AI · Holy Oracle
Unlocking Ancient Wisdom with AI
Archetypal symbol analysis and historical divination
// AI engineering
AI you can deploy, measure and trust.
We help companies put language models and coding agents to work safely: sandboxed, self-hosted where it matters, and tested against how they fail, not only how they demo.
- Good fit for
- Teams adopting coding agents or adding language models to a product
- Starts with
- A scoped proof of concept with success criteria written up front
- Tools
- Claude, OpenAI and Gemini APIs, vLLM, llama.cpp, Podman, Cloudflare Workers AI
// What we offer
Secure coding-agent setups
AI coding agents running in hardened, rootless containers, with a written threat model, a verification run on your machines and clear rules for review and production access. Based on our open-source agent-pod project.
Private and self-hosted models
Open-weight models served with vLLM or llama.cpp on your own or in-country infrastructure, monitored, and with documented failure behaviour, for data that should stay under your control.
Reliability and security reviews
We measure how your model or agent setup behaves when things go wrong: timeouts, lost nodes, recovery cost and gaps in isolation. You get measured findings, not a checklist.
Language models in your product
Summaries, extraction, classification and assistants built on the major model APIs, added to the web and mobile apps you already run.
Proof-of-concept sprints
A short experiment with success criteria agreed before we start, and a straight go or no-go at the end. A negative result is still a useful result.
// Research
We publish what we test.
Our AI work is backed by open experiments with predictions written before measuring, and results reported even when they are negative.
- github.comagent-pod Runs Claude Code in a hardened rootless Podman container, with a script that verifies the isolation on your own machine and a published threat model.
- github.comDistributed inference recovery What a single node failure costs an in-flight request on exo, vLLM and llama.cpp, measured across 107 trials.
- github.comDistributed inference resilience A simulation of pipeline-parallel model serving under node churn, calibrated against the Petals paper.
- naithanweigh.comAnatomy of an asymmetric persistence mechanism An essay on why cleaning up a compromised system starts with its anchor, and what that means for AI models that could persist across many devices.
Persistence is an asymmetry problem: the defender has to find everything, while the attacker only has to leave one thing behind.
AI · VOIYCE
AI-Powered Content Creation
Automating professional social media strategy
Platforms and systems · RizqX
Fintech platform for Islamic Finance
RizqX came to us with a web platform built in Plasmic that they wanted to keep. We integrated with it, took over their development and release pipeline, built their Flutter app for iOS and Android, and a year later moved them off Plasmic altogether.
// Questions
Questions
Will our data be used to train someone else's model?
We configure providers' business and zero-retention options where they exist. When data cannot leave your infrastructure at all, we host the model for you instead.
Where do we start?
With a call about the problem you want AI to solve. If a proof of concept makes sense, we scope it in writing, including what success looks like, before any work begins.
Is it safe to let AI agents write our code?
It can be, with isolation, review and limited access. That is exactly what our agent setups are built around: the agent never gets production credentials and every change is reviewed by a person.
// Next step
Have a system to build, fix or run?
Book a free 30-minute call. We'll tell you plainly whether we're the right fit, and if we're not, who might be.