Skip to content

Full stack developer · AI & LLMs · Potim, Brazil

Talles Amaral: Full stack engineer, from the data model to the last pixel.

I'm Talles Amaral, a Full Stack Developer at UNISAL with 3+ years shipping production systems — lately with LLMs. On my own time I build products like Lery, a voice-first AI English tutor, and Colete, an app for organizing weekly football games.

Currently
Full Stack Developer, UNISAL
Previously
Junior Developer, Editora Santuário
Focus
Web platforms, APIs, LLM systems
GMT−3

01/Selected work

A few systems I've designed and built end to end.

Lery AI cover: “The English tutor that talks back”, with the mascot and cards showing a Portuguese sentence, the bridge-mode level, the English reply, a score of 92 and the share of English spoken.

01Computer Engineering thesis2026

Lery AI

A screen-less, voice-first English tutor that runs on a Raspberry Pi.

Many learners freeze when they have to speak. A naive LLM tutor makes it worse: give it a long prompt and it answers an A1 student at B2 level. Lery treats the model as one component in a system that controls what it sees, when, and how its output is checked.

  • Layered context with explicit token budgets, plus just-in-time retrieval through tool calls instead of injecting full history every turn.
  • Compliance gate for beginner levels: typed violations are fed back to the tutor for one corrective retry before anything is spoken.
  • Every turn scored on four CEFR pillars using schema-constrained output, with eval fixtures and a synthetic student to test the pipeline.
  • OpenTelemetry traces showed two independent LLM checks running in series; running them in parallel removed a full call from the critical path.
Role
Design and development across the device voice loop, agent service, API and mobile app.
Stack
TypeScript · Fastify · Gemini · OpenTelemetry · PostgreSQL · Python · React Native
Postie cover in the style of its LinkedIn carousel: “From chaotic email to the right issue”, an airmail stamp with an envelope and “$0.00 license cost”, a priority scale (low, medium, high, critical) and the four steps: ingestion, AI triage, routing and deduplication.

02Personal project · Open source2026

Postie

An autonomous agent that turns chaotic support email into the right GitHub issue.

Triaging support email is invisible work: read the thread, open the screenshot, work out which repository the module belongs to, check whether an issue already exists, then file the card. Postie does all of it, end to end, and runs in the background in a container.

  • Multimodal triage: attached screenshots go to Gemini alongside the text, so the real error is read from the image and bugs are told apart from questions and out-of-scope mail.
  • Routing by declarative guardrails: each project in the catalog defines its scope and the actions the agent may take (create an issue, update priority, comment, auto-close).
  • Two-stage deduplication: lexical similarity with the project's own words filtered out, then the model as a semantic judge only when a close candidate exists, which keeps LLM calls cheap. Priority rises with repeat reports (3+ HIGH, 5+ CRITICAL) and never goes down.
  • Hexagonal architecture: the email provider (Microsoft Graph or IMAP), the model (Gemini or OpenAI) and the issue tracker sit behind ports and swap by configuration. The domain is pure TypeScript validated with Zod and covered by 19 Vitest tests.
Role
Design and development; open sourced with documentation in English and Portuguese.
Stack
TypeScript · Zod · Gemini · OpenAI · Microsoft Graph · GitHub API · Vitest · Docker
Colete landing page: headline, sign-in with Google and phones showing the live game screen.

03Personal product2026

Colete

A mobile-first web app for organizing weekly pickup football games.

Attendance, team draws, live scores, stats, rankings and awards usually live in a group chat and someone's memory. Colete puts the whole game night in one place, and makes the results shareable.

  • Statistics are derived from game events — goals, assists, awards — so there are no redundant counters to drift out of sync.
  • Authorization checked per group on every server action; private groups return 404 so their existence isn't revealed.
  • Invites use random tokens with revocation, expiry and usage limits.
  • Dynamic Open Graph images for public games, rankings and player cards.
Role
Product, design and full stack development.
Stack
Next.js 16 · React 19 · Server Actions · Prisma · PostgreSQL · Auth.js
NexusTrace cover: “Fraud hides in the connections”, with a graph of account communities, a high-risk account receiving transfers from 17 accounts, risk signals and a note from the AI analyst. Synthetic data.

04Integrator project · Team2025–26

NexusTrace

Fraud detection over a graph of financial transactions.

Fraud rings hide in relationships, not in single transactions. NexusTrace loads accounts and transfers into Neo4j, groups them into communities and scores every account by how money actually moves through the network.

  • Louvain community detection with Neo4j Graph Data Science, run on startup over synthetic account and transaction data.
  • Explainable risk scores from fan-in, volume z-score and cross-community behaviour, with configurable rules.
  • FastAPI service pushing anomaly events over WebSockets to a Next.js dashboard with an interactive force graph.
  • An AI analyst on GPT-4.1 that answers questions by calling the same risk API through tool calling.
Role
Built as a team integrator project at UNISAL; I maintain the public repository.
Stack
Python · FastAPI · Neo4j GDS · Next.js · WebSockets · OpenAI
Aresta Studio home page in dark mode: headline about turning complex ideas into real products, with an isometric illustration.

05Independent studio · Website2026

Aresta Studio

The website for my independent software studio, designed and built end to end.

A studio site has one job: make a business owner trust you enough to describe their problem. Aresta leads with problems instead of a list of services, shows real cases and turns the first contact into a structured brief.

  • Brief form as a Server Action with field validation, a honeypot and transactional emails built with React Email.
  • Isometric illustration system drawn in code, with light and dark themes and entrance motion that steps aside under reduced motion.
  • Dynamic Open Graph images and a route that renders Instagram carousels from the same components as the site.
Role
Brand, design and development.
Stack
Next.js · React · TypeScript · Motion · React Email

02/Experience

Building systems hundreds of people use every day.

  1. Apr 2025 — Present

    Lorena, Brazil

    Full Stack Developer

    UNISAL — Salesian University Center of São Paulo

    I work on the architecture, development and evolution of the institution's internal systems, from solution design to deployment.

    • SIGA, real-time queue system

      Queue platform with ticket kiosks, receipt printing, calling panels and operational metrics. Each step of a ticket's lifecycle — issued, called, transferred, finished, no-show — publishes an event pushed over Socket.IO to kiosks, panels and attendants.

    • Social Service platform

      Full stack platform in Fastify, TypeScript and Next.js (DDD, Clean Architecture) serving 678 students and handling 90 scholarship applications and 170 appointments a month, replacing manual work without growing the team.

    • Ombudsman modernization

      Migrated a legacy ASP.NET WebForms system to Next.js, with real-time chat and a file API with secure links.

    • Key management

      Replaced untracked spreadsheets with a Fastify + Vite app with permissions and a full audit trail — 350 keys and 312 checkouts a month.

    • AI training

      Taught the workshop “Artificial Intelligence and Prompt Engineering” to about 40 people during UNISAL's Employability Week.

  2. Dec 2022 — Apr 2025

    Aparecida, Brazil

    Junior Developer

    Editora Santuário — Publishing house

    Internal systems used every day across sales, HR and operations.

    • Sales CRM

      Cut contract generation from 10+ minutes to seconds for an operation with 5,678 clients and ~797 orders a month — a CRM in Next.js and .NET Core with Pagar.me integration and document automation.

    • LGPD-compliant digital signatures

      Digital signature platform with dynamic Markdown documents that issued 20,345 legally valid consent terms.

    • Help desk

      .NET and SQL Server ticketing system handling ~650 tickets a month, with departmental flows, prioritization and end-to-end traceability.

    • Intranet integrated with TOTVS RM

      Intranet portal connected to the TOTVS RM ERP, centralizing timesheets, shift management, announcements and real-time management dashboards.

    • Dynamic forms platform

      Customizable forms (Next.js, .NET Core) that gave HR and Workplace Safety autonomy over their own data collection.

Tools I reach for

Frontend & mobile

  • TypeScript
  • React
  • Next.js
  • React Native · Expo
  • Vite
  • Tailwind CSS

Backend

  • Node.js · Fastify
  • NestJS
  • Express
  • .NET Core
  • Python · FastAPI
  • Socket.IO

Data

  • PostgreSQL
  • SQL Server
  • MySQL
  • Neo4j
  • Redis
  • Prisma

AI engineering

  • Gemini · OpenAI · Claude
  • Context engineering
  • Tool calling
  • Structured output
  • RAG
  • Evals

Architecture

  • DDD
  • Clean Architecture
  • Hexagonal
  • Event-driven
  • SOLID

DevOps & quality

  • Docker
  • CI/CD
  • AWS
  • OpenTelemetry
  • Vitest · Jest
  • Playwright

03/About

I like software that holds up after the demo.

Talles Amaral

I'm a full stack developer from Potim, Brazil, with 3+ years building systems people rely on every day — first a CRM, help desk, intranet and digital signatures at Editora Santuário, and now academic and administrative platforms at UNISAL, where I own the work from solution design to deploy.

I care about the parts users never see: clear domain boundaries, authorization that holds up, performance I've actually measured, and tests where they matter. With AI, my focus is the engineering around the model — context, tool calling, structured output, evals and observability. I build, measure, and only then optimize.

Education

Bachelor's in Computer Engineering — UNISAL

Jan 2022 — ongoing

Portuguese (native) · English (B1/B2)

Languages

Courses

  • CS50HarvardIn progress
  • Node.jsRocketseat2025
  • React and Next.js(Certificate)Rocketseat2024
  • UI and UX Design — FigmaUdemy2023
  • Git and GitHubUdemy2022
  • SQL ServerWorkover Academy2022
  • Scrum FundamentalsWorkover Academy2022

04/FAQ

Quick answers about who I am and what I do.

Who is Talles Amaral?

Talles Amaral is a Full Stack Developer from Potim, Brazil. Since April 2025 he works at UNISAL, where he designs, builds and evolves the university's internal systems, and on his own time he builds products such as Lery, Colete and Aresta Studio.

Which technologies does he work with?

TypeScript, React, Next.js, React Native and Expo on the front end; Node.js (Fastify, NestJS), .NET Core and Python (FastAPI) on the back end; PostgreSQL, SQL Server, MySQL and Neo4j for data; Docker, CI/CD and AWS for delivery.

Does he work with AI and LLMs?

Yes. His AI work is about the engineering around the model: context engineering, tool calling, structured output, evals and observability. Lery, his Computer Engineering thesis, is a voice-first English tutor built this way. Postie, an open-source agent, applies it to routing and deduplicating support email, and NexusTrace pairs a graph-based risk engine with an AI analyst.

What has he built?

At UNISAL: SIGA, a real-time queue system, and a Social Service platform serving 678 students. Personally: Lery AI (voice-first English tutor), Postie (an open-source agent that triages support email into GitHub issues), Colete (app for organizing weekly football games) and Aresta Studio (his software studio's website). With a team: NexusTrace, fraud detection with graph analytics.

How much experience does he have?

Three-plus years in production. He was a Junior Developer at Editora Santuário from December 2022 to April 2025 (CRM, LGPD-compliant digital signatures, help desk and an intranet integrated with TOTVS RM) and has been a Full Stack Developer at UNISAL since April 2025.

Where did he study and which languages does he speak?

He is pursuing a Bachelor's in Computer Engineering at UNISAL (since January 2022). He speaks Portuguese (native) and English (B1/B2).

How can I contact him?

By email at contact@tallesamaral.dev. He is also on LinkedIn (linkedin.com/in/tallesamaral) and GitHub (github.com/tlsamaral).

05/Contact

Have a project, a role or a question?

Email is the fastest way to reach me. I'm also on GitHub and LinkedIn.