Sovereign open-source AI · financial services

Protect your data. Own your AI.

We help financial services and lending teams automate document-based workflows with AI. We deliver systems that you own end-to-end inside your own infrastructure. Data and decisions remain in house.

Book a 30-min fit check Read the reference architecture →

Information security posture: data never leaves your account · SOC 2 Type II (in progress) · GLBA- and FINRA-aligned controls · HIPAA controls where medical data is in scope · Trust & security →

Earlier track record · before the pivot

Equifax EVO Payments Splunk Intel Pax8 Chewy

What we do

Three productized service offerings with fixed prices.

Three options that work well in sequence beginning with an evaluation, then a deployed platform, and finally a workflow pilot.

01 — EVALUATION

OSS Model Evaluation & Selection

A written recommendation on which open-weight model to trust for a defined workload, with the eval results, harness code, and a reliable cost projection.

2–3 weeks · $50–75K→
02 — PLATFORM

Sovereign RAG-as-a-Platform

A retrieval platform deployed inside your own cloud account: Terraform modules, dashboards, runbook, eval harness, and a full knowledge and ownership transfer.

6–8 weeks · $150–300K→
03 — PILOT

Document Workflow Pilot

Extraction, classification, and agentic routing for a real document process using the best fit open weight model by operating price and eval performance.

8–12 weeks · $250–500K→

Proof

We benchmark against relevant use cases.

We're confident that open-source, open-weight models are the future. There are still real trade-offs and decisions to make today, so we run the benchmarks, publish the results, and open-source the harness.

Enterprise OSS LLM Index · Q2 2026
Open-weight models, self-hosted in-VPC on regulated-document workloads:
Model Extract Retrieval $/M out
Qwen3-VL-235B95.0%29%$1.30
Llama-4-Scout84.6%25%$1.01
GLM-4.7—37%$1.27
Kimi-K2.5—38%$2.18
DeepSeek-V3.2—35%$3.56
Extraction = F1 on degraded scans (vision models); retrieval under one fixed judge. Each within noise of the same weights on Bedrock. $/M out = US$ per million output tokens (serving cost), at spot pricing and peak utilization.
# reproducible · synthetic data · no customer data
Reference architecture — the full LLM/RAG-on-EKS deployment architecture, open on GitHub.→
Enterprise OSS LLM Index — quarterly, reproducible. Q2 2026 edition.→
Lighthouse case study — first financial-services deployment, publishing soon.SOON

Production track record

We operate open-weight AI models with GPU acceleration in production today on cloud infrastructure we deploy and our clients own — AWS today, with Azure and GCP supported.

How we're different

Independent, open, and honest about the trade-offs.

01

We publish the benchmarks.

A quarterly, reproducible index of open-weight models on the workloads relevant to financial services business processes. These benchmarks include quality, speed, and cost metrics.

02

We open-source the architectures.

Real reference architectures with SLOs and cost-per-token, on GitHub. A staff engineer can read them in five minutes and tell we know what we're doing.

03

We tell you when self-hosting is wrong.

A managed in-VPC option (Bedrock, Azure OpenAI, or Vertex AI) where it fits; sovereign self-host where sovereignty actually requires it. We're not tied to a model, so we have no reason to oversell one.

From the team

Cost & break-even reality check

The lock-in ratchet

Positive ROI now on use cases with frontier models begin a cycle of lock-in that ends poorly.

Read the argument→

Partnerships we're pursuing

NVIDIA Red Hat OpenShift AI AWS GPU clouds

Start with a fit check.

Tell us the workload and the constraint. In 30 minutes we'll tell you whether open-weight self-hosting is the right answer, and if it isn't, what is.

Book a 30-min fit check