SmartLic — Public Procurement Intelligence Platform

Live: smartlic.tech · Production v0.5 · Paid trials · Stripe live billing

---

What This System Does

SmartLic ingests Brazil's public procurement data from 3 government APIs (PNCP, ComprasGov, PCP v2), normalizes it into a unified DataLake, runs full-text search in Portuguese against 3.5M+ records, classifies opportunities into 20 industry sectors using a hybrid keyword + LLM pipeline, scores viability across 4 factors, and delivers results through a web app with alerts, exports, and billing enforcement.

Technically: a 3-layer data architecture (ETL ingestion → PostgreSQL full-text search pipeline → multi-tier cache) exposed through a FastAPI backend with 187 endpoints, a Next.js 16 frontend with 10k+ programmatic SEO pages, Stripe billing with atomic quota enforcement, ARQ background workers for summaries and exports, and a Prometheus + Sentry + OpenTelemetry observability stack.

For a detailed engineering narrative, see CASE STUDY.

---

Technical Overview

Architecture

flowchart LR
    subgraph Sources["Data Sources"]
        A[PNCP API]
        B[ComprasGov v3]
        C[PCP v2]
    end

    subgraph ETL["Ingestion (ARQ + Redis)"]
        D[Daily ETL<br/>27 UFs × 6 modalities]
    end

    subgraph DataLake["DataLake (Supabase PostgreSQL 17)"]
        F[Tenders<br/>1.5M rows · 400d retention]
        G[Contracts<br/>2M+ rows · historical]
    end

    subgraph AI["Classification Pipeline"]
        H[Keyword Density<br/>deterministic filter ~80%]
        I[LLM Arbiter<br/>GPT-4.1-nano · ambiguous cases]
        J[Viability Scorer<br/>4-factor assessment]
    end

    subgraph API["API Layer (FastAPI)"]
        K[Search Pipeline<br/>7-stage state machine]
    end

    subgraph Delivery["Delivery"]
        L[Web App]
        M[Email Alerts]
        N[Excel · PDF · Sheets]
        O[10k+ SEO Pages]
    end

    A --> D
    B --> D
    C --> D
    D --> F
    D --> G
    F --> H
    F --> K
    G --> J
    H --> I
    I --> K
    J --> K
    K --> L
    K --> M
    K --> N
    K --> O

Loading

Key Metrics

Metric	Value	Evidence
DataLake records	3.5M+ (1.5M tenders + 2M contracts)	Measured from Supabase pncp_raw_bids + pncp_supplier_contracts row counts, 2026-06
Full-text search latency	< 100ms p95	search_datalake RPC execution time, Sentry Performance p95 window
Test suite	5,131+ passing, 0 failures	CI result on latest main (.github/workflows/backend-tests.yml)
AI sector classification precision	≥ 85%	Validated against 15 labeled samples/sector, benchmark in tests/test_llm_arbiter_benchmark.py
AI sector classification recall	≥ 70%	Same benchmark; 20 sectors × 15 samples = 300-label evaluation set
SEO pages indexed	10,000+	Google Search Console indexed pages estimate + ISR revalidate=3600
API endpoints	187	OpenAPI schema auto-generated from FastAPI route decorators
PostgreSQL migrations	183+	supabase/migrations/ directory, applied via CI auto-deploy
Daily ingestion volume	~10,000 tenders	PNCP API daily publication rate, 27 UFs × 6 modalities
Billing	Stripe live, 9 plans	12 webhook events with idempotency, atomic quota via RPC

Stack

Layer	Technology
Backend	FastAPI 0.136, Python 3.12, Pydantic 2.13, httpx 0.28
Frontend	Next.js 16, React 18.3, TypeScript 5.9, Tailwind CSS 3.4
Database	Supabase (PostgreSQL 17), 48+ tables, RLS on all tables
Cache	Redis (L1 InMemoryCache 4h → L2 Redis 4h → L3 Supabase 24h)
Queue	ARQ 0.26+ (Python async job queue backed by Redis)
LLM	OpenAI GPT-4.1-nano (classification + executive summaries)
Billing	Stripe 11.4 (Checkout, webhooks, customer portal)
Email	Resend (transactional, domain smartlic.tech)
Observability	Prometheus + OpenTelemetry + Sentry + Mixpanel
Infrastructure	Railway (web + worker), Supabase Cloud, Redis (Upstash/Railway)

Testing

5,131+ backend tests + 2,681+ frontend tests + 60 E2E (Playwright). CI gate on every push.

Tests protect:

• Search pipeline — 7-stage state machine contracts, time budget invariants, dedup logic
• LLM responses — structured output parsing, hallucination checks, fallback on malformed JSON
• Billing — webhook signature verification, plan enforcement, quota atomicity, grace periods
• Auth — JWT validation (3 strategies), RLS enforcement, role checks
• Cache — SWR staleness windows, L1/L2/L3 fallthrough, key invalidation
• Ingestion — content hash dedup, checkpoint resume, 400-day retention enforcement
• API contracts — OpenAPI snapshot diff (openapi_schema.diff.json), Pydantic schema canary
• Security — pip-audit in CI, CodeQL analysis, secret scanning

---

Market Context

The Problem

Brazil's government procurement market moves $500B+/year. Most suppliers discover opportunities through fragmented portals, outdated PDFs, regional newsletters, and WhatsApp groups.

PNCP (the official national procurement portal) publishes ~10,000 tenders per day across 5,000+ agencies. Without classification, a supplier must manually scan hundreds of irrelevant listings to find the few that match their sector.

Why Now

• Lei 14.133/2021 mandated structured digital procurement data for the first time
• PNCP public API (launched 2023) made programmatic ingestion viable
• LLM inference costs dropped to fractions of a cent per classification — economically viable at scale

Business Model

SaaS, 14-day free trial, no credit card required.

Plan	BRL/mo	USD/mo (approx)
Pro (monthly)	R$ 397	~$80
Pro (annual)	R$ 297	~$60
Consultoria (monthly)	R$ 997	~$200
Consultoria (annual)	R$ 797	~$160

Traction

• Paid trials active — Stripe billing live
• Organic inbound via 10k+ Google-indexed SEO pages
• DataLake compounding daily (3.5M+ records, growing)

For detailed metrics, request the investor data room: tiago.sasaki@confenge.com.br

---

Team

Tiago Sasaki — Solo Technical Founder / Applied AI Engineer GitHub · tiago.sasaki@confenge.com.br

Solo technical founder. Built the full stack: DataLake architecture, AI classification pipeline, 187 API endpoints, Stripe billing integration, 10k+ programmatic SEO engine, observability.

CONFENGE Avaliações e Inteligência Artificial LTDA — CNPJ 52.407.089/0001-09.

---

Documentation

For Engineers (evaluating the system)

Document	Purpose
CASE STUDY	Engineering narrative: problem, architecture, decisions, incidents, lessons
AI Classification Pipeline	Hybrid keyword + LLM pipeline: anti-hallucination, cost control, structured output
System Architecture	Full module map, ERD, C4 diagrams, ADRs
Operational Reliability	SLO targets, incident history, MTTR, error budget
Backend README	FastAPI app structure, module map, test commands, env setup
Deployment	Railway, Supabase, environment variables, CI/CD

For Product / Business

Document	Purpose
PRD	Full product specification
Roadmap	Backlog and status
CHANGELOG	Version history

---

Operational Status

Signal	Target	How Measured
API latency	p95 < 2s	Sentry Performance
Uptime	> 99.5%	/health/ready probe
MTTR	< 30min	Time from Sentry alert to confirmed fix

• Sentry: https://confenge.sentry.io/projects/smartlic-backend/
• Health probe: https://api.smartlic.tech/health/ready
• Runbook: _reversa_sdd/operational-reliability-2026-05.md

---

License

© 2024–2026 CONFENGE AVALIAÇÕES E INTELIGÊNCIA ARTIFICIAL LTDA — All rights reserved.

Proprietary software. Contact: tiago.sasaki@confenge.com.br

---

Tags: govtech · b2g-saas · pncp · comprasgov · public-procurement · llm-classification · fastapi · nextjs · supabase · brazil · latam