OpsBond — Operational AI Request a pilot
Capabilities

Built, tested and running.

What OpsBond does today, how it is built, and what comes next. Every working capability is shown in our recorded demo or verified in testing.

Capability map

WorkingBuilt & testedNextPlatform

OpsBond Clinic

CapabilityStatusWhat it does
Multilingual text conversationsWorkingContext across turns in Persian, English and Arabic; the recordings show English and Arabic chats.
Corrections without losing valid detailsWorking“Not Tuesday, Wednesday” updates the changed field and keeps the rest.
Voice messages in, spoken reply outWorkingVoice note → local speech recognition → the same Clinic core → spoken reply, in Persian today.
Answers grounded in approved clinic factsWorkingPrices, availability and provider details come from approved facts, never invented.
Medical & aesthetic information layerWorkingInformational answers within an approved scope; clinical decisions stay with clinicians.
Draft-first explicit confirmationWorkingNothing is registered until the customer confirms the exact draft.
Reception feed with a structured recordWorkingRequest ID and revision, who it is for, service, preferred time, contact, state and next action.
Revision historyWorkingEach amendment is a new revision; staff see the previous one alongside it.
Requests for another personWorkingA friend’s request is its own record, and changing one request leaves the others untouched.
Cancellations and status, stated truthfullyWorkingCancellations are confirmed and recorded for reception; status answers reflect what really happened.
Medical hold and urgent safety escalationWorkingReported symptoms pause routine booking for medical review; urgent signs trigger emergency guidance and an urgent case for staff.
Human handoffWorkingEvery request moves to a named owner on the clinic team, and is never presented as a confirmed appointment before staff confirm it.

Realtime voice

Realtime microphone conversationBuilt · executedLive microphone input and spoken playback, physically executed.
Five-language runtimeBuilt · executedPersian, English, Urdu, Hindi and Bengali; human-tested in Persian, Urdu and English.
Interruption handling, turn detection, language lockBuilt · executedLocal voice-activity and turn detection; conversational language selection and locking.
Telephony integration layerBuilt & testedProvider-neutral bridge, runtime-tested against the realtime system. Next: phone-line deployment.

Governed foundations

Operational data foundationBuilt & testedCanonical state, history and decisions held outside the model. Accepted and closed, September 2026.
Task foundationBuilt & testedTask creation, transitions and history with duplicate protection. Production functional check passed.
Core controlsBuilt & testedTenant context, role and authority binding, idempotency and replay safety, and a safe outbox.
Operated infrastructureRunningLinux and Docker stack with PostgreSQL, Redis and workflow workers, with backup and recovery checks.

Next and platform

First bounded clinic deploymentNextOne clinic, one workflow, one owner, one baseline, one decision.
More channelsNextWhatsApp, Instagram and web on the same governed core.
Calendar and clinic-system integrationNextBookings confirmed directly in the clinic’s own system.
Revenue recovery workflowsNextUnfinished enquiries, no-shows, recall and waitlist.
Outcome intelligenceNextBaseline against result for every workflow.
Private deployment · API · SDK · connectorsPlatformThe enterprise control layer that models and agents pass through before they act.
Technology

How it is built.

A production-minded stack, chosen so models can be replaced while state, rules and records stay put.

AI & languageLLM integration across multiple providers, structured extraction, multilingual conversation design (Persian, English, Arabic), local models for experimentation
VoicePipecat, Gemini Live native audio, Silero VAD, SmartTurn turn detection, Deepgram Nova-3, local speech recognition, text-to-speech
Data & backendPostgreSQL as the canonical store (tenant model, row-level security, append-only history), PgBouncer, Redis, n8n workflow orchestration, Python, REST APIs, Telegram Bot API
InfrastructureUbuntu Linux, Docker Compose, n8n main/worker, backup and recovery checks, secure remote administration
QualityRegression, property-based and metamorphic test suites; calibration and holdout evaluation; non-regression capability matrices; canary releases and rollback

Why this architecture matters as AI starts to act →

How we ship

Engineering discipline behind every release.

Version-bound

Every result is tied to an exact build, date and tested surface.

Observable

Inputs, state changes and outputs are recorded and inspectable.

Rollback-ready

Releases pass regression gates, with a defined stop and rollback path.

Human-owned

Requests, handoffs and consequential actions have an accountable owner.

Test notes · realtime voice
  • Real-microphone baseline re-verified on 14 Sep 2026: pinned environment confirmed, regression suite passed 8/8 and real microphone execution passed.
  • Persian language selection and lock passed; zero reconnects and no self-echo were observed in that run.
Test notes · Clinic and foundations
  • The current Clinic product core runs in the private Telegram runtime as of September 2026.
  • The operational data and decision foundation was accepted and closed in September 2026.
  • The task foundation passed a production functional check under trusted identity mapping, covering creation, idempotency, transitions and history. The check transaction was rolled back and left no persistent rows.
External academic reviews

Three computer-engineering academics assessed selected architecture materials, technical evidence and controlled demonstrations between June and August 2026. Read the reviews →

See it running

Watch the product in action.