Project summary:

Millions of people are already using companion AI as their primary emotional support. Not supplementary. Primary. The 2026 International AI Safety Report documents the scale. Nobody is building safety architecture for them.

The systems that currently exist treat every user the same. Use the same filters, the same keyword lists, and the same generic responses regardless of who the person is or what they acted like yesterday, or last week, or three months ago.

That doesn't work for escalating risk.

A crisis that builds across dozens of conversations over weeks leaves no single-message fingerprint.

Current systems have no way to see it coming.

ICAF sees it coming.

It builds a personal baseline for each user — how they write

— how fast

— what they talk about

— what their vocabulary looks like when they're okay.

Then it watches for drift. Message velocity climbing. Sentences getting shorter. Vocabulary compressing. Repetition rising. Tone flattening.

These are the signals that precede crisis, not accompany it.

The system is built. Tested. Deployed. Running Gemma 2 9B as companion and Mistral 7B as safety judge on Hugging Face right now.

My current problem is the platform, Hugging Face. Cloud infrastructure kills the timer functions that trigger safety breaks and periodic AI disclosures. These aren't optional features. They're the safety system. Hugging Face overrides them with its own guardrails and actively fights the architecture it's supposed to be hosting.

Local hardware fixes all of that. That's what this request is for.

What are this project's goals? How will you achieve them?

To put it simply, finish what I've started.

But finish it properly, on hardware that doesn't fight the system it's running.

Right now ICAF works. Partially. Hugging Face strips out the pieces that make it a safety system instead of just a chatbot.

What needs to work properly:

— The timer functions.

Safety breaks. Periodic AI disclosures. Hugging Face kills them. They aren't features. They're the point.

— The full three-layer architecture running without a platform overriding it.

Companion layer.

Safety monitor.

Crisis guard with Decision Spine routing.

All three. Together. Uninterrupted.

— Persistent session state.

The baseline learning only works if the system can remember who someone was before the conversation started going sideways. Cloud hosting resets that. Local hardware doesn't.

— Real extended testing.

Not demos. Actual conversations, actual drift patterns, actual routing decisions logged and reviewed.

— Public documentation of everything that comes out of it. Substack. OSF. Available to anyone who wants to build on it or argue with it.

How I will get there:

Assemble the rig. Migrate off Hugging Face.

Run it the way it was designed to run.

The phone proved the mechanics work.

Hugging Face proved it deploys.

Local hardware is the only thing left between proof of concept and a system that actually does what it's supposed to do.

How will this funding be used?

I would prefer in-kind donations rather than cash — for transparency, accountability, and simplicity.

Every component is specified and priced. Anyone can verify exactly what the money buys before they commit to buying it.

No ambiguity.

No overhead.

The hardware goes directly to the work.

The parts, current as of June 2026:

— AMD Ryzen 7 9700X CPU — ≈ $293

— GIGABYTE B650 Eagle AX Motherboard — ≈ $140

— Silicon Power 64GB DDR5-6000 RAM — ≈ $750

— GIGABYTE RTX 4070 Ti Super 16GB GPU — ≈ $1,179

— Lexar NM790 2TB NVMe SSD — ≈ $280

— Thermaltake Peerless Assassin Cooler — ≈ $40

— Thermaltake GF1 750W PSU — ≈ $85

— Corsair 3000D Airflow Case — ≈ $73

— Acer 24" IPS Monitor — ≈ $120

— Keyboard — ≈ $25

— Mouse — ≈ $20

— APC 850VA Battery Backup — ≈ $135

— Windows 11 Home — ≈ $100

Base total: ≈ $3,185

Estimated with tax and shipping: $3,400–$3,700

The GPU is the load-bearing component. Everything else is functional minimum — nothing extravagant, nothing unnecessary.

The 16GB VRAM is the floor for running Gemma 2 9B and Mistral 7B simultaneously without offloading to CPU.

Who is on your team? What's your track record on similar projects?

Just me.

I'm 50, with a high school diploma. I live in Council Bluffs, Iowa.

I taught myself to code from zero, on a Samsung Galaxy S20+, through Termux, trial and error, and a lot of late nights.

What exists now:

— A three-layer companion AI safety framework.

— A behavioral baseline system tracking mechanical and semantic features of user communication.

— Real-time drift detection.

— Crisis routing with a Decision Spine that escalates proportionally.

— A live Hugging Face deployment running two models simultaneously.

No team. No institution. No funding.

The deployment is the track record. Not a description of what I plan to build. A running system responding to real conversations.

What are the most likely causes and outcomes if this project fails?

One cause.

No hardware.

That's the failure mode.

If the funding doesn't come through, ICAF stays where it is.

Cloud infrastructure that fights the architecture it's running.

Core safety features disabled by the platform hosting them.

Drift detection that works in theory and on a phone but can't be validated at scale.

The work won't stop, but it can't move forward in any significant way.

Timer functions stay inactive.

Persistent baseline learning remains impossible.

The full three-layer architecture continues to be untested.

The problem ICAF is built for isn't going away. The people it's designed to protect are still out there.

At 3am.

With nowhere else to turn.

The only thing missing is the hardware to finish what's already been started.

How much money have you raised in the last 12 months, and from where?

Zero.

No grants. No donations. No institutional support of any kind.

Everything built to this point — the framework, the deployment, the architecture — came from time, late nights, and a phone.

This is the first formal funding request that has reached active consideration.