A context layer that runs
your bank's decisions.
Pipe your data into the Graphical Context Layer and it self-manages decisions for every member and the whole portfolio. It infers micro and macro trends, runs a frontier model on your own hardware to reason over them, and acts inside the governance you set — without a single byte leaving your network.
What your members — and your desk — actually see
You've got room this month
Your NetStream renews Friday. Your balance trend says you're covered — want us to set aside $16 so it's automatic?
We paused a charge for you
A $4,300 charge on a new device was held. Reply YES if this was you, or NO to lock the card.
312 members likely to refinance
As rates dip, this segment's refinance propensity is rising. Pre-approve to retain an est. $4.1M in balances.
Replay speed measured from recorded pilot tasks. Inference runs on a self-hosted open model, so no customer data is sent to third-party AI providers.
The problem
Your best people spend their day copying numbers between screens.
A loan file gets re-keyed into three systems. An AML alert gets worked by hand, screenshot by screenshot. The procedures live in binders and in the heads of a few veterans who are one retirement away from taking them with them.
You can't hire your way out of it, and the big vendors' idea of "automation" is a six-figure, year-long project that still needs a consultant to change a single step.
Illustrative industry figures — to be replaced with your verified pilot data.
Research · Edge inference
Frontier-grade AI that runs at the edge, inside your walls, not someone's cloud.
Most AI vendors pipe your members' data to a cloud model and meter you by the token. For a bank or credit union, that is a compliance headache and a bill that punishes growth.
We do the opposite. Our Edge AI engine ships as a container that runs on your own hardware: frontier-grade automation with the data control of an on-prem system and a flat, predictable cost. The back office for banks and credit unions is the first thing we have built on it. See where we are taking it.
Already standardized on a vendor? If your institution has an approved AI provider, you can point the same engine at OpenAI, Anthropic, Azure, or Bedrock instead, switching from edge to your cloud model without rebuilding a thing.
Your data never leaves the building
The model runs on hardware you control, or on dedicated hardware we run for you. Member PII, account numbers, and documents are never sent to OpenAI, Anthropic, or anyone else.
No eye-watering AI bill
Cloud AI charges you per token, so the more work it does, the more you pay. Ours is self-hosted, so cost is fixed and predictable no matter how many tasks your team runs.
Runs in regulated, even air-gapped networks
Shipped as a single hardened container your IT team can run behind your firewall. No outbound calls required, so it fits the environments examiners and your security team already trust.
Benchmarks · Universal context database
The same frontier weights get sharply better with a banking graph behind them.
We take frontier models from the public usage leaderboard and evaluate each one twice on our banking suite — once from the prompt alone, and once with the universal banking context database attached. The database is model-agnostic: it grounds any model in your members, accounts, and policies, and the lift shows up at both the micro (single member) and macro (whole portfolio) level.
micro reasoning lift
member-level: identity, accounts, single-file decisions
macro reasoning lift
portfolio-level: cohorts, trends, subgraph aggregation
leave your network
retrieval runs on-prem, against your graph
Model names sampled from the OpenRouter usage leaderboard; scores are task pass-rates on our internal banking evaluation suite and are illustrative pending a published methodology. Each model is run identically — the only variable is whether the universal banking context database is attached as an on-prem retrieval layer.
Research · Banking context graph
A knowledge graph of your institution, built for the model to reason over.
Banking work is relational: a member has accounts, an account has transactions, a loan has documents and a policy. We build a private context graph that captures those relationships on-site, so the model reasons over connected facts — the way an experienced officer does — instead of guessing from loose text.
Entities, not documents
Members, accounts, loans, transactions, devices, and documents become linked nodes — so a model can traverse relationships instead of re-reading raw files every time.
Grounded, auditable answers
Every model output cites the exact nodes and edges it used. When a helper flags a wire or approves a file, you can see the precise context behind the decision.
An action context graph, not workflows
Actions attach to the same graph as first-class nodes. Instead of a brittle linear workflow, each decision traverses live context — an identity subgraph informs a rate, a prior investigation informs a new alert — so behavior adapts to the member in front of it.
Lives on your hardware
The graph is built and queried inside your network. It is never uploaded, and it is the memory layer that lets a compact edge model punch far above its size.
Research · Micro & macro
Macro priors live in the weights. Micro evidence lives in the graph.
The frontier weights carry what the model learned across the whole population — the macro prior on how a loan behaves. The context graph supplies what is true about this member — the micro evidence: an eight-year tenure, four linked products, a clean transaction history. The two combine cleanly, because in log-odds space evidence is additive. This is Bayes' rule written in logits.
The giare features read from the member's subgraph and the βi are learned weights. Because log-odds add, every edge in the graph is one more piece of evidence — nothing is re-estimated from scratch.
Macro only
application features, shared weights
5.73%
8.08%
+ Identity subgraph
micro evidence from the graph
1.21%
6.95%
Rate build-up: 5.50% cost of funds + 0.25% servicing + (PD × 25% loss-given-default) + 0.40% capital charge + 0.50% target margin. The default probability is the only term the graph moves, and it swings the expected-loss premium from 1.43% to 0.30%. Weights are illustrative; the additive-logit structure is exactly how the production model combines macro and micro evidence.
Research · Rules on the context layer
Rules read the graph — not a spreadsheet of hard-coded thresholds.
A rule is a condition over the context graph. Ask “is this member's propensity to re-subscribe to a streaming service high enough to surface an offer?” and the layer blends a macro category base rate with microevidence from the member's subgraph, in the same additive-logit space as every other decision. Rules only evaluate once their required data is piped in — no silent guesses on missing inputs.
rule "streaming_resubscribe_nudge":
requires pipe:
card_transactions, merchant_enrichment,
subscription_events, balance_history
when
propensity(member, "video_streaming") >= 0.60
and macro.trend("video_streaming", 90d)
is rising
then
surface offer → relationship officerRequired data pipes
propensity(member, “video_streaming”)
σ(Σ signals) · macro prior + micro evidence
65.3%
Rule fires
z = −1.27 + 0.40 + 0.90 + 0.60 = 0.63, so σ(0.63) = 1 / (1 + e−0.63) = 65.3% ≥ 60%. Change the threshold or the pipes and the same graph re-evaluates — no retraining.
It works like your best new hire
This isn't another system to learn. It's an AI teammate you bring on, train the way you already train people, and supervise like any other member of your team.
Hire
Pick a teammate for the job
Start from a ready-made helper in the SOP library — loan intake, fraud review, member onboarding — or create your own. Think of it as posting a role and filling it the same day.
Train
Show it once, no code
Walk through the task the way you do it today, or hand over an SOP you already have. It learns your exact steps across your systems — just like training a new hire on day one.
Supervise
It works, you stay in control
Your teammate does the work on its own and brings anything that touches a member, a dollar, or a decision back to a person for approval. Every step is logged for audit.
Not a tool your team has to operate. Not a brittle workflow your IT team has to build and maintain. Every action runs on the context graph — grounded in live member context — so it adapts instead of breaking the moment reality differs from the script.
Start from a ready-made helper
Common banking tasks, ready on day one
Browse a library of helpers built for the work your team already does. Pick one, see exactly what it does in plain language, and make it your own — no building from scratch.
New Loan Application Intake
Takes a new loan application and sets it up in your loan system, ready for an officer to review.
AML Transaction Review
Reviews flagged transactions against your AML rules and prepares a case for your BSA officer.
New Member Onboarding (KYC)
Verifies a new member's identity and opens their account once everything checks out.
Wire Transfer Verification
Double-checks outgoing wires for fraud signals before they're released.
Connects to the systems you already run
Your helpers work across the core, lending, fraud, and document tools your institution already depends on — so the work happens where it always has.
Core banking
- Core account systems
- Card & payment platforms
- General ledger
- Teller & branch systems
Loan origination
- LOS platforms
- Underwriting engines
- Decisioning services
- Servicing systems
Fraud & AML
- Transaction monitoring
- Case management
- Sanctions & watchlists
- Alert queues
Documents & e-sign
- Document management
- E-signature
- Content archives
- Imaging systems
…and more. If your team can do it in a browser, your helper can too.
Built for examiners, not just operators
We understand banking compliance. Every action is logged, explainable in plain language, and ready for audit — because compliance isn't a feature we added, it's the floor we build on.
US-Based Data Centers
Dedicated, SOC 2 Type II certified infrastructure. Your data never leaves US jurisdiction.
Zero Training Policy
Your data is never used to train our models. Complete isolation between every institution.
Enterprise Backup & Recovery
Encrypted, immutable backups through Iron Mountain secure facilities for full disaster recovery.
Closed-Loop Processing
Helpers run in isolated environments. No external calls with your data — everything stays inside.
Bring it on at your institution.
We are still early but shipping fast. If you are at a bank or credit union and want to put your operations on an action context graph, let us chat.
What you get
- Full platform access
- Create unlimited agents
- Upload your own SOPs
- Bring your own AI — works with your Claude or OpenAI keys
- Direct line to the founders
Private beta for financial institutions. We will reach out within a day.