Independent software engineerJoão Pessoa, Brazil  ·  Remote worldwide

PABLOCOSTA.

Python / Backend / Applied AIPortfolio © 2026
SOFTWARE THAT LEAVES THE DEMO.

Python backends.
Applied AI.
Reliable automation.

I'm Pablo Costa, a Python backend and applied AI engineer based in João Pessoa, Brazil. I build APIs, workflow automations, and AI integrations for real business operations.

01   Python backend
02   Process automation
03   Applied AI
04   Infrastructure
01 / SELECTED WORKReal projects. Documented decisions.

Less presentation.
More evidence.

A few examples of software I've built, with the problems, tradeoffs, and implementation details that matter.

01 / CONVERSATIONAL SYSTEMS
D
DentBotWhatsApp assistant · example flow
Hi! I'd like to know how your appointments work.
I can help with that. Would you like to book a consultation?

PROJECT / 001   —   AUTOMATION

DentBot↗

An AI customer-service product with a dental-clinic deployment in production, connecting WhatsApp conversations, business workflows and human operators.

OPENAI · WHATSAPP · N8N · DOCKER · POSTGRESQL
Built and operated since December 2025. WhatsApp, Chatwoot, n8n and LLM APIs connect to PostgreSQL/Redis-backed state and human operators. Structured model outputs support intent and handoff decisions; permissions, routing and external side effects stay under deterministic software control. Message debouncing consolidates fragmented inputs and helps prevent competing responses. DentBot Core is a separate Python/FastAPI modernization focused on typed contracts, domain rules and testability. The illustration contains no customer data.
02 / DATA PIPELINES
COURT NOTICE
→
RECORDSTATE
#001parsed
#002parsed
#003review

PROJECT / 002   —   DATA ENGINEERING

Court Data
Pipeline↗

A Python pipeline delivered to a São Paulo law firm, turning TRT2/PJe court communications into filtered, deduplicated Excel reports for review.

PYTHON · SQLITE · EXCEL · AUTOMATED TESTING

Validated with 93 automated tests. Filtering reduced the review surface by more than 96% on representative full-day datasets; this is not a measured time saving.

Delivered April–June 2026. Explainable rule-based filtering reduced the review surface by more than 96% on representative full-day datasets. The system was validated with 93 automated tests; these figures describe this project and its tested datasets, not a guarantee for other workloads. Implementation covers defensive parsing, API ingestion, SQLite persistence, Excel generation, retries and Windows packaging for non-technical users. Documentation covers API limitations, false-negative risks and operational behavior.
03 / DOCUMENT INTELLIGENCE

The answer is grounded in retrieved passages instead of relying on a model's memory alone.

SOURCE: DOCUMENT / SECTION 2[1] ↗

PROJECT / 003   —   APPLIED AI

RAG
Interactive↗

A project exploring retrieval-augmented generation and how source evidence makes model answers easier to inspect.

PYTHON · RAG · EVALUATION · RETRIEVAL
02 / THE LABA second way to explore the work
FOR THE CURIOUS AND THE TECHNICAL

Don't just read.
Try it.

Explore structured project data, inspect API responses, and see what an interactive engineering portfolio can become.

InterfaceInteractive playground
Data sourceRésumé + project notes
Current stateLocal prototype
Portfolio API SIMULATED DEMO / NO NETWORK
V1
02 / Response 200 OK

            
Simulated demo: responses are generated in your browser, without a real API. Job matching uses keywords, not AI. Skills reflect the supplied résumé; keyword matches do not verify proficiency or assess every requirement.
03 / ABOUT & APPROACHA little about the person behind the work

Practical by
design.

I'm a Python and applied AI engineer based in Brazil. Since February 2026, I've worked independently on backend integrations and AI automation for clients and my own products, from business rules and testing to deployment and maintenance.

Backend & operations

Python, FastAPI and typed contracts. Self-hosted Ubuntu/Docker infrastructure with Traefik, TLS, monitoring, encrypted off-site backups and tested restores.

Infrastructure case study ↗

Model evaluation

AI Trainer / Model Evaluator at Outlier since April 2026 and micro1 since August 2026. Structured review of instruction following, consistency, quality and naturalness, with documented failure patterns.

Education & languages

B.Tech in Internet Systems, UNIESP — completed July 2026. Native Portuguese and fluent professional English, following four years living and working in Australia.

Further training: IBM Generative AI for Software Developers Specialization; AWS Cloud Practitioner Essentials course; IBM RAG and Agentic AI course.

04 / CONTACTLet's talk about the actual problem