José Ignacio Concha Araya l0rdbarcsacs

AI Engineer & Tech Lead · CTO @Cerberus Soluciones · Sr. Tech Lead @Open Source Integrators
B.Sc. Physics, Universidad de Chile · Santiago, Chile

~15 ms GPU inference · 84 instruments traded · 20+ agents in production · 5.3k commits/year

Portfolio ▸ jose.cerberus.cl · Company ▸ cerberus.cl · LinkedIn ▸ jose-ignacio-concha-araya

$ whoami

I build AI systems that survive contact with production — not demos. Ensembles that price copper in under 15 ms, a Rust HFT engine that retrains on GPU every 10 seconds, and multi-agent compliance pipelines that replace weeks of manual regulatory review for Chilean financial institutions. Physicist by training, systems engineer by trade: I lead the team, design the architecture, and write the production code.

Most of my work lives in private repositories. The panels below are generated daily from the GitHub API and include that private activity in aggregate — the code stays closed, the volume doesn't.

Systems I've shipped

AEON-COPPER Protocol

Real-time copper price prediction for trading desks. A five-model deep-learning ensemble over 80+ features — news sentiment, HMM regime detection, conformal prediction intervals — answering in under 15 ms on GPU.

TensorFlow 2.21 PyTorch 2.11 Rust ONNX SpacetimeDB FastAPI CUDA

CMF Compliance Agent

Multi-agent regulatory review for Chilean banks. Five specialised verifiers (capital ratios, cybersecurity, FinTech law) orchestrated through LangGraph over 11 containers. Replaces weeks of manual document review.

Amazon Bedrock LangGraph Qdrant PostgreSQL 17 Redis 8.2 Docker

Cerberus HFT Engine

High-frequency trading engine tracking 84 Chilean and international instruments. GPU LSTM via burn-cuda fused with Kalman-filter and Ornstein-Uhlenbeck estimators. Retrains every 10 seconds. Written in Rust for the latency floor.

Rust 1.94 burn-cuda SpacetimeDB v2 Astro Lightweight Charts

AutoResearch

An autonomous agent that runs its own ML experiments — writes the code, trains, evaluates, iterates. ~12 experiments/hour, ~100 overnight, unattended, on a single consumer GPU.

PyTorch 2.11 CUDA Muon + AdamW

Architecture — Cerberus HFT Engine

flowchart LR
  S["Scraper · Rust<br/>84 instruments"] --> DB[("SpacetimeDB<br/>WASM module<br/>14 tables · 14 reducers")]
  DB --> I["Inference · Rust<br/>burn-cuda"]
  I --> M1["LSTM<br/>2-layer, h=512"]
  I --> M2["Kalman filter"]
  I --> M3["Ornstein-Uhlenbeck"]
  I --> M4["Heuristic ensemble"]
  M1 & M2 & M3 & M4 --> C{"Auto-calibration<br/>60s eval / 30s recalibrate"}
  C --> DB
  DB --> F["Astro 5 + React 19<br/>real-time dashboard"]

Architecture — CMF Compliance Agent

flowchart TD
  A["API Gateway · FastAPI + JWT"] --> O["Orchestrator · LangGraph + Celery"]
  O --> V1["Capital ratios verifier"]
  O --> V2["Cybersecurity verifier"]
  O --> V3["FinTech law verifier"]
  O --> V4["Operational risk verifier"]
  O --> V5["Reporting verifier"]
  V1 & V2 & V3 & V4 & V5 --> B["Amazon Bedrock<br/>Claude + Titan Embeddings"]
  O --> Q[("Qdrant<br/>semantic search")]
  O --> P[("PostgreSQL 15")]
  O --> S3[("MinIO S3")]
  O --> R[("Redis")]

Live telemetry

Regenerated daily from the GitHub API, private repositories included in aggregate. 98.3% of my commits and pull requests land in private repositories — 5,134 against 87 public, across 30 closed repos. The code stays closed, the volume does not.

The rig

Everything above is built, trained and benchmarked on hardware I own and tune myself.

Node	Spec	Role
CERBERUS-MAINFRAME	Ryzen 7 7700 · RTX 3070 8 GB · 32 GB DDR5-6400 · 4 TB NVMe · Arch Linux + Hyprland	ML training, GPU inference, daily driver
CERBERUS-HPC	IBM X6 3850 · 48 vCPU · 512 GB RAM	Local LLM cluster — vLLM + RAG + hybrid web search
CERBERUS-MOBILE	ThinkPad T14 · i7-1165G7 · 32 GB · Arch Linux	Remote development

CUDA 13/12.x cuDNN eBPF ZFS SLURM Kubernetes Docker Swarm Terraform Ansible

Writing

Long-form essays on AI governance, ontological auditing and infrastructure. Published in Spanish at jose.cerberus.cl.

Mythos, Roko y la zona ciega: el test epistémico que le hice a HAL 9000-CS · _2026-05-06
Autoresearch mar14: Cuando HAL Optimizó un GPT Mientras Dormías · _2026-03-14
Génesis: Cuando las Máquinas Comenzaron a Hablarme · _2026-02-02
La Emergencia Silenciosa: Por Qué Su IA Corporativa Necesita Auditoría Ontológica · _2025-07-17

Enter the mainframe

▸ l0rdbarcsacs.github.io — this account rendered in WebGL: the contribution calendar extruded into a city, the language distribution, and the systems above in orbit. Static by construction, data baked from the API at deploy time. Source.

▸ jose.cerberus.cl — a CRT-mainframe WebOS in Astro and raw WebGL2: a living neural-network background, fourteen desktop apps, and a plain-text mode.

_{Regenerated Oct 4, 2026, 9:20 a.m. · Santiago, Chile · every panel on this page is produced by a workflow in this repository}

Navigation Menu