Case — LIS (Demo for Libbs' Public Bid: AI-Powered Customer Service Ticket Triage Assistant)

Type: AutoU (demo for a Libbs public bid — pharmaceutical industry; the bid is still in progress, Libbs is not an AutoU client) Role: Full Stack Developer — solo project (demo/MVP built alone) Status: Live demo (awaiting the bid's outcome) Stack: Python, FastAPI, Poetry, React 19, TypeScript, Vite, TanStack Query, Zustand, Recharts, PostgreSQL, Gemini (google-genai) with rule-based fallback, Docker Compose, deployment on Oracle Cloud VPS

Confidentiality note: demo built by AutoU to compete for a public bid — validate what can be made public before exposing name/details. Do not present it as a contracted project or an AutoU client while the bid is undecided.

Context and problem

The pharmaceutical company's customer service (SAC) receives heterogeneous requests (medication questions, pharmacovigilance, complaints) that must be manually triaged before reaching the right agent — with regulatory risk when an adverse-event report takes too long to be classified. AutoU decided to compete for Libbs' public bid for this problem, and the LIS demo was built as a functional technical proposal — the bid is still in progress.

Solution

The LIS portal with two surfaces in the same application:

Public chat (own domain, e.g. chat.domain.com): the end user talks to the LIS assistant, which triages the request with AI and opens the classified ticket; cases requiring a human are escalated to support
Internal portal (protected login): a "Conversations and support" queue, ticket view with timeline, dashboards (Recharts)

Architecture and technical decisions

Graceful AI degradation: without GOOGLE_GENAI_API_KEY, LIS triages via local rules — the MVP is never down due to LLM unavailability or cost
One application, two domains: the same frontend serves the internal portal and the public chat, routed by the VITE_CHAT_HOSTS env var — less infrastructure, the same API and database, natural escalation between channels
React 19 + TanStack Query + Zustand: server state and client state separated by the right tool; tests with Vitest + Testing Library
Registry-free deployment: a PowerShell script builds images locally, sends images.tar + compose + .env via SSH and brings it up on the VPS — a simple, reproducible pipeline for the MVP
Architecture documented in HTML diagrams and technical opinions versioned in the repo (presentation material for the bid)

Challenges and solutions

Regulated sector (pharma): triage with a clear conversation→ticket trail and human escalation for sensitive cases
Fidelity to the design: screens implemented from CSS specifications extracted from Figma (chat and conversations/timeline view)
MVP cost: Oracle Cloud Always Free + DuckDNS, with a clear path to definitive infrastructure

Results and impact

Automatic SAC ticket triage with AI-assisted classification and deterministic fallback
Public support channel and internal portal delivered as a single deployment
Functional demo presented in the bid at zero infrastructure cost — bid outcome still in progress

Discovery-projected indicators (business-case projections — not measurements)

Baseline gathered with the stakeholders: ~50 tickets/day (~18–20k/year), of which 80% are simple; current handling time of 7 min/ticket (5 min drafting the answer + 2 min logging), with a response SLA of up to 24h. Projections presented:

Automatic resolution: ~72% of tickets with no human intervention (~36/day) — the AI resolves ~90% of the simple ones; complex cases go to a human with pre-collected context
Average handling time: from 7 min to ~2 min (~71% reduction) — simple cases answered instantly, complex ones in ~4 min because the context arrives already structured
Correct escalation: ~90% of sensitive cases routed correctly to a human, trending to 95%+ with the feedback loop
Operational load: ~4.2h/day saved (~1,050h/year), estimated savings of ~R$ 26k/year — not counting quality and rework
Intent-classification accuracy: ~82% at go-live → ~90% after the 1st feedback cycle → ~95% in mature operation (differentiator: continuous knowledge-base governance and agent feedback)
NPS: estimated gain of +25 to +30 points vs. traditional SAC

Automation with judgment, not a 100% automated SAC: simple cases (résumé, discount, institutional questions) resolved instantly with approved answers; sensitive ones (adverse event, quality deviation, technical complaint) identified, collected and safely escalated — respecting pharmacovigilance, LGPD and consumer-protection law.