pro
Brandon Grutkowski
Local-first AI engineer • ex-Amazon finance • Real products
New to Contra
Brandon is ready for their next project!
Local-first AI isn't a niche anymore. It's how I build everything.
Here's the full system architecture behind Aurora, my glass-box quantitative intelligence engine. 82 commits, 599 tests passing, 24+ analytical methods. No cloud. No API keys. No telemetry. Every piece runs on your hardware.
Let me walk through what's actually happening in this diagram.
The Substrate Layer
Aurora isn't a wrapper around an LLM. The core engine is a multi-stage pipeline: your data hits a preflight layer (schema validation, missingness detection, irregular-sampling checks) before any analysis runs. If the data has problems, you know before a single method fires.
24+ Research-Grade Methods, Not Vibes
The analytical engine runs Isolation Forest, Hampel z-scores, HMM Baum-Welch, Granger causality, persistent homology, SINDy (sparse identification of nonlinear dynamics), Gaussian processes, mutual information, VAR, DTW, BOCPD, Robust PCA, EMD, Kalman filtering, spectral entropy, and more. Each method either produces a cited finding or explicitly reports why it skipped (no time axis, negative values, cross-sectional data). No silent failures. Ever.
The Glass-Box Contract
Every finding is a typed, structured object: method, severity, threshold, evidence, citation. When Aurora says +448.6σ with p < 0E+0, that's a Hampel z-score on a specific row you can re-run yourself. The 0 fabricated chip isn't marketing. It's a contractual counter audited live on every run.
Knowledge-Grounded Synthesis
The "What This Means" narrative cites seed:* entries from a local knowledge bank. Newton (1701), Pierson & Moskowitz (1964), Torrence & Compo (1998), NIST, NOAA NDBC. Real papers. Real institutions. No invented citations.
Two Surfaces, One Engine
Aurora Copilot is the human-facing studio: six analytical lenses (Overview, Anomalies, Regimes, Motifs, Forecast, Physics), spacetime worldlines, phase-space projections, causal do-calculus. Aurora Cortex is the machine-facing API: an MCP server (7 tools, path-allowlisted, 2MB output cap), a Python SDK, Decision Contracts that fire webhooks to Slack/Discord/email when findings match programmable predicates, and streaming mode with Kafka + Postgres CDC connectors.
Fully Local, Fully Packaged
The desktop app (Tauri 2 shell + PyInstaller-bundled backend) is a double-click install. No Python, no venv, no terminal. Drop a CSV on the window and watch it analyze. The Docker path is one command: docker compose up. BYO-LLM with 5 pluggable backends (Ollama, Anthropic, OpenAI, Gemini, OpenAI-compatible), or run with no LLM at all. Aurora computes. It doesn't need to guess.
I built this because I spent years in Amazon finance watching decisions get made on black-box outputs nobody could trace. Cloud LLMs guess. Aurora computes. That's the whole thesis.
Open source. Apache 2.0. 100% free.
What's stopping you from running your AI stack locally?
1
42
Horizon: Local AI Quantitative Investment Engine (Powered by Aurora).
0
6
JobVend: Mobile Application for General Contractors to manage, and grow their business in real time. Real time Quote development, payment processing, work assignments to crew etc.
1
12
Fantasy Studio: AI Video Director and creator with real 3D rendering Software. 100% Free and Local.
0
14
Aurora: Local Glass Box Quantitative AI Intelligence Pipeline. 100% free and local
1
24