Klenio Padilha
Independent AI Researcher — Mathematical Guarantees for Regulated AI
Portugal·+351 939 436 961 (WhatsApp)·klenioaraujo@gmail.com·linkedin.com/in/kleniopadilha·kleniopadilha.com·@kleniopadilha

Profile

Self-taught technology professional with 25+ years of experience spanning four decades of computing — from legacy systems (Clipper, COBOL) through Delphi, C++, Java, PHP, Linux and databases to modern AI — with projects delivered across Europe (France, Germany, Portugal). Now leading independent research on deterministic vector search and AI with mathematical proof guarantees for regulated sectors, validated at 100M-vector scale and published as 23 DOIs on Zenodo. Philosophy: mathematical guarantees, not probabilistic heuristics — every claim has a metric, every benchmark is honest.

Professional Experience
Independent AI Researcher — WINNEX Research Program (Madhava) 2024 – Present
Deterministic vector search, agent security, pre-patent architectures
  • Designed a deterministic vector-search engine replacing probabilistic heuristics (HNSW, IVF) with Cauchy-Schwarz upper bounds — every discarded document carries the mathematical proof it could not be in the top-K.
  • Validated at 100M vectors (BIGANN-100M): NDCG@10 = 1.000 with 0 bound violations; determinism and an auditable evidence chain for regulated markets.
  • Authored WINNEX — a deterministic, training-free inference framework composing a spectral tokenizer, a manifold projector and the bound engine, shipped as three PyPI packages (C++20 cores).
  • Applied the guarantee to agent security (Madhava-Sec): 98.1% reduction in LLM calls with 0 violations on 416 AgentHarm scenarios.
  • Built pre-patent architectures for regulated domains: RAI multi-agent orchestration (Ed25519 audit trail), Solon (legal AI) and the Tracer product family for government, legal, healthcare and finance (recall@10 = 1.0000, 0 violations).
  • Published 23 DOIs (Zenodo) and 17 public Kaggle notebooks; 6 bugs fixed publicly with full transparency.
AI Research & Machine Learning Engineer 2020 – 2024
Machine learning, embeddings, vector search
  • Entered AI in 2020: machine learning, embeddings and vector search — foundations of the Madhava research program.
  • Built spectral and manifold-based inference components with explicit, verifiable mathematics.
Software Engineer — Distributed Systems 2015 – 2020
Enterprise systems, integration, advanced SQL
  • Built and integrated enterprise systems; deepened software architecture, advanced SQL and infrastructure.
  • Contributed to production projects in France, Germany and Portugal.
Linux / Database Administrator and Systems Engineer 2010 – 2015
Linux administration, automation, deployment, relational databases
  • Consolidated Linux administration (automation, deployment) and relational database design and optimization.
Software Developer — System Languages 2005 – 2010
C++, Java, object-oriented programming, Linux servers
  • Developed systems in C++ and Java; first production experience with Linux as a server platform.
Software Developer — Desktop Applications 2000 – 2005
Delphi, data modeling, database administration
  • Migrated to Delphi, building desktop database applications; learned data modeling and database administration.
Software Developer — Legacy Systems Early 2000s
Clipper, COBOL
  • Started career as a programmer in Clipper and COBOL — the foundation of corporate systems of the era.
Core Skills
Programming: Clipper, COBOL, Delphi, C++, Java, PHP, Python  ·  Linux: systems & server administration  ·  Databases: modeling, administration, optimization  ·  AI / ML: vector search, machine learning, multi-agent systems, C++20 + AVX2  ·  Research: honest benchmarks, reproducibility (Kaggle/Zenodo/PyPI), EU AI Act compliance  ·  Soft skills: self-taught learning, adaptability across 4 decades of change
Selected Publications & Research Output
WINNEX UB Width Forensic Supplement — the engine orchestrates proof and heuristics, exposing per-document exclusion source; honest benchmark: recall@10 = 1.0000 in 5/6 scenarios, 0 violations. DOI 10.5281/zenodo.21954024 (2026)
WINNEX — Weighted Inference Neural Network Enhancement — deterministic, training-free inference framework (spectral tokenizer + manifold projector + bound engine). DOI 10.5281/zenodo.21925977 (2026)
Winnex Tracer Series v5 — mathematical-proof retrieval for regulated domains (GOV/JUS/MED/GAP). DOI 10.5281/zenodo.21826594 (2026)
Madhava — BigANN C++ (100M vectors) — NDCG@10 = 1.000, 0 bound violations. DOI 10.5281/zenodo.21166403 (2026)  ·  Madhava Cascade — founding article. DOI 10.5281/zenodo.20970487 (2026)  ·  Winnex RAI Architecture — 21-page pre-patent, multi-agent orchestration with Ed25519 audit trail. DOI 10.5281/zenodo.21292595 (2026)
PsiQRH — Quaternionic-Harmonic Framework — reformulating transformer attention via quaternion algebra. DOI 10.5281/zenodo.17171112 (2025)  ·  Observing the performance of distribution systems with embedded generators — peer-reviewed, European Transactions on Electrical Power (2004)
Education & Languages
Education: Self-taught — continuous independent learning across four decades of computing  ·  Languages: Portuguese (native), English (basic), Spanish (basic)