Jorge Guberte

Principal AI Systems Architect

São Paulo, Brazil · Remote-friendly · hey@jorgeguberte.com

Software engineer with 12+ years engineering scalable platforms, distributed architectures, and high-throughput data pipelines. Current focus: applied AI systems architecture — moving foundation models from stateless chat into persistent, autonomous, context-aware products through memory substrates, tool orchestration, and generative UI.

12+

Years shipping software

Production

B2G AI platform live

2.8M ops/s

State-fork allocation

+50pp

Typed memory vs flat retrieval

Experience

  1. Co-Founder & Principal AI Architect · Hub Esfera

    Sep 2024 — Present — São Paulo, Brazil · Hybrid

    Architected Strata, a municipal intelligence platform for the public sector (B2G) with an embedded AI copilot — deployed on Cloud Run, serving production city profiles with live data pipelines and agent-driven decision support.

    • Designed multi-source public-data ingestion (BigQuery, PNCP, FNDE, INEP) into an embedded analytical DuckDB layer for ultra-low-latency agent queries.
    • Built a deterministic LangGraph agent with skill-based reasoning, sandboxed read-only analytical tools, and full LangSmith observability across education, procurement, and budget domains.
    • Engineered data-flywheel pipelines converting unstructured bureaucratic datasets (high-volume PDFs, official gazettes, Fundeb spreadsheets) into structured Parquet snapshots and agent-optimized Markdown.
    • Deployed scalable multi-municipality tenancy via environment-scoped configuration, Cloud SQL persistence, and Clerk-authenticated RBAC.
    • Prototyped generative-UI layers reducing cognitive load in high-density government dashboards.
  2. Lead Applied AI Architect · Independent AI Systems Research

    Jan 2023 — Present — Remote

    Ongoing applied R&D across agent memory, state management, generative UI, and multimodal AI — from experimental architectures to production-ready open-source libraries.

    • EPCG Memory Access Model: falsified flat retrieval as a universal memory interface; four typed operations beat it by +50pp in blind cross-domain validation (p < 10⁻⁷), with external AMA-Bench pilots audited end-to-end.
    • LoomDB: temporal memory as a formally bounded dissipative system (asymptotic saturation, sub-stochastic propagation) in Rust/WASM; hybrid retrieval policies optimized by evolutionary search instead of hand-tuned constants.
    • multiverse-js: Git-style branching state library — fork allocation at 2.8M ops/s, LWW merge optimized from O(N) scan to 44.9K ops/s.
    • CBR Audiotech / Prosodia: full async audio-intelligence pipeline (Next.js, BullMQ, Modal GPU, Cloudflare R2) with a prosodic adaptation engine for cross-cultural music generation; validated across two active music labels.
    • AYVU-Talian: decoder-only transformer from scratch on Talian — data curation, custom tokenizer, training, inference.
    • Pixie: embodied 3D web companion with Gemini Live API voice streaming, spatial tool execution, and amplitude-driven lip-sync.
  3. Senior Developer Analyst (Contract) · Bosch Digital

    2023 · 5 months — Campinas, Brazil · Remote

    Enterprise data integration and infrastructure modernization across complex SAP ecosystems, backend services, and distributed systems.

    • Architected impact-analysis protocols for critical data pipelines, presenting technical strategy directly to German headquarters.
  4. Software Engineering Foundation · SeePix · Martin Luz · EgypTeam · Interactus

    2006 — 2015 — Brazil

    Nine years of production engineering before AI: distributed delivery platforms, high-traffic campaign sites, cloud-native Python, and interactive media for global brands.

    • SeePix (2014–15): highly distributed digital-signage delivery with edge isolation and robust content sync (PHP/Symfony).
    • Martin Luz (2012–13): Tech Lead on high-traffic seasonal campaigns for Walmart and Sony; drove internal Agile and CI/CD adoption.
    • EgypTeam (2008–11): cloud-native Python on Google App Engine; foundational data pipelines for early RL environments (Unity3D).
    • Interactus (2006–07): interactive multimedia for major brands, including Nike's Joga3/JogaBonito global campaign.

Selected projects & R&D

Strata

A municipal intelligence platform with an embedded AI copilot.

Multi-source public-data ingestion (BigQuery, PNCP, FNDE, INEP) materialized into an embedded DuckDB analytical layer for ultra-low-latency agent queries. Deterministic LangGraph agent with skill-based reasoning, sandboxed read-only tools, and full LangSmith observability — serving real municipalities with agent-driven decision support.

LangGraph · DuckDB · Cloud Run · BigQuery · PostgreSQL

Multiverse

Git-style branching semantics for synchronous UI state.

A high-performance state management library addressing the Single-Timeline State Paradox: fork application state, explore alternate timelines, merge what works. Proxy lazy-evaluation plus a CRDT-lite patch engine make hundreds of concurrent isolated branches near-free.

TypeScript · Proxy lazy-eval · CRDT-lite patches

Pixie

An embodied 3D companion with bidirectional live voice.

Real-time bidirectional voice streaming over Gemini Live API WebSockets (16kHz/24kHz PCM), physics-based locomotion with Rapier, synchronous spatial tool execution, and 3D lip-sync driven by live audio amplitude. Observer mode builds pointer heatmaps for context-aware spatial commentary.

React · Three.js · Rapier · Gemini Live API

Ayvu-Talian

A language model trained from scratch for Talian.

A 5.39M-parameter decoder-only transformer trained for Talian, a Venetian-derived language spoken in southern Brazil. Full pipeline owned end-to-end: corpus curation, custom tokenizer, training, and inference — language modeling where the failure mode is cultural loss, not a lost benchmark point.

PyTorch · From-scratch transformer

EPCG

Typed memory interfaces for agents that must remember over months.

Emergent Personal Cognitive Graphs: a category-theoretic architecture for persistent cognition. Its first validated result: decomposing agent memory access into four typed operations beats flat top-k retrieval by 50+ percentage points in blind, cross-domain testing.

Rust · Category theory · AMA-Bench

Core skills

Languages
Python · TypeScript · JavaScript · Rust (systems integration)
AI systems
Agentic workflows & orchestration · Typed memory interfaces & temporal retrieval · Advanced RAG · Context engineering · Model Context Protocol (MCP) · Agent observability, benchmarks & behavioral QA
Frameworks
Mastra · FastAPI · Node.js · Next.js · React · LangChain/LangGraph · CrewAI · AI SDK · BullMQ
Data & storage
PostgreSQL · Redis · Qdrant · Memgraph (Cypher) · DuckDB · Cloudflare R2
Cloud & infra
GCP Vertex AI · AWS Bedrock · Azure OpenAI · Modal serverless GPU · Docker · GitHub Actions

Education

  • Extension: Complex Networks with Python and AI/LLM

    University of São Paulo (USP) · 2026

  • Data Structures & Algorithms Specialization

    UC San Diego (Coursera) · 2023

  • Technical Degree in Data Processing

    Colégio Marcel Proust · 2002

Languages

  • Portuguese — native
  • English — fluent; extensive distributed international team experience