Long-Term AI Engineer — Autonomous Agent Orchestration (Claude Code + Open Models)

We are a company running multiple brands — AI video production, AI-powered real estate technology, and a network of medical compliance clinics. We're hiring an AI engineer who builds with autonomous agents the way developers used to build with frameworks. THE HONEST PITCH Work that used to cost us $100k/year in dev contracts now ships in a week with the right person and the right stack. We're looking for that person — not someone who "uses AI for coding," but someone who's internalized that the harness around the model is the actual product, and who treats Claude Code, Kimi K2, GLM, Minimax, and Qwen as a routed orchestra, not competing chatbots. WHAT YOU'LL ACTUALLY BUILD - Multi-agent workflows that run unsupervised overnight (planner / generator / evaluator architectures, not single-shot prompts) - Real estate data pipelines pulling MLS, HUD records, and public sources into Supabase with webhook triggers - Internal tools and dashboards for non-technical operators across our brands (we have a backlog) - Custom MCP servers when none exist for what we need - WordPress / Next.js sites cranked out fast with AI-native workflows - Voice and video automation pipelines (Whisper, Seedance, fal.ai — bonus if you've touched these) - Whatever the brief is next week — the work is varied and we move fast WHAT WE PROVIDE - Stable, recurring, long-term work — predictable hours and weekly pay, not feast-or-famine gig hunting - Paid access to every coding model and tool that matters — Claude Max, Kimi, GLM, Minimax, Codex, GPT, Gemini - Infrastructure budget (Supabase, Vercel, Cloudflare, Runpod, VPS) - Direct line to ownership — no project manager between you and the brief - A real growth path — top performer becomes the lead and brings on junior builders under them WHAT WE DON'T WANT - "I use ChatGPT for code" people - Anyone whose AI workflow stops at copy-paste - WordPress-only developers with no scripting depth - Gig-hoppers looking for a quick one-off — we're investing in someone for the long haul - AI-polished proposals that open with "I am thrilled to apply for this exciting opportunity" HOW TO APPLY — READ CAREFULLY The first round is a video. Record a 5–10 minute screen-share (Loom or unlisted YouTube) walking through your actual setup, in your own voice. We're not looking for production value — we want a builder who can show us their real stack, live. If it sounds like you're reading an AI script off a teleprompter, that's as obvious to us as an AI-written cover letter, and we'll pass. In the video, cover: 1. Who you are, your time zone, and your daily driver — Claude Code? OpenCode? Cline? Something you rolled yourself? Show the actual environment. 2. How you route your models — orchestrator brain, primary coder, fallback, and how you decide. If you have a router or config, put it on screen. 3. What agents you run — subagents, skills, slash commands, hooks, MCP servers. Show the harness, not a description of it. 4. One or two real things you've shipped. Bias toward what we care about: overnight multi-agent workflows, real estate / public-data pipelines into Supabase, custom MCP servers, or voice/video automation. Screen-share the repo or live product — show the commit history. 5. The honest part: a production failure. What broke, what you did, what you learned. Proposals without a video link will be declined. We watch every video that comes with honest answers. BONUS (NOT REQUIRED, BUT NOTICED) - Self-hosted observability (Langfuse, Helicone, claude-code-hooks-multi-agent-observability) - Understanding of the "lethal trifecta" and prompt-injection mitigation - FFmpeg / video pipeline work - Real estate, mortgage, or regulated-industry experience - Public writeups, blog posts, or X threads on agents - GitHub commits in the last 30 days DETAILS - Hourly, long-term, ongoing — measured in months and years, not weeks - Consistent 30–40 hrs/week with predictable weekly pay - 3–4 hours of daily overlap with US Eastern required - Where you land in the range depends entirely on what you demonstrate in your video and proposal. We set the rate by what we see, not what's asked — and strong performers grow over time. What You Need to Have Actually Done (proof needed) Built full agentic coding harnesses — orchestrator → subagent → CI gate → merge loops, end to end Deployed Daytona (or equivalent: E2B, Modal) as an isolated sandbox execution layer for AI-generated code Run parallel Claude Code agent sessions via Git worktrees — multiple simultaneous subagents on scoped tasks Built or configured a self-hosted LiteLLM gateway routing to multiple models (Claude, GPT, Gemini, DeepSeek, Kimi or similar) Wired up MCP servers (file system, SQL, Atlassian, Slack, or similar) as tool layers for agents Structured agent harnesses using CLAUDE.md, COMMON_MISTAKES.md, subagent role definitions, hook scripts, and settings.json — or a directly comparable framework Implemented Plan Mode or equivalent human-in-the-loop approval gates before agent execution Built something like the 7-agent feature factory pattern (Researcher → Spec Writer → Builder → Validator → Verifier) or Dynamic Workflows / multi-agent orchestration equivalent Used Temporal, n8n, or similar for durable long-running agent workflow execution Treated CI green as the mechanical enforcement gate for agent output quality