AI and automation / Jan 2026 to Present

Sellavie AI

Businesses lose revenue when social enquiries are answered late, business context is fragmented, and follow-through depends on manual coordination. Automating replies alone is not enough: the platform must preserve tenant boundaries, understand catalogue and policy context, hand conversations to people, and keep commercial actions predictable.

I started Sellavie as an independent product and built it across channel ingestion, tenant-aware state, AI orchestration, catalogue knowledge, human takeover, commerce, subscriptions, reporting, web and mobile surfaces, compliance, deployment and production support. Generated language remains separate from deterministic actions that change orders, payments, access or subscription state.

Role: Designer and engineer
For: Independent project (Sellavie)
Stage: Maintained

Walkthrough

Watch it run

Product walkthrough / 1 min 44 s / Silent, captioned on screen

Signed in to a local build with Sellavie's sample workspace: a grounded reply, an order and invoice created from the chat, a bank-transfer claim that pauses the AI, and a person resolving the handoff.

My part

What I built

Translated the product idea into its technical and operational architecture.
Built backend services, the business workspace, administrative surface, mobile application and shared product workflows.
Implemented inbound social-event handling, tenant-context resolution, conversation state, guarded AI responses and outbound delivery.
Developed catalogue, business knowledge, bookings, customers, handoffs, orders, invoices, notifications, analytics and team collaboration.
Built subscription lifecycle, plan limits, renewal recovery, promotional access, usage reporting and administrative controls.
Own provider integration, deployment, failure investigation, production hardening, runbooks and staged modularisation.

Decisions

The choices that shaped it

AI writes language; domain services perform actions

A model can interpret intent and draft a response, but order, payment, billing, account and integration changes pass through deterministic services with independent validation and authorisation.

Business context comes before intelligence

The connected business, authorised account, catalogue, settings and conversation state are resolved before an AI request is built. Generation is never used as a substitute for tenancy.

Human takeover is a durable workflow

Handoff includes pause, acknowledgement, assignment, notes, follow-up expectations and resolution so the business can operate the conversation after automation stops.

Provider failure is expected

Stable internal contracts, idempotency, bounded retries, fallbacks and operational visibility prevent one external service failure from corrupting the wider workflow.

What has been shown to work

Validated onboarding, account verification, channel setup, catalogue configuration and live conversation handling.
Exercised event verification, duplicate delivery, business routing, provider failure, fallback response and outbound delivery paths.
Tested purchase, variant, shipping, booking, invoice, handoff and manual-reply behaviour around AI conversation.
Validated subscription checkout, renewal recovery, expiry, downgrade, promotional access and plan-limit enforcement.
Checked 188 frontend API calls against 295 backend routes and crawled 22 workspace routes for accessibility.
Repaired a webhook deduplication regression so provider retries no longer risk duplicate AI replies.

Current stage and limits

Sellavie moved into maintenance mode in October 2026. The site is still online and accepts sign-ups, but active feature development has paused.
It is an independent project. No revenue or customer figures are claimed for it.

Technical detailOpen the system view

SYSTEM ARCHITECTURE

Tenant-aware conversational commerce boundary

Social messages enter through verified platform callbacks, resolve to business-scoped context, pass through guarded AI orchestration and reach deterministic commerce or human workflows without granting the model authority over business state.

actorCustomer

Instagram, Facebook or WhatsApp message

externalPlatform webhooks

Signature, event and channel context

serviceTenant router

Business, role, catalogue and conversation state

workerAI orchestration

Grounding, provider fallback and proposed response

controlAction boundary

Guardrails, handoff and deterministic domain rules

dataOperational state

Conversations, orders, billing, teams and audit

01CustomermessagePlatform webhooks
02Platform webhooksverified eventTenant router
03Tenant routergrounded contextAI orchestration
04AI orchestrationproposed text / intentAction boundary
05Action boundaryauthorised transitionOperational state
06Operational statereply / handoff / outcomeCustomer

Security and privacy

Inbound platform events are verified before downstream processing.
Business data and integrations are scoped to tenant and role context.
Web, native, administrative and provider paths use separate trust boundaries.
Sensitive configuration is protected and omitted from ordinary client responses.

Built with

Production backendWeb dashboardMobile appTenant-scoped dataConversation automationAI response orchestrationCommerce workflowsOperational deploymentAccount controlsConfiguration protection

Links

Live site Technical case study

More work

Gazelle's KitchenA WhatsApp business gatekeeper that qualifies genuine orders, enforces business and safety rules, contains off-topic or abusive conversations, and only surfaces the small set of cases that genuinely need the owner.Pilot AI Automation Control PlaneA runnable multi-tenant control plane for AI-assisted business workflows, combining FastAPI, Next.js, PostgreSQL, Redis, n8n, provider routing, human approval, idempotency, replay, cancellation, audit traces, and deterministic local demonstrations.Maintained Webhook RelayWebhook reliability tool for receiving, verifying, inspecting, replaying, and forwarding events with bounded retries and a live inspector UI.Maintained

Loading portfolio

Sellavie AI