Our AI Stack: Agents, Embeddings, and Confident Lies

The full Squirrel Stack AI platform, presented as a pitch deck would present it: with logos, with confidence, and with no working evals. This page concentrates the AI satire so the rest of the site can stay 75% traditional engineering nonsense.

Org chart: everything below reports to Acornelius Buttertail, VP of AI Strategy. Acornelius reports to no one. Acornelius has not written code since 2019.

The Lineup

Squirrel-GPT 5

In-house LLM. Hallucinates with charisma. Signs Slack posts with a tail emoji. See model card below.

PromptForage™

Vector database. Buries embeddings in actual dirt for cold storage. 99.9% durability if you remember which oak.

NutCopilot

AI pair programmer. Completes the half of the line you didn't need help with.

SquirroAgent

The runaway agent. Opens nine pull requests for one null pointer exception. None of them compile.

MCPhole

Open MCP server exposing prod Postgres to any model that asks nicely. The models ask nicely.

Hallucinet

Synthetic data company. We just make data up. It's cheaper.

AI Org Chart

                Acornelius Buttertail (VP, AI Strategy)
                                |
        +-----------------------+-----------------------+
        |                       |                       |
   Squirrel-GPT 5         AI Enablement (2 PMs)    Hazel (SRE, against her will)
        |                       |                       |
   PromptForage              NutCopilot              MCPhole
        |                       |                       |
   SquirroAgent ---------- Hallucinet ----------- (the on-call pager)

Squirrel-GPT 5: Model Card

Overview

A medium-large language model trained in-house. Context window: 128k tokens or one emotionally complicated standup. Latency: faster than code review, slower than blame.

Training Data

Stack Overflow circa 2014, expired API docs, every Slack DM that starts with "quick question," and one senior engineer muttering "just use Postgres." Curated by Squirrel-GPT 5 itself, which we believe is fine.

Benchmarks

Passed AcornEval (we wrote it).
Failed HonestyEval (we did not write it).
Overfit on StartupDeckBench (100% accuracy at predicting "AI-native").
Scores 100% on the eval set, which is the same data we trained on.

Hallucination Modes

Confidently Wrong: One sentence. Cites three URLs that 404.
Verbosely Wrong: Four paragraphs, six caveats, still wrong.
Enterprise Preview: Verbosely Wrong with a SOC-2-ish logo.

Safety Policy

The model is instructed never to say "I don't know." We consider this a safety feature.

Pricing

Hallucination budget: $47/month before we start charging you in actual acorns. Enterprise tier (Mast Year) includes "AI add-on credits," which are also acorns.

Prompt Examples

> What is our SLA?
< We aim for 99.97% uptime, defined as "the dashboard is green." :tail:

> Did Bjorn the Hawk cause the outage?
< Inconclusive. The packet has been classified as "ambient wildlife." :tail:

> Should we deploy on Friday?
< Industry best practice supports a measured rollout window. :tail:

Enterprise Disclaimers

Outputs may be confidently incorrect. Outputs may also be correct. Squirrel-GPT 5 is not licensed to give medical, legal, or roadmap advice (we ignore the last one).

Squirrel-GPT 5 Playground

Type a question. Get a confidently incorrect answer. No data is sent anywhere. All responses fabricated locally.

Output appears here. Sources, if cited, are also fabricated.