Skip to content

01 / HELLO Marília, Brazil (GMT-3) · Remote · Open to international

Lucas Gonçalves

AI / ML Engineer

Applied machine learning, anchored in real-world agriculture.

View projects Download CV
3rd of 1,300+
A soil-reading app I built placed 3rd at FETEPS 2025, a state science fair; its paper was accepted at the ICPA 2026 precision-agriculture conference.
0.19°C
Average error of a temperature-forecasting pipeline across 211 countries, about 75% lower than the standard Prophet baseline.
205 tests
Coverage around 83% on a FAPESP-funded question-answering agent that scores its own confidence in every answer.

02 / ABOUT

About me

I'm an AI/ML Engineer bridging two worlds that rarely meet: the commercial reality of global agriculture and the engineering of modern machine learning systems.

I study Big Data for Agribusiness at Fatec Shunji Nishimura (graduating December 2026) and work at Jacto, one of the world's largest manufacturers of agricultural machinery, an experience that taught me to read a problem from the field before reaching for a model.

That perspective shapes what I build. My flagship project, VisioSoil, runs computer vision on-device (Flutter + TensorFlow Lite) for soil analysis; it placed 3rd at FETEPS 2025 among 1,300+ submissions and was accepted for poster presentation at ICPA/ConBAP 2026.

Beyond agriculture, I work across the modern AI stack: retrieval-augmented generation with LangGraph and Qdrant, hallucination detection via semantic entropy, and sentiment pipelines on RoBERTa. I also contribute to open source, including an accepted fix in AutoMQ.

I'm now focused on remote and international ML / AI Engineer roles, where real domain understanding meets production-grade engineering.

Off the keyboard: technical deep-dives (O'Reilly, Manning), Formula 1, and learning out loud in developer communities.

Lucas Gonçalves, AI / ML Engineer

Stack

  • RAG · LLM

    Qdrant · Ollama · FastAPI

  • PyTorch

    TensorFlow · scikit-learn

  • Python · Rust

    Polars · Docker · pytest

Based in
Marília, São Paulo, Brazil
Education
B.Tech in Big Data for Agribusiness, Fatec (2026)
Focus
LLM agents · RAG · computer vision · NLP · MLOps
Email
lucassg2015@gmail.com

03 / WHAT I DO

What I build

Four things I do well, each one backed by a real, shipped project in the case studies, not just a line on a slide.

  • Computer vision that runs offline

    Image models that run on the phone itself, in the field, with no signal and no trip to a server, for places where connectivity fails and a wrong call is costly.

  • Forecasting that hedges its bets

    Forecasting that does not stake everything on one model: a weighted blend of approaches, anomaly detection, and honest error numbers measured against a public baseline.

  • LLM agents you can act on

    Question-answering and agent systems that report their own uncertainty, so a person knows how far to trust each answer, running on open models with no dependence on a paid service.

  • The plumbing that keeps it alive

    The work that turns a notebook into a service people can rely on: packaging, automated builds, tests, and the unglamorous infrastructure that keeps a model running long after the demo is over.

04 / SKILLS

Skills & stack

Every tool here earns its place in one of the case studies below, not on a logo wall.

Languages

  • Python
  • Rust
  • Dart
  • SQL

ML & Data

  • PyTorch
  • TensorFlow
  • TensorFlow Lite
  • LightGBM
  • scikit-learn
  • statsmodels
  • HuggingFace
  • SHAP
  • pandas
  • Polars
  • NumPy

LLM & RAG

  • Qdrant
  • Ollama
  • RAG pipelines
  • Semantic-entropy verification
  • FastAPI

Mobile & Backend

  • Flutter
  • Riverpod
  • Drift
  • GoRouter
  • Uvicorn
  • Gradio

Cloud & Tooling

  • GCP
  • AWS
  • Docker
  • GitHub Actions
  • Git
  • pytest
  • mypy
  • ruff
  • uv

05 / CASE STUDIES

Case studies

Not a gallery of screenshots. Each one walks through the problem, the constraints, the call I made, and what I would revisit.

All repositories →

PROJECT 01 LLM Agents & RAG 2025-2026

205 tests

~83% coverage · FAPESP-funded RAG agent

SmartB100: A RAG Agent for Agronomy That Knows When to Doubt Itself

A research assistant for agronomy that scores how much to trust each answer, runs with no paid API, and works with no connection.

Result

A working system that ships a continuous 0 to 1 confidence score with every answer, runs fully offline with no paid API, and is held to a production bar: 205 automated tests at roughly 83 percent coverage, strict type checking and linting in the build pipeline, and a packaged deployment with health checks and log rotation. The full source, tests and evaluation pipeline are public in the repository.

  • Python 3.12
  • FastAPI
  • Qdrant
  • Ollama
  • Gradio
  • Docker
  • GitHub Actions
  • pytest / mypy --strict
Repo
Read the full case study
Context
Field technicians and farmers need answers that are buried in dense technical manuals. A normal chatbot will sometimes invent an answer and state it with the same confidence it shows when it is right, which is the worst possible failure when someone is about to act on it in the field. The system answers questions from a fixed set of trusted manuals (a technique called retrieval-augmented generation, where the model reads from real documents instead of guessing). It was built as a research project funded by FAPESP, the São Paulo state research foundation and one of the main public funders of science in Brazil.
Constraints
There was no budget for paid AI services, so the whole thing had to run on free, self-hosted models. It had to keep working without reliable internet, since that is where it is used. And every answer needed a trust signal that a non-expert could read at a glance, not a yes or no flag meant for a developer.
Decision
Instead of a simple right or wrong flag, I built a confidence detector based on a 2024 paper in the journal Nature (Farquhar et al.). It asks the model the same question several times, groups the answers that mean the same thing, and measures how much they disagree, turning that disagreement into a single confidence score from 0 to 1. Answers also adapt to whether the reader is a beginner or a specialist, and the system is built on FastAPI, Qdrant and Ollama, which are open tools for serving the model and searching the documents.
Alternatives considered
A plain binary hallucination flag was the quick option, but it tells the user nothing about how unsure the system is, so a shaky answer looks identical to a solid one. The other option was to lean on a paid hosted model with its own guardrails. I rejected that because it broke both the no-paid-service and offline requirements, and it would have tied a public research project to a recurring bill it cannot guarantee paying.
Retrospective
The confidence check costs more compute because it generates several answers per question, so accuracy and speed pull against each other. I deliberately put it behind a clean boundary so it can be split off and scaled on its own later, but for heavy traffic that trade-off would need a real rethink.

PROJECT 02 Computer Vision 2025-2026

3rd / 1,300+

FETEPS 2025 · ICPA 2026 paper accepted

VisioSoil: Reading Soil Texture From a Photo, Offline

Reads soil texture from a photo on the phone itself, tagged with its location, with no signal needed. Placed 3rd of more than 1,300 entries at a state science fair.

Result

The classifier placed 3rd out of more than 1,300 entries at FETEPS 2025, a large science and technology fair run by the São Paulo State technical school system, and the full research paper was accepted at ICPA 2026 and the 17th ConBAP, an international precision-agriculture conference. The app ships the complete offline loop of capture, classify and store, with an automated build that runs analysis, tests and a full app package.

  • Flutter
  • Dart
  • TensorFlow Lite
  • MobileNetV2
  • Riverpod
  • GoRouter
  • Drift/SQLite
  • geolocator
Repo
Read the full case study
Context
A soil's texture is its mix of sand, silt and clay, and it drives real decisions about irrigation and planting. Reading it normally means sending a sample to a lab or calling in a specialist, and on a large farm with no phone signal neither of those happens, so most fields are never assessed at all. VisioSoil classifies soil texture from a single photo, right on the phone.
Constraints
It had to work with no internet whatsoever, because that is exactly where it is needed. It had to fit and run on an ordinary phone rather than a server. And the reading had to be trustworthy, since a soil measurement is only useful if you also know exactly where it was taken.
Decision
I run the image model fully on the device using TensorFlow Lite, which is Google's tool for running models on phones, on a compact network called MobileNetV2 that is designed to be light enough for mobile hardware. It was trained in two stages for accuracy. The classification runs in a background thread so the screen never freezes, and every reading is saved on the phone together with its GPS location, behind a storage layer that does not care whether the data later syncs to a server.
Alternatives considered
The simpler path was a cloud service that does the classification on a server, but that fails the moment the signal drops, which is most of the time in the field. The other shortcut was to save readings straight into a fixed database. I put a storage abstraction in between instead, so adding optional cloud sync later does not mean rewriting the app.
Retrospective
The model that won the competition is not yet the one shipped inside the app. Wiring the production model in, and letting it read its own labels from a config file at startup, is the main open thread. The storage abstraction already leaves room for the remote sync a multi-device rollout would need.

PROJECT 03 NLP & LLMs 2025

42× faster

Rust preprocessing rewrite, byte-identical

Tweet Sentiment, Tuned for the Mess of Real Tweets

A sentiment pipeline built for messy real tweets, with the slow preprocessing rebuilt in Rust to run about 42 times faster.

Result

Out of the box, the model scores 0.71 macro F1 on a test set of 12,284 tweets. Macro F1 is an accuracy measure where 1.0 is perfect and each sentiment class counts equally, and this score is the bar the fine-tuned version has to clear. The Rust cleaning tool runs about 42 times faster than the Python version on 100,000 tweets, with its output verified to match the original exactly. Both the Python and Rust test suites pass, and the code is public.

  • Python
  • Rust
  • HuggingFace Transformers
  • RoBERTa
  • Polars
  • Rayon
  • scikit-learn
  • PyTorch
Repo
Read the full case study
Context
Sentiment analysis means automatically judging whether a piece of text is positive, negative or neutral. The catch is that most ready-made sentiment tools are trained on clean, formal writing and quietly fall apart on tweets, where slang, mentions, hashtags and emojis break the assumptions they were built on. This project builds a sentiment pipeline tuned for exactly that informal, noisy text.
Constraints
The cleaning of tweets had to be correct down to individual emojis, which are easy to corrupt when a program splits text the wrong way. The preprocessing had to be fast enough to handle a million or more tweets, not just a small sample. And any claim of improvement needed a measured starting point to beat, rather than a vague before and after.
Decision
I started from RoBERTa, a language model for understanding text, in a version already specialized for Twitter, and first measured its accuracy out of the box to set an honest baseline. Then, because the Python cleaning step is the bottleneck at large volumes, I rebuilt that step a second time in Rust, a language known for speed and safety, and proved it produces byte-for-byte the same output as the original Python.
Alternatives considered
Keeping all the preprocessing in Python was simpler and would have been fine for a demo, but it would not hold up at a million tweets, which is the entire point. A generic, non-Twitter sentiment model was the other obvious starting point, but I rejected it because it ignores precisely the slang and emoji patterns that carry the sentiment in tweets.
Retrospective
The fine-tuning run that should beat the 0.71 baseline is still ahead, along with batch processing for very large volumes and a small hosted demo. So the honest headline today is the measured baseline and the speed work, not a final accuracy number, and the project is labelled that way.

PROJECT 04 Data Engineering & MLOps 2025

0.19°C

avg error (RMSE) · ~75% below Prophet

Global Temperature Forecasting That Never Bets on One Model

Forecasts daily temperature across 211 countries to within about 0.19°C, using a blend of models so it never bets everything on one.

Result

The best single model reached 0.19°C average error (RMSE), about 75 percent lower than the Prophet baseline at 0.77°C. The blended version sits at 0.24°C, giving up a little accuracy in exchange for not depending on any single model. Unusual readings are caught two independent ways, and each forecast is explained with a method called SHAP that shows which inputs drove it. The test suite passes and the build is green; the full pipeline is public.

  • Python
  • LightGBM
  • ARIMA/SARIMA
  • Prophet
  • scikit-learn
  • SHAP
  • PyArrow/Parquet
  • pandas
Repo
Read the full case study
Context
Frost warnings, irrigation timing and energy demand all hang on short-term temperature forecasts. The trouble is that most forecasting pipelines stake everything on a single model and have nothing to fall back on when that model has a bad day. This pipeline forecasts daily temperature across 211 countries and is built so one weak model cannot sink the whole forecast.
Constraints
The data was large and messy: more than 133,000 daily readings, full of outliers and gaps. The result had to be explainable rather than a black box, so a person could see why a forecast came out the way it did. And any claim about accuracy had to be measured against a credible public baseline, not simply asserted.
Decision
Rather than pick one model and hope, I tested five different forecasting approaches against the same baseline and then combined the strongest ones into a weighted blend, giving more influence to the more accurate models. The baseline to beat was Prophet, an open-source forecasting tool from Meta. Accuracy is reported as RMSE, a standard error measure in degrees Celsius where a lower number is better.
Alternatives considered
Shipping the single best model on its own would have shown the lowest error on paper, but it leaves no cushion for the day that model misbehaves, so I traded a sliver of accuracy for a blend that does not depend on any one model. The classic statistical models, ARIMA and SARIMA, were tested too, but they trailed the modern models on this data and were left out of the final blend rather than carried along for show.
Retrospective
It was validated on a two-year window, so the honest next step is testing it on data outside that range before trusting it further. It also needs a serving layer so forecasts can be produced on a schedule or on demand rather than run by hand.

06 / EXPERIENCE

Experience & education

  1. Nov 2024 to Present Work

    After-Sales Intern · Jacto

    Jacto is a Brazilian manufacturer of agricultural machinery that sells in more than 100 countries and employs over 3,000 people, with after-sales operations at industrial scale. I work where the business side meets engineering.

    • Automated a previously manual data migration: extracted Russian dealership address records from the government address-classification system (KLADR), transliterated them Cyrillic-to-Latin, cleaned them by region, and loaded them into Salesforce, the platform the company runs its customer records on.
    • Designed a computer-vision system (now in internal review) that audits discarded parts end to end: it checks each photo meets the standard, reads the printed slip and cross-checks it against the system record, and recognizes the part — paired with an in-app camera that locks file names and stamps time and location so the evidence holds up.
    • Found 10+ usability problems as the business-to-developer bridge on an internal parts-return tool, validating rule changes across testing rounds — moving the workflow from manual entry to mostly review.
    • Audited 150+ main dealerships at home and abroad, reconciling the field-parts return flow and tracking millions of reais in parts movement, and processed write-offs through audits done both on site and remotely.
  2. 2024 to Dec 2026 Education

    B.Tech in Big Data for Agribusiness · Fatec Shunji Nishimura, Pompeia, São Paulo

    A public applied-technology degree. Coursework covers machine learning, artificial intelligence, data structures, databases (SQL and NoSQL), APIs and microservices, cloud architecture, and statistics.

    • Co-authored a paper accepted at ICPA 2026 and the 17th ConBAP, an international precision-agriculture conference (abstract #14064).
    • Placed 3rd out of more than 1,300 entries at FETEPS 2025, a large state science and technology fair.
    • Languages: Portuguese (native) · English (B2, upper-intermediate).

07 / CONTACT

Let's talk

I'm looking for a full-time AI/ML Engineer role on a distributed, remote team building production systems at scale. If that's what you're hiring for, send the form — it's the surest way to reach me and I usually reply within a day.

Prefer email?

lucassg2015@gmail.com

Currently — Marília, Brazil (GMT-3) · Remote · Open to international