Products · REMIT
Four products that read, decide, check and review, and a proof record that proves each decision. Runs entirely inside your data centre.
Decide what you can prove. Defer what you cannot. Explain both.
Status. REMIT is in active development. The first solution, banking complaints and alerts, is being built and benchmarked on public regulator data; design-partner deployments run as forward-deployed engagements. Nothing here reports a production deployment.
FOLIO
Reads the documents
SIFT
Decides the clear cases
BOUNDS
Knows when to stop
BENCH
Puts a person on the rest
The four products
Reads the documents
Messages, voice notes, slips, claims, statements and policies become one case record, every field traced to the line it came from.
Decides the clear cases
A small model trained on your workflow chooses from your list of outcomes. It cannot invent an option and it cannot write free text.
Your list of outcomes
Same input, same answer. Nothing outside the list.
Knows when to stop
Measures, against a standard set on your own past cases, whether the system is sure enough to act. If not, it hands the case to a person. This is the core of REMIT.
Puts a person on the rest
Your reviewers see the evidence first, a suggestion clearly marked as a suggestion, and the same list of outcomes. They decide; the trail records it.
On every decision
A signed record of what was read, what was decided, by whom, and against which standard. A regulator can verify it with your public key, without us in the room.
What a proof record contains
Records are append-only in your own audit store and export as JSON or PDF. The public key and the record are enough.
Sector solutions
Each sector solution adds your documents, your list of outcomes, an accuracy standard measured on your past cases, and a compliance package for your regime. The platform underneath never changes.
The standard is measured, not assumed, so the underlying model can change and the proof survives.
Every case a person resolves becomes a labelled example for the next measurement round.
The record says which examples came from people and which did not, so an auditor can weigh them.
For model-risk and audit teams
BOUNDS uses split-conformal prediction over the workflow’s option set. Given n calibration cases held out from training, it emits a prediction set that contains the true option with probability at least 1 − α, for any model and any data distribution. The assumption is exchangeability between the calibration set and live cases; under drift the bound is void, so drift monitors and a scheduled re-calibration are part of the product, and a stale calibration routes every case to a person. Coverage is a frequency over cases, not a per-case promise.
The evidence base, with citationsTokens. Never. No metering on usage.
The base models. Open weights. You own the right to run them.
Hand-offs to people. Stopping when unsure is the product, not an overage.
Your data leaving your walls. Because it doesn’t.
Open inference infrastructure
Two open pieces that make a 7B-class model run on an Apple device, measured in public, including where they fall short.
Mixed-precision quantization, chosen by the Hessian
Sub-3-bit quantization usually stops a model being a model. FORGE is a post-training pipeline that decides per tensor rather than uniformly: it measures Hessian flatness to find the weights that cannot survive ternary, holds those at higher precision, and crushes the rest. On Falcon-H1-7B that is a 4.4x reduction to 3.48 GB with factual recall intact.
Read the detail
Metal prefill kernel for Mamba-2 hybrids
State-space models compute a sequential recurrence that defeats most mobile compilers. HELIX reassociates it into a chunked associative scan, so the work lands on Apple's matrix hardware instead of a serial loop, and keeps the carried state in threadgroup memory instead of spilling it to device memory on every chunk. It drops into llama.cpp behind one interception point and declines any shape it cannot serve, so it can never be a regression.
Read the detailThat is what REMIT is built for. The first paid step is small.