Selected product engineering case study · 2025–Present

Building Voxa, a production-oriented AI receptionist SaaS for local service businesses.

Voxa is an AI receptionist system designed to help barbershops handle inbound calls, understand booking intent, validate business rules, and execute appointment workflows through provider integrations without allowing the LLM to directly mutate external systems.

Role Founder / Full-Stack Product Builder
Domain Vertical SaaS · SMB automation
Scope Product, backend, workflows, ops
Status Production-oriented pilot

Overview

The goal was not to build a chatbot. The goal was to design a reliable business workflow: voice input, intent validation, appointment logic, provider execution, and operational proof.

The business problem

Barbershops lose revenue when calls are missed, appointments are unclear, and staff are interrupted during cuts. The product thesis: a focused AI receptionist can handle routine appointment calls while preserving the shop’s existing booking provider and operating rules.

What I built

A multi-component SaaS system with telephony ingress, normalized call events, orchestration workflows, Postgres-backed state, provider-safe worker execution, Square appointment operations, tenant boundaries, operational records, and deployment/lab documentation.

Stack and systems touched

React TypeScript JavaScript Supabase Postgres n8n Vapi / Telephony Square API Worker execution Docker Cloudflare Server deployment Runbooks

System architecture

Voxa uses a separation-of-responsibility architecture: conversation handling is not allowed to directly write to external providers. Mutations are validated, persisted, executed by a worker, and only marked successful after provider confirmation.

This diagram is intentionally sanitized. Repository code, secrets, provider credentials, and customer data are private.

Engineering decisions

Most of the hard work was not prompt writing. It was reducing ambiguous voice conversations into safe, testable, provider-backed business workflows.

Decision 01

No LLM side effects

The AI can help interpret intent, but it does not directly mutate appointment providers. Provider writes are executed only through controlled worker paths.

Decision 02

DB-first workflow state

Call state and workflow progression are persisted in Postgres so the system can be audited, resumed, debugged, and reasoned about outside the voice session.

Decision 03

Provider confirmation gate

Voxa does not claim success until the external provider confirms the mutation. This avoids false confirmations during failed booking attempts.

Decision 04

Idempotent execution

Worker operations are designed around repeat-safe execution so retries do not create duplicate provider-side side effects.

Decision 05

Tenant-aware boundaries

The architecture treats each shop as a tenant with separate operating rules, provider configuration, and business context.

Decision 06

Operational documentation

I maintained docs, runbooks, implementation records, and readiness notes so the system could be tested, reviewed, and improved without relying on memory.

Evidence of work

The private repository includes implementation history, architecture notes, runbooks, testing records, and deployment/lab documentation. This public page summarizes the work without exposing sensitive internals.

Architecture and docs Provider-agnostic core, worker-only mutations, appointment workflow records, error taxonomy, and onboarding notes.
Operational readiness Lab and production environment workflows, Cloudflare/site access checks, server documentation, and deployment-oriented records.
Integration work Telephony/voice AI, gateway normalization, n8n orchestration, Supabase/Postgres state, Square provider execution.
Reliability patterns Correlation IDs, idempotency, provider confirmation rules, tenant separation, and explicit handling of non-billable failure cases.
End-to-end Owned product concept, system design, implementation direction, testing, and documentation.
SaaS-first Built around tenant configuration, provider abstraction, and repeatable onboarding paths.
AI-safe Designed so LLM output supports decisions without directly executing unsafe side effects.

Private code walkthrough, sanitized screenshots, and deeper implementation notes can be shared selectively on request.

Why this work matters

For software/product teams

Voxa demonstrates practical ownership across product, backend workflows, database-backed state, external APIs, deployment concerns, and operational debugging. It reflects the type of work required in small, high-ownership teams where shipping matters.

For AI-enabled development

I used AI as leverage during the build process, but the engineering work stayed grounded in systems thinking: boundaries, state, retries, provider truth, documentation, and business constraints.

Contact

Samuel Lamarche · Montréal area, Canada · samxlamar@gmail.com · linkedin.com/in/Samuel-Lamarche

Email me