About · the sample
About this sample
A working product sample for one RIA transition workflow. It is not a Savvy product and it does not touch real client data.
Transition Copilot is a concept for the operations layer an RIA needs when an adviser joins with a book built somewhere else. The sample starts with six synthetic documents and a CRM export. It ends with a plan, a review queue, and an evidence trail.
This page lists what runs, what is simulated, where the guardrails are, and what is missing. The sample is built so those claims can be checked against the data and the code.
Provenance
What is real and what is simulated
What is real
Runs in this build
- A server-rendered FastAPI app with SQLite state. Draft status, evidence, consent, and the audit trail persist across requests.
- A real extraction pass over the six synthetic documents. Every field has a confidence value and a source span.
- A plan with account-level due dates, owners, phases, and a seeded NIGO risk register.
- Role-based approval with an evidence file, an artifact hash, and an append-only audit event.
- The hours arithmetic on the value page. Every step can be checked by hand.
What is simulated
Invented for the demo
- Every name, household, custodian, account number, address, phone number, and email address.
- The policy references and custodian rule names. They are labelled synthetic in the plan.
- Task dates, completion states, and the six hard findings. They are seeded to show a mid-transition snapshot.
- The 45 minutes per account, the 100-client reference book, and the 25 to 35 percent reduction. They are modelled assumptions, not a time study.
- The compliance outcome. This is a product sample, not legal or compliance advice.
How the AI is used
Deterministic by default, live by choice
Extraction is a deterministic parser.
It reads the statement layout with fixed rules. There is no model call in this run, and no field is invented when the text is missing.
Drafting is deterministic by default.
Each draft is assembled from seeded facts. The provider interface also has a cached path that replays recorded model output by key, and an optional Bedrock path behind an environment switch. The cache exists so a live run cannot make the demo drift. This run used no model calls.
Every numeric claim must resolve.
A claim table points each number or status back to a book, plan, or policy reference. An uncited number fails a policy check and blocks approval.
No model approves anything.
A human role approves. The model can draft and check. It cannot seal an evidence file or write an approval as a person.
Compliance map
Five controls, one screen
| Control | What the product does | What it does not do |
|---|---|---|
| Explain versus advise | Client drafts explain what the record shows and what happens next. Recommendation, transaction, and guarantee language fails the advice check. An advice question routes to the adviser. | The sample does not recommend a security, a rollover, an allocation, or a timing decision. |
| Protocol consent | The book record carries the five Broker Protocol fields and the date they arrived. Client contact waits for resignation. Each household has a consent status. A pending household blocks its client-facing draft. | The sample does not use account numbers, balances, holdings, or statements as protocol data. It does not contact a client before consent. |
| Records and retention | Approval writes an evidence file with approver, role, version, audience, channel, disclosures, retention date, and artifact hash. Records are firm-held in the demo SQLite file. Five-year retention is the target. | The sample does not claim a vendor retention agreement. A production deployment would need a zero-retention carve-out in the vendor contract. |
| Privacy | All data is synthetic and local. Account numbers are masked. There are no tax identifiers, no analytics, no third-party requests, and no external fonts or images. | The sample does not process real client data, connect to a custodian, or send telemetry anywhere. |
| Human approval | A named role is required. A failing policy check disables approval on the server, not only in the button. Approve, reject, and consent each write an audit event. | The sample does not auto-send, schedule, or deliver a draft. There is no mail or custodian integration. |
Limitations
What this sample does not do
- No custodian API, no ACATS submission, and no account opening integration. Transfer paths are a plan, not an execution.
- The documents are synthetic text files. There is no PDF ingestion, OCR, or document storage.
- The extraction eval is small and in-sample. A high score on 51 labels describes this dataset, not the next statement.
- No public NIGO rate and no public days-to-first-funded series exist. The product instruments both from day one instead of inventing a benchmark.
- The policy checks are drafting controls. They are not a legal opinion and the rule references are synthetic.
- No authentication, no multi-user roles, and no production database. State is one local SQLite file.
- The value model is an assumption set. It is not a time study and not a measured result.
Not affiliated with Savvy
This is an unofficial concept built only from public information. It is not endorsed by, connected to, or supported by Savvy. If you represent Savvy and want it removed, contact me through the repository and I will take it down.
Who built this
Anthony Stachowitz
Product management work sample · November 2026
I built this to show how I think about transition operations. The record comes first. The plan follows from the record. The approval trail is part of the workflow, not a report added at the end.
I did not work at Savvy and did not use any non-public Savvy information. The data is generated, the custodians are fictional, and the policy references are synthetic. The brief covers the market context, the competition, and the decisions I would defend in a review.