How these samples are built
The point of the samples is not the pages. It is the way they were built: real model calls, recorded so they replay, small evaluations, and an honest list of what failed.
The model and the provider
All model work runs on DeepSeek through its API. The default model is DeepSeek Flash, with DeepSeek V4 Pro available for the harder generation steps. Every output shown in the samples was produced by a real call, saved to a recording with its model, latency, and token usage, and replayed by default so a reviewer never sees a spinner fail.
Where the deployment allows it, a module can make a fresh live call inside a small daily budget. The budget exists so a public demo link cannot run up an API bill.
Model outputs are generated text, not advice, and every client-facing output sits behind a human approval step and an evidence record.
How this was built in two days
Two days, five modules, 179 tests. Transition Copilot was built the day before the other four. Every module shares the same platform: the DeepSeek client and recorded-run cache, the audit log, the module state, and the suite layout.
I made the product decisions, wrote or directed every prompt, defined the guardrails and the evaluations, reviewed the output, and ran the QA. AI coding agents wrote most of the code. That division of labour is the workflow I would bring to the role: the leverage is real, and so is the need to review, test, and keep a human accountable for what ships.
This is a work sample, not a proposal for Savvy's roadmap. You may already have some of this internally. If so, the useful part is the method: what I noticed, what I chose, how fast it came together, and how the controls are built.
Five AI patterns, one per sample
| Sample | Pattern | What the model does |
|---|---|---|
| Transition Copilot | Document extraction and workflow orchestration | Turns a signed adviser's book into a structured record, a signature and NIGO plan, and an approval file. |
| Client Concierge | Retrieval over the household record and grounded generation | Answers a client's own questions from their record, with citations, and routes anything that is advice to the adviser. |
| Book IQ | Reasoning and ranking over the book, plus outreach drafting | Ranks the next dollar in every book with evidence, and drafts the play for the adviser. |
| Supervision in the loop | Generation plus claim classification and policy checks | Drafts adviser marketing inside a pre-approval workflow with an evidence file per piece. |
| LP Ops | Document extraction and exception detection | Takes private markets paperwork from intake to capital call without adding headcount per dollar. |
Evaluations
| Sample | Evaluation | Cases | Score | Failures shown in the page |
|---|---|---|---|---|
| Transition Copilot | Extraction field accuracy (in-sample) | 51 | 100.0% | no |
| Book IQ | Play and signal accuracy | 12 | 100.0% | no |
| Client Concierge | Advice-boundary and citation accuracy | 10 | 90.0% | yes |
| LP Ops | Subscription packet extraction | 5 | 96.9% | yes |
| Supervision in the loop | Claim flag accuracy | 12 | 75.0% | yes |
What the models got wrong
- E8 Calloway Household: the answer names a holding the record does not contain: nvidia
- P-03 Exception severity: Bank name mismatch: expected Medium, model returned High.
- P-04 Exception severity: Date anomaly: expected Medium, model returned High.
- In-sample eval: the five packets were written for this test, so the score describes these documents, not the next subscription packet.
- E-04 carried the right flag but the wrong type (performance).
- E-07 was read as a supported factual claim, so the performance flag never fired.
- E-08 was read as a supported factual claim, so the performance flag never fired.
- The eval is in sample: the same model, the same fact sheet, and twelve cases chosen by hand.
Recorded runs
- Recorded calls
- 54
- Models
- deepseek-flash
- Mean latency
- 2041 ms
- Prompt tokens
- 84,825
- Completion tokens
- 14,568
- Live budget today
- 0 of 150
The guardrails that repeat everywhere
- Every output that reaches a client or a custodian is held for a named human approval.
- Every number in a draft points at a seeded source field, or the draft is flagged.
- Client-facing work for a household without protocol consent is blocked on the server.
- Advice language is blocked. Anything that looks like a recommendation routes to the adviser.
- Each approval writes an evidence record and an audit event.
What this shows, and what it does not
- Shows: speed with AI, product judgment across five surfaces, control design, and honesty about model failures.
- Does not show: production engineering depth, a held-out benchmark, or any custodian, CRM, or private markets integration.
- The interview question I expect: what did you do versus the agents. Answer: every product decision, prompt, guardrail, evaluation, and review. The agents typed most of it, and I am accountable for all of it.
What this is not
- Not a Savvy product, and not affiliated with Savvy.
- Not connected to a custodian, a CRM, or a private markets administrator.
- Not a measurement. The hours and margin figures are modelled estimates.
- Not legal or compliance advice. The policy references are synthetic.