01 / Frontier voice agents↗
Persistent operators for calls that matter.
Sub-second turn-taking, durable memory, escalation contracts, and tool access for care, billing, and field operations.
STOIC LABS — RESEARCH OPSv III · MMXXVI
● BUILDING THE CONTROL PLANE FOR HIGH-STAKES WORKVOICE · DOCUMENT · BROWSER · ERP OPERATORSPRODUCTION LOOPS — NOT DEMOSRESEARCH-LED · OPERATOR-FIRSTFOCUSED ON CARE · CLAIMS · INTAKE · MANUFACTURINGVEDANTAGARWAL@JOINORTHOGONAL.COM● BUILDING THE CONTROL PLANE FOR HIGH-STAKES WORKVOICE · DOCUMENT · BROWSER · ERP OPERATORSPRODUCTION LOOPS — NOT DEMOSRESEARCH-LED · OPERATOR-FIRSTFOCUSED ON CARE · CLAIMS · INTAKE · MANUFACTURINGVEDANTAGARWAL@JOINORTHOGONAL.COM
SOTA agent systems for high-stakes workflows
Orthogonalby Stoic Labs
Orthogonal is a research-led AI lab building durable agent systems for regulated, emotional, and procedural work: calls, claims, faxes, approvals, audit trails, and the glue between every system the enterprise already bought. We are not a chatbot studio. We are a control plane for real operations.
No pilots.
We enter production or we don't enter.
No dashboards.
Software that watches itself isn't doing the work.
No model worship.
Frontier models are rented. Control is the product.
VoiceVisionBrowserERPFaxIntakeEvalsMemoryPolicyAuditEscalationReplayRecoveryVoiceVisionBrowserERPFaxIntakeEvalsMemoryPolicyAuditEscalationReplayRecoveryVoiceVisionBrowserERPFaxIntakeEvalsMemoryPolicyAuditEscalationReplayRecovery
/ MANIFESTOFiled by the operator class
01
— Frontier compute is a component, not a moat.
02
— Pay what you want once the system makes work better.
03
— An operating loop, not a license.
How we workWe build the layer that decides, acts,
We build the layer that decides, acts,
remembers, and proves itself.
Orthogonal does not ship demos. We embed inside the workflow until the system is accountable for outcomes. Four steps. No shortcuts.
STEP 01
Enter the workflow
We sit inside the actual work.
Calls, faxes, claims, approvals — we observe the operation as it runs, with the operators, until the bottleneck is named.
STEP 02
Instrument the bottleneck
Encode the policy. Stand up the evals.
We turn tribal rules into machine-readable policy. Every decision gets a measurable target before any model is allowed to act.
STEP 03
Ship the first loop
One accountable operating loop.
A single end-to-end loop in production with a human checkpoint, full audit trail, and a kill switch. No demo lanes.
STEP 04
Earn autonomy
Expand only after trust compounds.
Autonomy is granted by evidence, not by roadmap. Loops only widen as their evals, recovery, and escalation hold under load.
SOTA technical stackModels are rented.
Models are rented.
Operating advantage is engineered.
We compose frontier models with memory, permissions, evals, observability, recovery, browser control, voice, and domain policy. That's where enterprise AI becomes a company-building technology instead of a demo.
02 / Document intelligence↗
Faxes, PDFs, portals — all become state.
Extract, verify, cite, and route messy operational documents into systems of record with a complete audit trail.
03 / Browser & ERP operators↗
Agents that use the tools you already bought.
We coordinate browser actions, ERP writes, and human approvals without waiting for a perfect API to appear.
04 / Evals & control plane↗
SOTA is a component. Control is the product.
Every workflow gets evals, replay, escalation thresholds, permissions, and observability before it earns autonomy.
Operating standard
Production loops, not investor decks.
We're early and building in the open. This is the bar every loop clears before it earns autonomy — not a traction scoreboard.
Every decision
Traceable
Logged, attributable, and replayable by design.Every loop
Checkpointed
A human checkpoint and a kill switch before any autonomy.Every deployment
Production
We enter the real workflow — never a demo lane.Every model
A component
Frontier models are rented. Control is what we build.Solutions
Specific systems for work that cannot be abstracted away.
We don't sell horizontal AI. We pick narrow operating problems where the math is brutal — and ship the system that owns them.
CASE / 01●
Senior care
Voice operators for medication, transport, and family escalation.
Persistent agents that call seniors, confirm medication routines, hail rides, coordinate pharmacies, and alert caregivers when the signal is off.
Adherencethe outcome we own
CASE / 02●
Health insurance
Claims, denials, and the payment integrity gap.
Audit-grade loops for intake, coding, payer rules, appeal packets, and the revenue leakage hiding inside black-box vendors.
Payment integritythe problem we own
CASE / 03●
Durable medical
From fax intake to clean, compliant order.
Document intelligence for faxes, authorizations, patient records, Medicare requirements, and the exception routing nobody documents.
Fax → clean orderthe loop we own
CASE / 04●
Manufacturing
Parts, specs, naming, and the conventions only humans know.
Validation systems that encode tribal conventions, catch bad data during design, and approve routine decisions in real time.
Design-time validationthe loop we own
The refusal
We don't build the theater.
Six categories of product we will not ship — under any contract, at any margin.
R-01Chatbots— Theater dressed as product
R-02Co-pilots— Liability with a paperclip
R-03Dashboards— Watching is not working
R-04Pilots— Production or nothing
R-05Demo theater— Pretty videos, broken loops
R-06Per-seat SaaS— Renting users to themselves
Human-centric research lab
AI should be connective tissue. Not a subscription tax.
We study the enterprise until the workflow disappears. Compensation should track value: pay what you want after the system makes work better. The lab publishes its operating theories — and the failures that came with them.
Build where software stopped.
We take five new operators per quarter. Bring an unsolved workflow, the system that should own it, and the patience to do it properly.