Custom AI Service Bureau

Vertical AI surfaces with their own curated knowledge, billed at the floor.

Generic AI tools answer everyone's question with everyone's data. Kevin runs domain-tuned surfaces — clinical, civic, research, IoT — each with a fine-tuned corpus and multi-LLM routing that keeps token cost at the floor, not the ceiling. One contract, many verticals.

Try a vertical Talk to us

What Kevin is

Domain-tuned, not generic.

Kevin is a custom AI service bureau. Each surface is a narrow, domain-tuned chat tool with its own curated corpus behind it: a clinical surface tuned by clinicians, a meeting surface tuned by deal context, an IoT surface that knows the firmware. Behind every surface, a compression layer routes each query through least-cost inference and keeps the per-token bill at the floor.

Customers buy access to the surfaces they need. They don't buy GPUs, they don't curate corpora, they don't watch token spend. The corpus is included; the billing is a tier, not a meter.

The family

Each member owns one job. Together they're the platform.

Orchestrator
Marco
Orchestration layer. Reads directives, coordinates the family, routes work to the right surface.
Meetings · transcript · counterparty alerts
Saul
Real-time meeting intelligence with a fine-tuned KB per engagement. Counterparty-aware alerts, live.
Knowledge · RAG router
Annie
Routes every query to the right corpus and the lowest-cost LLM that can answer it. Multi-domain by design.
Observability
Mona
Sees everything. Fleet-wide health, alert pipeline, and the dashboard customers watch.
Hardware · firmware
Max
ESP32-class devices on a LoRa mesh. Same widgets and bindings as the rest of the platform.
Keystore · vault
Hank
Reactive key-value store. Every key is subscribable; every write is auditable.
Auth · entitlements
Grant
Says where anyone goes. Per-customer tokens, per-domain scopes, pluggable across products.
Compliance · rules
Trixie
Policy enforcement layer. HIPAA, regulatory, deal-specific clauses — turns a corpus into a guardrail.

How a query flows

Three paths. The same bus.

Knowledge query

customer chat
Annie picks domain →
vector retrieval
cost-routed LLM
citations + answer

Telemetry stream

device
MQTT
Mx bus (LVC + history) →
Mona dashboard
alerts on rule match

Holon dispatch

directive
holon mailbox
capability-routed worker
produced commit
review & merge

Customer surfaces

Pick the one that fits the use. Switch tiers without changing wiring.

Starter
$9 / mo
  • 500 queries / month
  • Embeddable chat widget
  • One vertical surface
  • Best-effort hosting
Pro
$29 / mo
  • 5,000 queries / month
  • Embeddable widget + API
  • Two vertical surfaces
  • Curated corpus updates
Vertical
$99 / mo
  • Unlimited within domain
  • Hosted app at {domain}.iiiai.us
  • Per-engagement KB
  • Counterparty-aware alerts
Dev
$0.02 / query
  • Pay-as-you-go
  • Or $15/mo for 1,000
  • Raw API access
  • Bring your own corpus

How we build

Four layers. Customers see one product; we run the whole stack.

Directive layer
Operator's directive becomes a graph of work — no slide decks, no drift between intent and execution.
Graph + dispatch
Capability-routed holons. Each work item carries its required skills and lands on a worker that can do it.
Worker fleet
Workers pull, execute, produce a result with citations. Every step is observable, every artifact is traceable.
Fabric telemetry
Routing decisions, cost savings, latency, error rates — captured as receipts customers can verify.

Where to start

Three doors, depending on what you need.

Try the demo

Hosted vertical you can interact with right now — research desk for US presidents, fully cited. See the surface before talking to anyone.

presidents.iiiai.us →

Embed a widget

One <script> tag drops a chat box on any page, scoped to a domain you choose. Starter tier covers small sites.

Get a domain key →

Spin up a vertical

Send your seed corpus, we host the surface at {your-domain}.iiiai.us within a week. Per-engagement KB, counterparty-aware.

Start a trial →