Hi, I'm

Fernando Velasco Lozano

Senior Applied Data Scientist & AI Tech Lead

Over a decade turning mathematical ideas into production-grade AI systems — neuro-symbolic architectures, LLM orchestration, ASR, and Computer Vision pipelines. I lead AI squads through ambiguous problems. This is where I write up the model internals and research I can share publicly.

See projects Download CV Get in touch

Featured projects

Fante — RPG Game Orchestrator

Game-loop orchestrator for a voice-and-text RPG built for a 2-year-old. Coordinates LLM narration, game rules (via MCP), and player input through a ports-and-adapters architecture designed to grow with the player.

Python 3.12+

MCP

core-llm-bridge

Provider-agnostic LLM abstraction layer. One interface for Ollama, OpenAI, and Anthropic — sync, async, streaming — with conversation state, tool calling, and typed errors.

Python 3.12+

Pydantic v2

httpx

Copper — AI-maintained knowledge bases

An LLM Archivist reads raw sources — notes, PDFs, Obsidian vaults — and compiles a structured markdown wiki. No embeddings, no RAG; just folders, text, and LLM workflows (store / tap / polish). Supports multi-mind queries and bidirectional linking across knowledge bases.

Latest writing

From Bytes to Ideas: Language Modeling with Autoregressive U-Nets

A byte-level language modeling proposal where autoregressive U-Nets learn a hierarchical token representation — no tokenizer required.

HELMET: A Fresh Look at Context-Length Evaluation in LLMs

A benchmark system for evaluating LLMs across long contexts — summarization, long-doc QA, RAG, re-ranking, and generation with citations — beyond synthetic needle-in-a-haystack tests.

Titans: Neural Long-Term Memory for Enhanced Sequence Modeling

A paper that proposes test-time learnable long-term memory for LLMs, with a 'Surprise' metric that decides what in the context is worth remembering.

See all posts →

Current research pursuits

Much of my recent work is proprietary, but these are the frontiers I'm actively exploring:

Neuro-symbolic AI — hybrid architectures combining the reasoning and traceability of symbolic AI (knowledge graphs, rule-based logic) with the perceptual power of neural networks. Focused on solving the “black box” and hallucination problems in enterprise data insights.
LLM-as-Judge & evaluation frameworks — robust, automated metrics for model coherence and depth, moving beyond simple benchmarks to guarantee high-quality personalization in production.
Agentic orchestration — designing systems where specialized LLM agents collaborate on complex, multi-step tasks, optimizing for both accuracy and computational efficiency.

Awards & recognitions

Bronze Medal · Spanish NLP Hackathon 2022

Technical Lead and Lead Researcher. Novel T5-based architecture for gender-neutral Spanish text generation, outperforming baselines by 22% in human evaluation.