog
- Operate an MCP platform serving more than 10 million weekly tool calls, connecting conversational AI to production systems.
- Root-caused a multi-service outage in minutes across a five-hop request chain, halting a harmful redeploy and directing the recovery.
- Demonstrated that an incident remained active after apparent recovery because of timeout asymmetry, changing how the team validates service health.
- Designed and shipped a same-day in-cluster route that delivered a double-digit p95 latency reduction while preserving the public path as a safe fallback.
- Found and corrected a CI isolation defect across 383 tests, restoring zero-network unit-test guarantees.
- Resolved an MCP HTTP 421 outage using APM evidence, then added a dependency-host guard, pinned lockfile, monitors, and an audit of sibling services.
- Eliminated order-of-magnitude retry amplification and validated the change under live production traffic.
- Shipped three customer-context MCP tools, including the platform's first MCP-to-MCP workflow, within service-level objectives.
- Used production log analysis to isolate and correct a timestamp race, then verified the fix under load.
- Designed a split test for a model migration covering a significant share of assistant traffic.
Verifications & Underwriting
2020–Jun 2026 · .NET, Azure, Kafka, Kubernetes
- Founded an MCP server and orchestrator in December 2025. I wrote roughly 6,650 lines in six days and supported it with a 445-line operator guide. It became the foundation of the organization's LLM platform.
- Built an autonomous delivery pipeline that completed 17 runs over 10.6 hours and produced 20 reviewable commits.
- Created an internal marketplace of nine agent plugins and introduced recursive language-model techniques for work beyond a model's normal context window.
- Designed the operating contract for a custom Claude Slack integration used in production investigation and engineering workflows.
- Architected a fraud-analysis service with 13 ADRs, C4 diagrams, and five runbooks to support a sixfold growth target.
- Led a workload-identity migration spanning five shared packages and more than five services; wrote the 1,000-line implementation playbook used by other teams.
- Reduced advocate insurance work from more than 20 minutes to under four, and delivered an OCR proof of concept in three days with an error rate below 2%.
- Cut one service's memory ceiling from over 1.5 GB to under 243 MB, reducing total memory utilization from 60–80% to below 10%.
- Contributed across services to a top company rearchitecture supporting co-buyers, while mentoring engineers on domain workflows and dependencies.
- Authored 15 architecture decisions and multiple engineering standards and root-cause analyses. Opened 143 pull requests and reviewed 177 across 20 repositories in 28 weeks; 58% of the work was self-scoped.
CoStar Group (formerly STR)
Software Engineer
- Served as technical lead for Instant Insights, decomposing monolithic services into reusable .NET microservices and building a React and Redux analytics experience.
- Joined a three-person team integrating STR's global hospitality data into CoStar Suite, a product serving more than one million daily searches.
- Built a report-generation platform with AWS Lambda and Step Functions, plus reusable translation and currency-conversion modules.
Writing & Open Source
WilliamBratz.com
Nine essays on production engineering and AI systems, including “Production Reveals the System,” “From Chatbots to Persistent Organizational Memory,” and “Engineering in the Age of Agents.”
vaultkit
A personal, LLM-maintained knowledge engine: roughly 6,400 lines of Python, more than 320 tests, three scheduled maintenance cadences, and a scaffold adopted by four teams. The reusable vault-scaffold workflow is available on GitHub.
Billy's AI Skills
Cross-client Recursive Language Model plugins for Claude Code and OpenAI Codex, with bounded fanout, dry-run planning, local model support, automated tests, and a credential-free public demonstration.
Hugo
A self-hosted Slack reading assistant that curates technical RSS feeds with Claude, summarizes URLs and threads on demand, and delivers a configurable daily digest.
Independent products and open source
Built Neverending Story Pointer, a free real-time estimation app, and How Unique Is a GUID?, an interactive probability experience also available for UUIDs. Contributions to RestSharp and NUnit.
Technical Skills
- Languages
- Python, C#, TypeScript, JavaScript, SQL, GraphQL
- AI & agents
- Agent orchestration, MCP server/client/gateway federation, FastMCP, OpenAI and Anthropic APIs, PydanticAI, tool calling, structured output, RAG with Pinecone, prompt experiments, evaluations, observability, Claude plugins and skills
- Backend
- FastAPI, asyncio, Pydantic v2, .NET Core, REST, JSON-RPC, gRPC, Kafka, Azure Service Bus, event-driven systems
- Data
- Cosmos DB, Redis, PostgreSQL, Azure SQL, SQL Server, DynamoDB
- Platform
- Kubernetes (AKS/GKE), Docker, Helm, Tekton, Flux, GitHub Actions, Azure DevOps, Lambda, Step Functions, Workload Identity, OIDC
- Observability
- Datadog, Splunk, Prometheus, Grafana, OpenTelemetry, SLOs, Locust
Education
Arizona State University
B.S. Software Engineering · In progress
Volunteer State Community College
A.A.S. Computer Programming · 2022
Phi Theta Kappa · Dean's List