Case — Sol (Omnichannel IT Service Desk with AI and CRM)

Type: Freelance (client: Mirassol — internal IT support) Role: Full Stack Developer (full authorship: product, backend, frontend, AI, infrastructure and deployment) Status: In production (with a separate UAT environment) Stack: Python 3.11, FastAPI, async SQLAlchemy, Alembic, Angular 20 (standalone + Signals, strict TypeScript), PostgreSQL 16 + pgvector, Redis, Evolution API / Meta WhatsApp Cloud API, Microsoft Teams app, e-mail (messages), Gemini (google-genai) with a planned migration to Vertex AI, Docker Compose, deployment on a VPS + Google Cloud

Context and problem

The client's IT team received support requests in a scattered way (personal WhatsApp, e-mail, word of mouth), with no queue, no history and no prioritization. Support needed to be structured across the channels users already use — without forcing anyone to learn a new portal — and give management real visibility into service quality and volume.

Solution

An omnichannel service desk with AI, complete management and integrated CRM:

• Multichannel support: tickets handled via WhatsApp, Microsoft Teams, a web platform and e-mail (by messages) — same queue, same history, regardless of the incoming channel
• AI triage bot (Gemini): converses naturally with the user, understands the problem, classifies it and opens the ticket with the right data; free-form conversations are also supported; fallback to local rules when the LLM is unavailable
• Complete ticket management panel (Angular 20): queues, real-time conversations, per-ticket timeline, management dashboards — everything related to the ticket's lifecycle in one place
• Service and satisfaction gauge: AI-driven analysis with conversation sampling, generating a per-department assessment and actionable insights for management
• Automated reports on service operation and quality
• Integrated CRM module: campaigns and contact intake via WhatsApp on the same platform — service and relationship on the same data
• Semantic search with pgvector: history and knowledge indexed by embeddings to support triage

Architecture and technical decisions

• Async FastAPI with async SQLAlchemy and Alembic; quality locked in with pytest, Ruff and mypy
• Angular 20 standalone with Signals and pure SCSS, strict TypeScript — fine-grained reactivity in the real-time conversation panel
• Isolated channel layer: WhatsApp (Evolution API in the MVP, with a planned migration to the official Meta Cloud API), Teams app and e-mail plug into the same ticket domain — adding a channel doesn't rewrite the business logic
• PostgreSQL 16 + pgvector for operational data and embeddings in the same database (less infrastructure, single transactions); Redis for conversation state and cache
• Sampled quality analysis: instead of processing 100% of conversations with an LLM (prohibitive cost), statistical sampling generates the satisfaction gauge and per-department assessments at a controlled cost
• Two environments: production and UAT with separate deployment scripts (deploy.ps1, deploy-homolog.ps1, vps-setup*.sh) — a reproducible pipeline from local build → tar → VPS
• Security: external SQL Server integrated with a restricted user/role; secrets kept out of version control

Challenges and solutions

• Unifying heterogeneous channels (WhatsApp, Teams, e-mail, web) into a single ticket flow: a channel abstraction in the domain, with user identity resolved across channels
• AI that can't block operations: triage gracefully degrades to local rules when the LLM is unavailable
• Measuring quality without explosive LLM cost: gauge based on conversation sampling with per-department assessment
• Dates, time zones and message ordering in asynchronous conversations: handled as specific fix plans with tests
• Visual evolution: UI overhaul planned and executed with an in-house design system

Results and impact

• IT tickets now originate structured from any channel (WhatsApp, Teams, e-mail, web), with a unified queue and history [volume TO CONFIRM]
• Management gains quality visibility: satisfaction gauge, per-department assessments and automated reports — previously there was no measurement at all
• Integrated CRM turns the same support channel into a campaign and contact-capture channel
• Operation with a UAT environment, versioned migrations and strict typing/lint — maintainable by third parties