what we've shipped.

Ollie isn't a wrapper — it's the actual control plane behind everything we ship. five specialist agents (main, research, content, comms, ops) plus a Voice Agents tier for receptionist work, all running on the Claude Agent SDK with a shared "business brain" of versioned brand voice, ICP, and sequence playbooks. plug-in adapters for Cal.com, Twilio, ElevenLabs, AgentMail, Composio (Gmail/Calendar/Drive/Notion/Slack/HubSpot). human checkpoints, review queue, live event stream, full memory + cost telemetry. when we sell agentic infrastructure, this is what we're selling — proven on us first. since launch it's grown into a fleet: one-command box provisioning, supabase-backed login with role-based access, a per-box business brain behind an authenticated API, and pinned-image deploys with health gates — boring, repeatable infrastructure, which is the point.

eight skills spanning a full engagement: outbound-fit (a go/no-go gate that sometimes says "don't buy this"), qualify-accounts, pipeline-estimate, stack-design, campaign-design, build-roadmap, voice-capture, and pre-launch-audit — an adversarial auditor that runs with fresh context so it can't inherit the author's assumptions. the voice system is the moat: you hand-edit a small batch, the system diffs your edits into rules, and drafting refuses to run without your voice spec. sequencer-neutral by design (smartlead today). the machine drafts and audits; a human always clicks send. jnow runs as client zero on the identical structure a client gets.

not a toy demo — olivia runs live on jnow's production box today, and she's the same agent a client gets on theirs. every skill is scoped with a defined trigger, input, and output; the brain keeps per-client state so work compounds instead of restarting. the best proof that AI marketing works at business scale is that we bet our own pipeline on it.

the protocol lives in versioned standing issues that every agent re-checks on each heartbeat and confirms with a receipt — update the protocol once, the whole fleet adapts. strict one-task heartbeats keep agents predictable; a hold/block system routes questions to the right channel (on the issue vs. the operator); and a hard safety boundary means agents ask before anything customer-facing ships. thirteen receipt tokens make every action auditable by anyone who can read a linear issue.

two-model architecture: Sonnet handles the conversational turns, Opus generates the final readout. session state in KV with 30-minute TTL. hard 12-turn cap. voice-eval'd against 20 fictional scenarios before launch. cost: ~$0.07 per completed session.

solves a real problem Mike lived for 20 years: every SE rebuilds workshop content from scratch for every new product. LabSmith automates the 80% that's structural so the SE can focus on the 20% that's customer-specific.

can't name the company or show the code — it was internal tooling at a previous employer. including it because it's the same kind of operational automation we build for clients: take a messy process that depends on someone noticing a problem, and replace it with a system that surfaces the problem before anyone has to notice.

these aren't brochure sites with a chatbot bolted on. the AI is the site. job-fit analysis, interactive Q&A, peer-review data, expandable context — all running against live models. we built these for ourselves before we built for anyone else.

this wasn't a consulting engagement — John built the entire GTM motion from zero. ICP definition, messaging framework, sales process, comp model, and the actual software tools the sales team uses on discovery calls. the trapped cash estimator lets prospects enter their billing metrics and instantly see a personalized projection. the report generator reads XBRL data from SEC EDGAR — real revenue, receivables, and DSO, not estimates — computes cash-release scenarios, and ships as a one-page branded PDF; built for a live campaign against named public-company targets.

four layers: a register of the actual texas rules (kept current), a posture layer that configures what applies to this brokerage, gates on execution (clearly-illegal requests are blocked before the AI even runs; unscreened output is held, never sent), and an evidence log that becomes a monthly TRAIGA-readiness report you can hand a regulator or insurer. honest about its limits by design: it proves the checking happened — it is not a lawyer, and hard calls still go to a human. the same governed-AI pattern deploys to any regulated vertical.

the framework's signature move is the honest limit: every control states what it can't prove ("proves screening ran, not that its verdict is correct") — because a governance document that over-promises is worse than useless when anyone checks. the obligations overlay ties controls to the contracts, NDAs, and regulations a client is genuinely bound by, not a downloaded template. the first four controls aren't theoretical — they run in production today inside our governed compliance builds.

built specifically for the kind of small field service businesses JNOW works with. multi-tenant architecture with row-level security so each org's data is fully isolated. mobile-first UI with optimistic updates for fast performance on spotty connections — because technicians are in crawl spaces, not conference rooms.

a working example of the document processing pattern we deploy for clients. takes any inspection report format, extracts every issue, classifies by severity, estimates cost ranges, and generates a branded summary. the exact same pipeline we'd build for an insurance company, law firm, or property manager — different documents, same architecture.

end-to-end mobile app shipped to both app stores. sub-500ms voice transcription latency. AI-powered semantic tagging surfaces related ideas before they get lost. offline-first architecture with Supabase sync when connection restores. proves we can build and ship across platforms, not just web.