Mike.
Goolean's own LLM. Never downloadable. All processing and data operations run locally on our appliance, in your environment.
Mike is a large language model stack developed by Goolean for document-heavy, regulated work: a domain-tuned vision-language model, per-firm adapters, and a governed runtime — typed task contracts, a default-deny policy engine, a hash-chained evidence ledger and an evaluation harness. It comes pre-installed on an appliance we deliver to your building. Every document is processed there.
- Raw data never leaves the boxRemote backends refuse raw input at preflight. Mike Cloud sees synthetic or de-identified samples only — never a firm's documents.
- Observation-only outputsAny model output referencing a token not on the page is rejected outright. No confidence scores, no rectangles, no invented numbers.
- Safety proven by tests608 automated tests; every guard has a mutation test that proves removing it turns the suite red.
- Signed releases, on-box acceptanceUpdates ship as signed manifests and are gated by an acceptance run against that firm's adapter before activation. Rollback is the previous manifest.
First vertical: debt-collection law firms — intake, approval, redaction, packet assembly, e-filing and post-filing, as a 37-task catalog. Reference appliance: a Mac with 128 GB unified memory; NVIDIA supported.
Architecture
Mike runs in two places. Your data lives in one.
Inference runs on the appliance inside your network. Mike Cloud exists to keep the appliance current — signed model releases, knowledge packs, fleet health — over an outbound-only channel the appliance opens. Mike Cloud never receives a document, a page, a token, or a firm's adapter weights. That boundary is enforced in code and covered by tests, not by a policy PDF.
- Domain-tuned VLM + your firm's adapter
- OCR, classification, redaction, extraction
- Policy engine · evidence ledger
- Signed model releases + update manifests
- Knowledge packs (licensed / synthetic data only)
- Fleet health from enum-coded telemetry
How we engineer
Rules we can prove, not values we claim.
These are lifted from the operating manual of our own LLM runtime. Each one is enforced by a test in CI — the code shown is representative of the real thing.
def test_foreign_token_is_rejected():
out = observe(page, tokens)
assert not references_unknown_token(out) # BLOCKINGProvenance
Where Mike comes from
Anonymised where the client asked for it. Every outcome listed is one we can substantiate.
Join the alpha.
We're onboarding a small number of firms and partners for the alpha. Tell us your page volume and environment; we'll size the appliance we lease you.