# Jorge Guberte — Principal AI Systems Architect

> Personal site and applied AI systems research lab: https://jorgeguberte.com

Jorge Guberte is a Principal AI Systems Architect and Co-Founder & CTO of Hub Esfera based in São Paulo, Brazil. With 12+ years shipping production software and distributed systems, his work focuses on applied AI systems architecture: agent orchestration, cognitive memory architectures, temporal retrieval, context engineering, and generative UI.

## Core Thesis
"AI systems should retain context, adapt over time, and remain useful beyond a single interaction."
Most AI products optimize for the single session and suffer total amnesia. Jorge Guberte designs and ships architectures that accumulate understanding, manage decay, maintain identity, and operate reliably in long-horizon production environments.

## Production Systems
- [Strata](https://jorgeguberte.com/systems/strata): Municipal intelligence platform (Hub Esfera · Production B2G). Embedded AI copilot on top of live public-data pipelines (BigQuery, PNCP, FNDE, INEP) materialized into an embedded DuckDB analytical layer. Deterministic LangGraph agent with skill-based reasoning, sandboxed read-only tools, and LangSmith observability. Live serving city profiles including Porto Alegre/RS.
- [Multiverse](https://jorgeguberte.com/systems/multiverse): Open-source state management library with Git-style branching semantics (`multiverse-js`). Fork allocation benchmarked at 2.8M ops/s (430ns); LWW merge at 44.9K ops/s inside a 60fps frame window. MIT license. GitHub: https://github.com/jorgeguberte/multiverse
- [Pixie](https://jorgeguberte.com/systems/pixie): Embodied 3D AI companion with bidirectional live voice streaming over Gemini Live API WebSockets (16kHz/24kHz PCM), Rapier physics locomotion, synchronous spatial tool execution, and audio-amplitude-driven lip-sync. Built with neurodivergent-first interaction constraints.
- [Ayvu-Talian](https://github.com/jorgeguberte/ayvu-talian-base): 5.39M-parameter decoder-only transformer trained from scratch for Talian (Venetian minority language spoken in southern Brazil). End-to-end tokenizer, corpus curation, PyTorch training.

## The Lab (Research Programs)
- [EPCG (Emergent Personal Cognitive Graphs)](https://jorgeguberte.com/lab/epcg): Category-theoretic architecture for persistent agents. Falsified flat top-k retrieval (`retrieve(query) -> top-k`) as universal memory interface. Memory Access Model v0 decomposes recall into four typed operations (`EpisodicAt`, `StateAt`, `StateHistory`, `SemanticRetrieve`), beating flat retrieval by +50 percentage points in blind cross-domain validation (N=200, p < 10⁻⁷). External AMA-Bench pilots: EPCG-T 72.5% vs BM25 70.0% and LongContext 50.0%.
- [LoomDB](https://jorgeguberte.com/lab/loomdb): Graph-based temporal memory engine modeled as a nonlinear dissipative dynamical system. Features asymptotic saturation, lazy decay with consolidated stability, sub-stochastic propagation (bounded energy), and evolutionary routing policies. Rust core compiled to WASM.
- [Y2K Sensory](https://jorgeguberte.com/lab/y2k-sensory): Maximalist aesthetic unified with neurodivergent cognitive accessibility via dual-layer CSS architecture and AttentionGuard presets.

## Links & Profiles
- Website: https://jorgeguberte.com
- Lab: https://jorgeguberte.com/lab
- Systems: https://jorgeguberte.com/systems
- Writing & Essays: https://jorgeguberte.com/writing
- CV: https://jorgeguberte.com/cv
- Work with me: https://jorgeguberte.com/work-with-me
- GitHub: https://github.com/jorgeguberte
- LinkedIn: https://www.linkedin.com/in/jorgeguberte
- ORCID: https://orcid.org/0009-0001-1198-1065
- Email: hey@jorgeguberte.com