Sporfie
AI Platform, Full Stack & DevOps ArchitectFeb 2016 – presentsporfie.com
Since 2025: from LLM-assisted debugging of production issues to an agentic AI platform, designed and shipped in 2026 and live in production. It has two halves: Sporfie Agents, where the team investigates and operates, and CodeDev, where coding agents build.
Sporfie Agents (Python / FastAPI, AWS Bedrock, MCP): a multi-agent system for investigation and operations. An LLM router sends each question to one of five specialist agents (infrastructure, operations, design, development, tech lead), each a persona with a fail-closed tool allowlist and its own case memory. 19 read-only MCP servers over Kubernetes, CloudWatch, Prometheus, GitHub, SQL, Jira and more; any change goes through a signed plan, a human approval in an AI-free app and a separate executor, with a full audit trail. Reachable from the web, Slack, Claude Desktop (MCP) and GitHub, where a Tech Lead agent reviews every PR. One LiteLLM gateway gives each agent a key, a budget and a cost per conversation.
CodeDev (TypeScript / Node, React, Kubernetes): a cloud workstation where Claude Code (Agent SDK), Codex and OpenCode build and run the real stack and hand back a branch, a PR and a verified preview URL. Per-branch preview environments, an epic mode that runs parallel agents in git worktrees against a server-checked definition of done, deterministic commit and push guards as hooks, and an MCP so other agents can delegate coding work.
Evals on both halves, so behaviour is checked by code, not by reading transcripts. Sporfie Agents runs a behaviour-regression suite (real model, mocked tools) before every release; CodeDev admits a model only after it passes graded coding and design tasks with a blind review. The target state, which both are built toward: a versioned eval set per agent and per skill, pass rates tracked per model and per prompt change, and a prompt or model swap gated the same way a code change is.
Platform and back end (2016 onward): building and running the platform end-to-end since day one of this live-streaming SaaS. Led the move to Kubernetes in 2020 with a strong GitOps + IaC culture; introduced Karpenter for smarter autoscaling, built a Kubernetes-native serverless framework to bypass Lambda limits, and run GPU workloads for streaming and AI tooling.
Own multiple back-end services (Java/Go/Python/Node), lead performance and scalability work, create load-testing suites, and implement multi-layer caching (including CloudFront for less-dynamic public APIs). Hands-on with CDN architecture and operations: cache strategy, invalidation, edge behavior, origin design, and the observability needed to troubleshoot real-world delivery issues.
Stack: Kubernetes, ArgoCD, Karpenter, Prometheus, Grafana, Docker, PostgreSQL, Java / Spring Boot, Node.js, Python (OpenCV, FFmpeg), Go, Firebase, GraphQL, REST, Redis / Redisson, HAProxy, CircleCI. AWS: Lambda, CloudFront, S3, ElasticTranscoder, CloudWatch. AI: AWS Bedrock, MCP, Claude Agent SDK, LiteLLM, Python / FastAPI.