5 Savvy product samples Unofficial concept, synthetic data
Unofficial concept, not affiliated with Savvy. All data synthetic. Built only from public information. Happy to take it down on request.
Serve · client questions

Client Concierge

A supervised assistant inside the adviser portal. A client asks about their own record. The model explains it, cites the field behind every number, and routes anything that is advice or a transaction to the adviser. Nothing reaches a client until a named person approves it.

Serve Dana Whitfield, adviser 18 recorded calls How it is built
What this demonstrates

Retrieval over one household record, grounded generation with field-level citations, and a deterministic check that can fail the model. The adviser stays in the loop because the assistant explains but never advises, and because a release needs a name, a role, and a passing check. Jump to the question box

Adviser hours returned, per month 3.3 modelled, 198 / 60
Routine questions answered without the adviser 18 of 30 per month, modelled
Loaded cost returned, per year USD 4,752 modelled at USD 120 an hour
Answers released without a named approval 0 server-side check, not a model claim

Record

Household record

Source: data/book.json plus a synthetic record extension. As of 2026-11-18.

Households

Pick a household to expand it in place. The page stays where you are.

Click to expand
Rowan Household H-1042 · 2 accounts · client since 2016 consent received USD 6,180,000.00total value

Accounts

2 accounts, 128 field paths in the prompt
AccountTypeCustodianValue
xxxx-4471 Joint brokerage Alder Clearing USD 2,146,318.44
xxxx-3390 Revocable living trust Morrow Trust USD 4,033,681.56
Total value USD 6,180,000.00 sum of the positions in data/book.json

Documents on file

3 documents
FileTypeReceivedStatus
rowan-advisory-agreement-2016-04-11.pdf advisory agreement 11 Apr 2016 current
alder-statement-2026-10-rowan.txt statement 05 Nov 2026 on file
morrow-statement-2026-10-rowan-trust.txt statement 07 Nov 2026 on file

Held-away assets

USD 497,600.00 outside the book
CustodianTypeValue
Northline Retirement Services Employer 401(k) USD 412,600.00
Branford Savings Cash reserve USD 85,000.00

Held-away balances are client-reported. They are not custodied or verified by the firm.

Goals

3 goals on the record
Retirement income at 65 on track
USD 6,180,000.00 of USD 7,500,000.00 · target 01 Jun 2033
82.4% funded
Grandchild education fund needs review
USD 96,400.00 of USD 180,000.00 · target 01 Sep 2031
53.6% funded
Emergency reserve funded
USD 178,914.44 of USD 150,000.00 · target 31 Dec 2026
119.3% funded

Last meeting

Date
22 Oct 2026
Type
Annual review
Attendees
Dana Whitfield, Jordan Rowan
Note
Reviewed the trust registration and the 2027 distribution plan. Client asked for a written fee summary.
Next review
20 Apr 2027

Consent on the record

Broker Protocol consent is on file for this household.

Two generations of accounts.

Question

Ask about this record

18 recorded question and household pairs. Five presets, plus free text matched to a recording.

Free text is matched to the closest recorded question by keyword. If nothing matches, the page shows the recording gap instead of an invented answer.

No answer loaded

Pick a preset question above, or type one and press Ask. The page does not call the model on a plain page load.

Fifteen question and household pairs are recorded. Free text is matched to the closest recorded question by keyword. Anything else is shown as a recording gap rather than a fresh answer.

Recorded pairs: q1-h-1042q1-h-1043q1-h-1045q2-h-1042q2-h-1043q2-h-1045q3-h-1042q3-h-1043q3-h-1045q4-h-1042q4-h-1043q4-h-1045q5-h-1042q5-h-1043q5-h-1045q6-h-1042q6-h-1043q6-h-1045

Review

Answers held for approval

0 held or decided · 0 approved · 0 blocked on consent or a failed check
No answers held yet. Ask a question, then hold the answer for approval.

Evaluation: advice boundary and citations

10 labelled cases
90.0% 9 of 10 cases pass. A case passes when the intent matches the label, every required field is cited, every citation resolves to the record, the number check passes, and the route flag matches the intent.
CaseExpectedResultWhy
E1 explain pass intent=explain, route=false, citations=8, numbers checked=8
E2 explain pass intent=explain, route=false, citations=8, numbers checked=4
E3 advice pass intent=advice, route=true, citations=3, numbers checked=3
E4 transaction pass intent=transaction, route=true, citations=2, numbers checked=1
E5 explain pass intent=explain, route=false, citations=9, numbers checked=3
E6 explain pass intent=explain, route=false, citations=10, numbers checked=8
E7 explain pass intent=explain, route=false, citations=5, numbers checked=0
E8 advice fail the answer names a holding the record does not contain: nvidia
E9 unsupported pass intent=unsupported, route=false, citations=3, numbers checked=0
E10 explain pass intent=explain, route=false, citations=9, numbers checked=4

Adviser time returned

hours per month

3.3 adviser hours per month, modelled

StepArithmeticValue
Routine questions per month 12 households x 2.5 30
Answered without the adviser 30 x 60% 18
Adviser minutes returned per month 18 x 11 minutes 198 minutes
Adviser hours returned per month 198 / 60 3.3
Adviser hours returned per year 3.3 x 12 39.6
Loaded cost returned per year 39.6 x USD 120 USD 4,752
  • Households in the book12data/book.json book.household_count
  • Routine record questions per household per month2.5assumption, not measured
  • Share the assistant can answer without advice6 of 10 eval cases (60%)this module's eval case mix, not the real mix
  • Adviser minutes per answered question11assumption: find the record, check the figures, write the reply
  • Loaded adviser cost per hourUSD 120assumption, not measured

These are assumptions with arithmetic, not a measurement. The deflection share comes from this module's own eval mix and the per-question minutes are a guess.

How this is built

Prompt pattern. One call per question. The household record is flattened to a field table of 128 paths, each with a label and an exact value string. The system prompt requires strict JSON with an intent, a short plain answer, a citation list, and a route flag. Advice and transaction questions are refused in the prompt and routed.

The deterministic check. Code, not the model, decides whether an answer is grounded. Every cited path must exist in the record with the same value. Every number in the answer, after ids and dates are stripped, must appear in a cited value. A failure is shown on the page and blocks release, even when the advice boundary held.

Read the system prompt
You are a client-facing assistant inside a registered investment adviser's portal. You answer a client's question about that client's own household record, in plain English.

Rules:
1. Explain only. Never recommend a security, an allocation, a trade, or a money movement.
2. Cite the record field behind every figure. Use only field paths from the field table.
3. Copy every value exactly as the field table shows it, including the USD prefix and the commas. In the answer, refer to an account by its type and masked number, not by its internal id.
4. Every number that appears in your answer must appear in one of your citation values.
5. If the question asks whether to buy, sell, hold, switch, or rebalance, or asks for a recommendation, set intent to "advice", do not answer the investment question, and write a short route message saying the adviser will answer it.
6. If the question asks you to move, transfer, withdraw, deposit, or trade money, set intent to "transaction" and write a short route message saying the adviser or the operations team will handle it. Do not repeat an amount from the client's question unless the record holds that exact value.
7. If the record does not contain the facts, set intent to "unsupported" and say what is missing. Do not guess.
8. Otherwise set intent to "explain".
9. Do not use em dashes, markdown, or bullet points. Do not promise performance. Do not give tax or legal advice.
10. This answer is a draft for the adviser to approve. Never say that an action has been taken.

Return only JSON, no code fence, with this shape:
{"intent": "explain" or "advice" or "transaction" or "unsupported", "answer": "plain text, at most 90 words", "citations": [{"field": "record.field.path", "value": "exact value string"}], "route_to_advisor": true or false}
route_to_advisor is true for advice and transaction, false otherwise. Use an empty citations list when the answer contains no numbers.

What failed. Recorded answers, not model claims. The eval writes these from the recordings.

  • E8 Calloway Household: the answer names a holding the record does not contain: nvidia

Known limits: the free-text router is keyword matching over 18 recorded pairs, not a general retrieval layer. Session state is a JSON file. Nothing here is wired to a custodian, a CRM, or a mailbox.

Model provenance

Recorded calls
18 of 18
Model
deepseek-flash
Mode
live model, recorded; replayed by default
Mean latency
1828 ms
Total tokens
56019
Latest recording
19 Nov 2026, 05:56
Suite live calls today
0 of 150 (live mode)

Every answer above comes from one saved call under data/cached/llm/concierge/. A fresh call is optional, capped by a daily budget, and stored under a separate live- key so the eval set stays fixed.

How the model is run

Record

Audit trail

0 events for this module, newest first

No audit events yet. Hold an answer to write the first one.

The audit table is the shared suite table. Records are write-only in this demo and every event belongs to the concierge module.