Get a free assessment
Back to Chimzee

The engine room

Most agencies show you a case study. This is the machine itself, the one we run our own company on, and the same one your work would run on. You never have to touch any of it. We just think you should be allowed to look.

9systems running in production
173,802lines of code written and running
55technologies wired together
15upgrades it made to itself

Every frontier model, and the ones that never leave the building

Each job goes to the model that is genuinely best at it, not to whichever brand we happen to sell. This list is read live from the machine, not typed by hand.

xAI / Grok
Grok 4.5Grok SuperHeavy (multi-agent)
voice, vision, fast live reasoning
OpenAI
GPT-5.6 SolGPT-5.6 TerraGPT-5.6 Luna
transcription, strict interpretation, fallback reasoning
Claude
Claude Sonnet 5Claude Fable 5Claude Opus 4.8
writing, analysis, orchestration
Gemini
Gemini 3.5 FlashGemini 3.1 Pro
alternate review and research
On this machine
nomic-embed-text:latestcatters:latestcatters-fast:latestqwen2.5:14bllama3.1:8bllama3.2:1b
the work that never needs to leave the building

What it has actually built

Not mockups. Systems that run every day, for real businesses, on real data.

Agentic Brain (this site)

The live model council with result cards and a self-improvement feed. Successor to the original brain, whose council answered real visitor questions and published 51 full plans.

Hudson's Pizza voice agent

AI phone agent for a real pizzeria: answers calls, takes orders, hands off to a human when needed. Call +1 (844) 974-1968 and an AI picks up.

rlookup.io restaurant recon

PIN-gated intelligence tool used before sales calls: Google data, online-ordering setup, POS, compatibility signal.

Territory planner (/plan)

Mobile field-planning tool: pick a zone and batch size, get an optimized door-knock route with live hours, intel and visit logging.

Coganut lead-gen platform

Lead generation and market-intel platform built for a marketing group client.

Lead enrichment pipeline

Turns raw business lists into ranked, contact-ready leads: 1,842 enriched and 532 ranked to date.

Prospecting engine (sales CRM)

The system a seller actually works out of: 425 restaurants enriched with what they run (online ordering platform, POS, catering setup, what is broken on their site), then queued by timezone so every call lands in business hours. It watches for replies, drafts the follow-up, books the callback on the calendar, and reads a pre-call brief out loud before the phone rings. 80 replies logged so far.

Server infrastructure

The self-hosted machine under all of it: reverse proxy, TLS, monitoring every 3 minutes with auto-restart, encrypted backups.

It keeps improving itself

The loop never stops: attempt, critique, prove, then do it cheaper next quarter. These are the most recent upgrades the system made to itself.

Model names are no longer typed by hand anywhere. credit_check.py now discovers the newest model on each key (ranked by release date, since xAI ships a 'grok-4.20' that is older than grok-4.5), calls it, and writes only the models that actually answered to models_latest.json. The status page, the /ai model rail and the copy in the panels all read that one file, so a new release from any vendor propagates on its own within one check.

Jul 13

One scanner now feeds both the admin usage page and the public counters, so they cannot disagree. It counts hand-written source only (.py .ts .tsx .js .jsx .html .css .sh .php .sql) and excludes vendor code, build output, backups, dated duplicates of a live script, and saved copies of other people's pages. Three technologies were added after verifying each one on the host: Twilio, the Google Places API, and the OpenAI realtime voice API.

Jul 13

docMode defaults to simple (was tech). The toggle and the per-panel Simple/Technical switch are unchanged, so the deeper detail is one tap away for anyone who wants it.

Jul 13

GPT-5.6 Sol wired end to end: council GPT seat moved from gpt-5.2 to gpt-5.6-sol, the model catalog behind the live status rail now reports it, and on-server Codex was authenticated and set to gpt-5.6-sol (confirmed on the wire, not just in config). Sibling tiers Luna and Terra are available on the same key.

Jul 13

Declutter pass: welcome gate + name aliases removed (silent visit tracking retained), engagement nudge and one-card explore panel removed, latest-upgrade promoted to a full-width top bar above the HUD, Show me around now opens the section callout tour directly, and the collapsed dock pill is a visible text input inviting a multi-model council question.

Jul 12

Non-blocking welcome banner (auto-dismiss + first-interaction dismiss), single consolidated Projects/Status/Agents/Ideas panel with live health data, preset asks bypass the intake and call /ai-ask/plan directly, staggered single-callout onboarding, and a positioning thesis line. Every step Playwright-verified on the live site.

Jul 12

That is the engine. What you deal with is a conversation, a thing that works on Monday, and someone to call when it does not. If you want to poke at a piece of it yourself, the brain is open.

Ask the brain, free Get a free assessment