Two production systems for one staffing firm. Five case studies. The products on our shelf, and the systems under all of it. Every number on this page maps to a receipt: its measurement method, its scope, and the date it was verified. Open any marked number to read it. No receipt, no number.
The flagships: two production systems, one staffing firm
The client stays unnamed. The work speaks. Both systems run for the same staffing firm: the agent platform that operates the back office, and the marketplace that gates its subvendor supply chain. The vertical, in full: the staffing page.
Flagship 01 · Agent platform
An agent fleet runs the back office.
Finance, sales, recruiting, account intelligence, federal capture: specialist agents plus an apex that sees the whole firm, speaking to operators over chat with voice, documents, and vision. The governance is mechanical. Agents read through views only, every table carries row-level security, and every outbound message rides a human approval bound to the exact content hash. LLMs advise. Code and humans decide.
Its public face is the eye above: a deterministic WebGL iris, seeded once and frozen, each specialist owning a true angular sector of the whole. Built and gate-tested; the redesign ships next.
sevenspecialized agents
1,209runtime test cases
94governed tables
Flagship 02 · Compliance marketplace
Roster: a compliance gate fused to a supply chain.
The same firm's subvendor marketplace, live in production. Requisitions release only to vetted partners. AI reads the insurance, tax, and banking documents; a deterministic rule engine issues the verdict; a human approves; the agreement e-signs in production. Placements ride an append-only event spine with a hard ownership lock, and a margin firewall keeps every bill rate out of the portal and out of every model prompt.
Document to decision in ~5 s on the compliance plane. The splash above is live today; the data on it is illustrative by design, because a compliance product does not fake dashboards.
106krequests, zero failures, 200 VU
1,160test cases gating CI
12 mindowntime, live DB migration
Case studies
A study lands here only after every number in it has been verified against a receipt. Newest first.
Our own production voice agent. It answers the phone on the work page, books real meetings on a real calendar, leaves a receipt for every call, and was audited against the catalog we sell before we sold it.
One agent substrate, two fleets: a seven-agent operational platform for a client, and the system that runs our own portfolio. Same memory, same governance, same uptime discipline.
A subvendor marketplace for a staffing firm, live in production: AI document review feeding a deterministic rule engine, humans on the exceptions, and a placement lifecycle riding an append-only event spine behind it.
email to decision, measured end to end
requests, zero failures, 200 VU
test cases gating CI
The shelf: products we run
The public tier of the portfolio. The live ones open right now; every number carries its receipt. The screenshots live in the case studies, where they are evidence.
Marrow open source
semantic memory for production agents
Persistent, citing, honest memory over an operating history. Hybrid retrieval with reciprocal rank fusion; answers cite their sessions or abstain. It is the layer under everything else on this shelf, and under our own seven specialized agents. The engine is now open source, Apache-2.0.
A deterministic state kernel under the model: full event-sourced history, a 36-table spine, tested and containerized, carrying a novel-scale canon. Proof the stateful discipline holds at narrative depth.
Client and internal systems, so described rather than linked. Every one of these is whiteboard-able, in depth, on a call.
01
Vendor-compliance platform
The document plane behind Flagship 02 above. Email to decision in ~5 seconds, proven end to end. AI document review: vision extraction → deterministic rule engine → human-review queues → e-signature → ERP sync. Multi-tenant, AV-scanned, fully audited.
02
Cognitive runtime
A production AI memory and identity substrate running since Q1 2026: persistent semantic memory, scheduled reflection cycles, multi-channel (chat, terminal, API), row-level-security firewalls. Recently live-migrated between databases with 12 minutes of downtime. Fail-closed security proven on prod, rollback armed throughout. AI systems with state are a different sport, and we play it daily.
03
Agentic control plane
The substrate under Flagship 01 above. Seven specialized agents over an operational warehouse: event bus, per-agent cognitive schemas, governance rules enforced in the database, and honesty contracts in code. The API returns “source unavailable” instead of fabricating when upstream is dark.
04
Autonomous PM agent
Runs our own portfolio: scheduled digest and drift-detection cycles, an AST-validated SQL sandbox, role-level cross-database firewall. The fleet manages itself so the humans can build.
05
Stateful AI engine
Event-sourced AI applications where state, memory, and consequence persist across long sessions: a deterministic state kernel under the model → full event-sourced history → dozens of normalized domain tables, with a complete test suite and container pipeline. The hard kind of AI: it does not reset between sessions or change without a trace.
06
Real-time data-ingestion pipeline
Forty-plus live external APIs → multi-source normalization → an algorithmic model layer → ranked, explainable decisions on an automated daily cadence. The same ingest–normalize–decide pattern we put on invoices, claims, market signals, and intake docs.
07
Autonomous orchestration harness
Complex work run as structured, checkpointed phases: fan-out → adversarial verify → consolidate → resume-on-failure. The control pattern behind agents that finish multi-step jobs without a human in every loop.
