Matt Butler

Production, Zillow-style SaaS for pre-foreclosure and distressed-property data: an interactive geospatial map with tiered access, live at app.foreclosurebase.com. Designed, built, deployed, and operated solo.

• Full-stack: Next.js (App Router) + TypeScript + MapLibre GL frontend; FastAPI (async Python), async SQLAlchemy 2.0, and Pydantic v2 backend; PostgreSQL + PostGIS for GiST-indexed geospatial bounding-box search
• Data pipeline: ingests county recording feeds over SFTP, entity-matches filings to parcels, geocodes, and scores leads; idempotent upserts on natural keys; Celery + Redis task queue on a daily schedule
• Server-side tier gating that strips restricted fields on every path that can emit them (not just CSS); NextAuth (Auth.js) Google-OAuth JWT auth; per-tier Redis rate limiting; Stripe billing; strict CSP/HSTS security headers
• Deployed GitOps-style on Kubernetes (Oracle OKE) with Argo CD: self-healing, SHA-pinned images, zero-downtime deploys; automated daily Postgres backups to R2 with a verified restore drill; observability via OpenTelemetry, structlog, and Grafana Cloud (traces + RED metrics)

Next.js, TypeScript, FastAPI, PostgreSQL/PostGIS, SQLAlchemy, Pydantic, Celery, Redis, Stripe, NextAuth, Kubernetes, Argo CD, Cloudflare, OpenTelemetry, Docker

AI operations agent that monitors real-time Flink/Kafka pipelines, detects anomalies, diagnoses root causes with a multi-agent LLM system, and escalates with human-in-the-loop, recommend-never-execute safety.

• Streaming layer: Apache Flink 2.0 job (keyed state; hybrid static-threshold + EWMA-baseline anomaly detection keyed per source/metric) over Kafka (KRaft) with Protobuf events; Java 17
• Multi-agent LLM system (Python): a hub-and-spoke coordinator spawns context-isolated diagnostic and report sub-agents; agentic loop driven by Claude’s stop_reason; tools exposed via a Model Context Protocol (MCP / FastMCP) server over the Flink REST, Kafka consumer-group, and Prometheus APIs
• Trustworthy-AI patterns: structured outputs (Pydantic), claim-source attribution with conflict escalation, confidence scoring, and a draft-only (recommend-never-execute) contract enforced at the prompt, schema, and escalation layers, with a persisted incident audit trail
• Prometheus + provisioned Grafana dashboards; Docker Compose stack; 200+ tests (JUnit 5 + pytest) with CI secret scanning

Apache Flink, Kafka, Java 17, Python, Claude API (Anthropic), MCP / FastMCP, Pydantic, Prometheus, Grafana, Protobuf, Docker

Core developer on a distributed big data streaming platform comprising ~56 microservices (.NET, Node.js, Java). Owned the Prism Bet Events App — a high-throughput Apache Flink streaming application serving as the totalizator, processing tens of thousands of bets per second across distributed systems. Designed, built, and automated real-time event processing pipelines and infrastructure provisioning serving downstream market odds, regulatory, and analytics systems.

• Owned the design and evolution of a 48-operator Flink DAG with 13+ keyed state descriptors, serving as the real-time totalizator — aggregating all bets across wagering channels into pool totals that drive live odds on the tote board
• Designed and implemented deterministic market closure system replacing time-based mechanism, tracking validator consensus with RocksDB-persisted state and crash recovery
• Diagnosed a checkpoint OOM crash caused by a GCS buffer leak spanning four bugs across Google’s client library, the Hadoop GCS connector, and Flink’s state uploader and GCS plugin, each failing to release 67MB upload buffers on checkpoint failure. Resolved by replacing the GCS checkpoint path with Flink’s S3A plugin targeting GCS’s S3-compatible API
• Designed and built a Flink State Inspection CLI and Web UI for reading, analyzing, and diffing broadcast and keyed state from production savepoints and checkpoints stored on GCS. Built on the Flink State Processor API, the tool runs locally without cluster access. 113 unit tests, Dockerized with Helm chart for GKE deployment
• Built Kafka topic archival system with race-condition protection and deletion barriers — zero data loss across graceful restarts and crash recovery scenarios
• Identified Kubernetes HPA misconfiguration (30:1 CPU limits-to-requests ratio) causing pods to scale from 2 to 10 replicas; prevented cluster resource exhaustion
• Implemented protobuf deserialization resilience handling multiple wire formats including .NET Confluent.Kafka producers, eliminating poison-pill job failures
• Fixed regulatory compliance gap ensuring collation matrices emitted for all Pick 4/5/6 markets regardless of bet volume
• Designed adaptive logging infrastructure with ring buffer and automatic error escalation, reducing production log volume while maintaining visibility during critical issues

Flink, Kafka (Confluent Cloud), RocksDB, Protocol Buffers, Confluent Schema Registry, Kubernetes (GKE), Istio, Flink Kubernetes Operator, Docker, Helm, Google Cloud (GCS), Prometheus, Java, Tilt, JIRA

• Contributed to the research, design, architecture, and delivery of a platform enabling Amwell partners to register for WebHook/CloudEvent notifications when their patients use Amwell’s telehealth services
• Events encompassed revenue, visit scheduling, notifications, user analytics, and more via a pipeline of Redis Streams → Kafka → Flink (conform) → Flink (enrich) → event publication

Flink, Kafka, Redis, Kubernetes, DynamoDB, WebHooks, CloudEvents, Google Tag Manager / Matomo / MySQL, FHIR Query, Google Cloud / FHIR

• Built a Scala-based parser, interpreter, and high-performance execution engine for CQL and FHIR standards on RocksDB — scans 400 million records and computes rules in approximately 7 milliseconds

Kafka, Flink, Scala, Java, RocksDB, Debezium, ClickHouse, Confluent Cloud, KStreams, .NET

• Implemented WsFed SSO via IdentityServer4 / .NET Core
• Developed .NET 4.6 MVC REST API and third-party integrations (BambooHR, Box.com)
• Built Docker stack for Couchbase and Elasticsearch

• Implemented authentication, authorization, and federated services using LDAP, SAML, and Siteminder
• Developed “What We Sell” — a content management platform serving ~107,000 users, winner of the ITSMA Marketing Excellence Award
• Responsible for DevOps of 30+ servers and 45+ databases
• Developed backend for HP Experience Center and Cloud Maintenance Utilities for HP Helion