A plain-language walkthrough of how the AI applications we build on top of Mesh handle your data, your end-users, and the AI layer behind every workflow — designed for review by Trabian InfoSec, Compliance, and bank/credit-union procurement.
Application data is hosted by four managed US-based services — all SOC 2 Type 2 compliant. No on-premises component, no data stored in Trabian or partner FI infrastructure, and no replication to third-party regions. Mesh integrations sit alongside this boundary, not inside it — Mesh handles the core/banking layer, this layer handles the AI application data.
Convex is the application backend — not “serverless glue.” It runs on dedicated AWS infrastructure (EC2 + RDS Postgres) in us-east-1, hosting our schema, queries, mutations, HTTP actions, and scheduled jobs. Mesh sits beside it as the governed core/banking integration layer; every other service talks to Convex, and the US boundary contains them all.
Each AI feature ships with a documented purpose, a documented payload, and a documented region. The pattern is the same every time: structured, de-identified inputs → one model call → one cached result. No general-purpose chat over your bank's data, no model fine-tuning on your bank's data.
What the model sees: aggregate transaction categorizations, normalized account-type labels, workflow step identifiers, anonymized free-text the user has typed into the application, and the structured outputs of Mesh integrations after PII has been stripped. AWS Bedrock does not use customer prompts or responses to train or improve its underlying models.
Anonymized, structured workflow data only.
Protected data — never sent to any AI model.
Application data is retained for the term of the FI engagement. FIs can purge member-level data at any time, on demand, end-to-end.
AI-generated outputs (insights, summaries, drafts) are cleared whenever the underlying member record is cleared. Bedrock retains nothing — there is no model-side cache to flush.
If your FI has a specific retention schedule (BSA/AML, NCUA, FFIEC), we will configure to it.
Application data is structured and retrievable for audit and exam responses. No member data is sold, licensed, or shared with outside parties — ever.
Analytics events (PostHog) are behavioral and never include member content. All vendor sub-processors are SOC 2 Type 2 audited and US-resident.
DPA, GLBA / Reg P alignment, BSA/AML guardrails, FFIEC IT examination support — happy to walk any of it with your team.
The short version — the promises behind everything above.
Convex, Cloudflare R2, Cloudflare Pages, Clerk, and AWS Bedrock all hold data within US data centers. Every sub-processor is SOC 2 Type 2 audited.
AI receives only anonymized, structured workflow data. No member PII, account numbers, or untokenized core/banking records are ever sent to the model. AWS Bedrock does not use your data for training.
Member records can be purged on demand, end-to-end. We will conform to your FI's BSA/AML, NCUA, or FFIEC retention schedules.
We provide DPAs, complete vendor questionnaires, support FFIEC IT exams, and align to GLBA / Reg P. Mesh's existing SOC 2 and FI-grade posture extend cleanly to this layer.