Lucas Biagetti

Lucas Biagetti

AI/ML Engineer & Researcher

Building production AI systems. Asking hard questions about evaluation, interpretability, and reasoning.

GitHub LinkedIn Hugging Face

Research Interests

Questions I keep returning to

📊 Evaluation Without Ground Truth

How do you distinguish real learning from memorization when there's no correct answer? I faced this building poetry generators and market simulation systems.

PoetGPT PhantomLabs

🔍 Interpretability for High-Stakes Decisions

When a model assists investment or education decisions, the "why" matters as much as the output. Building systems with full traceability and auditability.

Xylence RAG Systems

🧠 Reasoning & Chain-of-Thought

Can explicit reasoning compensate for model size? Exploring this question through experiments with CoCoNut and years of coaching mathematical problem-solving.

CoCoNut Math Olympiad

Research Projects

Open-source experiments and demos

PoetGPT

Spanish poetry generation using EleutherAI's Pythia model. Built a dataset of ~5,000 poems and a custom tokenizer with special tokens for line/stanza structure. The core challenge: how do you evaluate quality in creative generation?

Demo → Code →

BorgesBot

Literary style transfer fine-tuned to imitate Jorge Luis Borges. Created my own dataset from his complete works and built the training pipeline from scratch. An experiment in capturing stylistic patterns.

Demo → Code →

CoCoNut + GPT-2 Experiment

Implementation of the CoCoNut paper applied to GPT-2. Exploring whether explicit chain-of-thought reasoning can compensate for model size—trading parameters for structured thinking.

Code →

Experience

Building AI systems in production

Xylence 2025 – Present

Head of AI/ML Engineering

Leading the predictive engine that turns unstructured inputs (pitch decks, interviews, web signals) into decision-grade structured data for VCs.

  • Built technical foundation from scratch: no GitHub to product in beta in <3 months
  • Architecture with full traceability: every output is auditable and explainable
  • Evaluation harnesses for retrieval quality, latency, hallucination rate
  • Leading a team of 5 engineers
Santex Mar – Sep 2025

Sr. AI Engineer

Rapid prototyping and production deployment of AI solutions.

  • Voice agents with Whisper + ElevenLabs (<500ms p95 latency)
  • "Prospect cloner" system analyzing interview patterns to create AI personas
  • ~2 proofs of concept per week; 30% conversion to product
NeuralSoft 2021 – 2025

AI Specialist

Built production RAG systems deployed to 800+ enterprise clients.

  • Custom RAG system achieving 93% accuracy across ~2000 queries
  • Domain-agnostic anomaly detection in pure Java with Fourier transforms
  • Time series forecasting and sentiment analysis systems
  • Reduced AI integration costs by 80%

Background

Competitive Mathematics

  • IMO Selection (2012)
  • Iberoamerican Mathematical Olympiad (2012)
  • Tournament of Towns (2008, 2012, 2013)
  • Coached olympiad students, 2013–2016

Education

  • Master's in Pedagogical Sciences — 110/110 with honors, Università della Calabria, Italy
  • Bachelor's & Teaching Degree in Educational Sciences — 8.45/10, Universidad Nacional de Rosario
  • Systems Analyst — 66% completed, UNR

Academic Research

  • 4 publications in educational anthropology (2016–2018)
  • Including a chapter with ISBN
  • Presentations at UBA, UNC, UNR, CONICET
  • Researcher at Núcleo Antropológico Educativo (UNR, 2014–present)