Lucas Pacheco · AI Research Scientist · Belém ⇄ Bern
Post-training, evaluating & shipping LLMs at national scale.
PhD in Computer Science (University of Bern) turned applied LLM engineer. Today I help build Brazil's sovereign AI platform at CPQD — research that ends up serving millions of citizens through gov.br.
223+citations
9h-index
15+papers
R$390Mprogram
Best PaperPerCom '24
Current role
Building sovereign AI for 200 million people.
2025 — present · CPQD, Brazil
AI Research Scientist on Project INSPIRE
Core researcher on Brazil's R$390M national AI-for-government program (FINEP/MCTI & Ministry of Management) — 94+ active workstreams across data, AI platforms, intelligent applications and cybersecurity.
Vale
Nitro Química
Gaia Greentech
TrinitySys
gov.br
- Co-designed the national AI governance methodology and project frameworks adopted by federal agencies
- Built technical training curricula on RAG, LLMs, LLMOps and knowledge bases for federal practitioners at every maturity level
- Technical consultant on the Unified Federal AI Platform — sovereign infrastructure serving millions of citizens through gov.br
- Applied AI research delivered to industry clients in mining, chemicals and greentech
Expertise
The full LLM lifecycle, end to end.
From preference optimization to production serving — with the evaluation rigor of a researcher and the deployment scars of an engineer.
post-training
LLM Post-Training
SFT, DPO, RLHF/RLAIF, reward modeling, and parameter-efficient methods (LoRA/QLoRA). PhD work in federated optimization transfers directly to distributed training — FSDP, DeepSpeed, mixed precision, scaling behavior.
evaluation
Evals & Red-Teaming
Agentic and multi-turn evaluation harnesses, tool-use simulation, LLM-as-judge, reward-hacking detection, regression suites and custom domain benchmarks. Red-teaming for prompt injection and jailbreak robustness.
inference
Inference & Serving
vLLM-based serving, quantization, batching strategies and latency/throughput tuning. Self-hosted, high-volume inference — down to the edge, where my research career started.
agents
Agentic & RAG Systems
Multi-agent orchestration with LangGraph and CrewAI, RAG pipelines and knowledge bases, tool/function calling, guardrails, memory and context management — production-deployed in government and industry.
PyTorchJAXHugging Face · PEFT/TRL
Federated LearningKubernetes · Docker
Vertex AIPythonC++Rust
Fortinet Certified (Cybersecurity)
PT native · EN fluent · DE intermediate
Trajectory
From the edge of the network to the center of national AI.
2025 — present
AI Research Scientist (Pesquisador Especialista I)
CPQD — Centro de Pesquisa e Desenvolvimento em Telecomunicações
National AI governance, sovereign platform consulting, LLM training programs, and applied agentic AI for federal government and industry.
2025
Postdoctoral Research Fellow
Federal University of Pará
Federated learning optimization and security for vehicular edge networks; international collaboration with the University of Bern; graduate student supervision.
2020 — 2025
PhD Researcher
University of Bern, Switzerland
Pioneered distributed aggregation for vehicular federated learning and layer-selection algorithms for partial FL (Best Paper, IEEE PerCom 2024). Published 15+ peer-reviewed papers with teams across Europe and South America.
PhD · 2020–2025
Computer Science
University of Bern & UFPA (co-tutelle) — thesis on federated learning in mobile edge computing
MSc · 2019–2020
Electrical Engineering
Federal University of Pará — mobility & cloud management in heterogeneous 5G networks
BEng · 2015–2019
Computer Engineering
Federal University of Pará — CNPq research scholarship recipient
Research
Peer-reviewed, cited, and awarded.
Skipping-based Handover Algorithm for Video Distribution over Ultra-dense VANET
A Distributed Aggregation Approach for Vehicular Federated Learning
Full list on Google Scholar and Lattes CV — additional papers accepted and under review (2025–2026).
Open source
Code that backs the claims.
fl-ns3
C++NS-3-based federated learning simulation framework — network-realistic FL experiments at scale.
FedNetSim
Python · TFTensorFlow-based federated learning simulation tool for rapid algorithm prototyping.
mysql2postgres
RustHigh-performance database conversion tool — because migrations shouldn't take a weekend.
mycelium_search
RustFull-text search engine built on Tantivy — fast indexing and retrieval from scratch.
92+ public repositories across federated learning, edge computing and security tooling → github.com/lsiddd