Franco
Fitte

AI engineer: I train and evaluate language models, build machine learning systems and write the software that puts them to work, from lab data to apps people actually use.

Born in 1999 Buenos AiresPhysics and Data Science · UBAEnglish C1

Email me Instagram

What I've done since 2025.

Early 2025 — present
Remote · USA

Turing

AI Trainer · LLM evaluation and model quality

Training language models such as Atlas, Gemini, Codex and Claude: rubrics, data labeling and test cases for RLHF pipelines.
Evaluating responses on reasoning and code-generation tasks; automated evaluation workflows to detect bias and improve quality.
Docker-based code for business use cases, model evaluation and training LoRA adapters.
Working fully in English, in a distributed international team.

LLM evaluationRLHFPrompt engineeringDockerLoRAGemini APIPython

2025 — 2026
Buenos Aires

IFIBYNE — CONICET

ML Engineer / Data Scientist · bee behavior

End-to-end machine learning pipelines over more than 1.5 billion data points from 300+ hours of high-frequency recordings: ingestion, features, training, inference and quality control.
Vision and deep learning: pose estimation with DeepLabCut (ResNet-50, 91.6% mAP), convolutional networks, transformers and latent embeddings with CEBRA to classify behavioral states.
Spectral analysis of time series (VMD, CWT, HHT), clustering, anomaly detection and persistent homology.
Real-time inference wired to Arduino and servo hardware to detect sleep and deliver stimuli automatically; parallel batch GPU processing.
Language models in the analysis: RAG with ChromaDB, local LLMs with Ollama, Gemini API and Vertex AI.

PyTorchCNNTransformersLLMDeepLabCutCEBRAHDF5GCPArduino

2026 — present

dilidili

Founder · design and development

A studio that designs and builds websites, custom apps and management systems for businesses and entrepreneurs in Argentina.

Node.jsPostgreSQLSupabaseCloudflare WorkersVercelGodotWebGL

Projects of my own.

Franco sitting with his laptop on his lap

Pulso

Real-time credit bureau shared between fintechs that detects loan stacking (3 or more lenders within 48 h). Privacy by design: the national ID never travels (HMAC-SHA256 and OPRF), immutable events and k-anonymous responses.

Node.js · Cryptography · private

Neural rendering distillation

Independent GPU research: a 285,652-crop dataset to distill a DLSS-style neural rendering teacher into CNN + linear-transformer students, multi-objective search with Optuna, INT8 quantization and kernel profiling on an RTX 3090.

PyTorch · CUDA · private

Smooth Motion RTX30

Enables NVIDIA Smooth Motion on RTX 30-series cards (Ampere SM86): a one-click Windows utility, no per-game DLSS-FG integration needed.

Bee behavior analysis

Automatic classification of behavioral states over 300+ hours of video: pose with DeepLabCut, spectral decomposition, CEBRA and real-time inference with Arduino.

Python

manuscript2latex

From scanned physics manuscripts to validated LaTeX: OCR, embeddings, semantic retrieval with ChromaDB and generation with Gemini and Vertex AI, with four layers of consistency checks.

OCR · RAG

spectroprint

A library that computes a 9-feature spectral-entropic fingerprint and matches it against a 65-entry atlas (44 mathematical families and 21 NIST spectra); it returns ranked hypotheses instead of a single answer.

Python

Resonant GPU

GPU performance is path-dependent: operational hysteresis, T* operations and the limits of the Von Neumann abstraction.

Python

Education and tools.

Franco thinking, with an atom in his thought bubble

2020 — in progress

B.Sc. in Physics

University of Buenos Aires

In progress

B.Sc. in Data Science

University of Buenos Aires

Languages

Native Spanish · English C1, with full professional experience at Turing

Tools

PythonSQLJavaScript / Node.jsBashPyTorchScikit-LearnOptunaRAGEmbeddingsOllamaChromaDBGemini APIVertex AIBigQueryGCPPostgreSQLSupabaseDockerGitPlaywrightGitHub Actions