01Overview概览
A coach for the why behind the buy.
一个看懂"为什么买"的理财教练。
WhimTell is an AI money coach. It reads your bank transactions (read-only), spots the emotional ones, and asks you to confirm a card or two each morning. I started it at an AI PM bootcamp, led an 11-person team to a first-place MVP, and am now rebuilding it for launch.
WhimTell 是一个 AI 理财教练：只读你的银行交易，找出其中情绪化的那几笔，每天早上请你确认一两张卡片。我在 AI PM 训练营发起它，带 11 人团队做出获第一名的 MVP，现在正重做、准备上线。
Try it: tap an answer, or swipe the card.
示例卡片：右滑 = 冲动消费，左滑 = 没问题。示例数据。
1st placeAI product pitch competitionAI 产品路演比赛第一名
11People led from interviews to a working MVP带 11 人从访谈做到可用 MVP
37% → 70%Detector precision, four weeks of retraining检测精确率，四周重训
17% → 4%Hallucination rate, weekly coaching report周报幻觉率
Budget apps answer "what". The problem is "why".记账 app 回答"花了什么"。问题在"为什么"。
Most money tools treat emotional spending as a math problem: a category, a limit, a red number. In our interviews, three moments kept coming up — and one sentence.
大多数理财工具把情绪化消费当数学题：一个分类、一个限额、月底一个红字。访谈里有三个时刻反复出现——还有一句话。
63%
of Americans say emotions have influenced how they spend.
63% 的美国人说情绪影响过自己怎么花钱。
The 1 a.m. delivery order after a long week.漫长一周后凌晨一点的外卖。
The "I deserve this" cart the weekend after payday.发薪后那个周末"我值得"的购物车。
The purchase forgotten by noon.中午就忘了的那一笔。
"Don't block my card. Don't lecture me. Just make me see it."One-on-one interviews一对一访谈原话"别封我的卡，别说教。让我看见就好。"
Two people, one habit.两个人，一个习惯。
Both have tried budgeting apps and quit. Neither wants a parent.
两人都试过记账 app 又放弃了。都不想要一个"家长"。
Maya, 28Maya，28 岁
The budgeting-app dropout记账 app 的放弃者- Who
- Salaried; has quit a YNAB- or Mint-style app within weeks, more than once.有工资收入；不止一次几周内就放弃了 YNAB、Mint 这类 app。
- Pain
- Categorizing felt like homework. The red number at month end felt like a verdict.给消费分类像写作业；月底的红字像判决。
- Wants
- To stop repeating regrettable purchases without running a budget. Thirty seconds a day.不做预算也能不再重复后悔的消费。每天 30 秒。
"I know where the money goes. Not why it leaks.""我知道钱去了哪，不知道为什么漏。"
Dev, 33Dev，33 岁
The pattern-blind spender看不见规律的消费者- Who
- Spends on impulse at predictable moments: late nights, the weekend after payday, a bad day.在可预测的时刻冲动消费：深夜、发薪后的周末、糟糕的一天。
- Pain
- Regret the next morning, forgotten by noon. The pattern shows up months later, if ever.第二天后悔、中午就忘；规律几个月后才浮现，甚至从不。
- Wants
- A personal baseline ("unusual for me") and one plain-language reason per flag.一条个人基线（"对我来说反常"）和每次标记一句大白话理由。
"Show me the pattern while it's still small.""趁规律还小的时候让我看见。"
Regret arrives on the statement, not at checkout.后悔出现在对账单上，而不是付款时。
So WhimTell never intervenes in the moment. It steps in the next morning, when regret is fresh and the stakes are low.
所以 WhimTell 不在当下打扰，而是在第二天早上介入——后悔还新鲜、代价还很低。
1Trigger触发
2Purchase下单
3Next morning第二天早上
4Month end月底
5Resolve, then quit下决心，然后放弃
Doing做什么
11 p.m. after a long week; the Friday after payday.漫长一周后的 11 点；发薪后的周五
A delivery order; the "I deserve this" cart.一单外卖；"我值得"的购物车
Moves on. Forgotten by noon.照常过日子，中午就忘
Opens the statement or the budget app.打开对账单或记账 app
"Next month I'll do better." Deletes the app."下个月一定改。"删掉 app
Feeling心情
"I've earned this""这是我应得的"
relief, then "oops"松一口气，然后"糟了"
mild regret一点后悔
"what happened?""怎么回事？"
resignation认命
WhimTellWhimTell 介入
Nothing. Never in the moment, never a block.不介入。不当场打扰，不封卡
The transaction enters tomorrow's pipeline.这笔交易进入明天的流水线
One card: "$38.50, 1:32 a.m., your 3rd late-night order this week. A whim?" One swipe.一张卡："$38.50，凌晨 1:32，本周第 3 单深夜外卖。冲动消费吗？"滑一下
The Sunday recap already said it: 5 whims, $186, 3 after 11 p.m., one thing to try.周日周报已经说了：5 次冲动、$186、3 次在 11 点后、下周试一件事
Fewer cards as the model learns; progress = money not spent on confirmed patterns.模型学会后卡片变少；进步 = 在已确认的模式上少花的钱
Metric指标
—
—
Card confirm rate (precision)卡片确认率（精确率）
Money not spent vs. own 12-month average比自己 12 个月均值少花的钱
—
Sample amounts.金额为示例。
v1 proves it sees. v2 proves it helps. v3 goes native.v1 证明看得准，v2 证明帮得上，v3 做原生。
One gate per stage. Nothing scales before the gate is met.
每个阶段一道门槛，没过门槛不放量。
| v1 · Nov 2026 · webv1 · 2026 年 11 月 · 网页版 | v2v2 | v3 · iOSv3 · iOS | |
|---|---|---|---|
| Scope范围 | Detect → confirm → Sunday recap.检测 → 确认 → 周日周报。 | Coaching content; report auto-verification; subscription audit.教练内容；周报自动核验；订阅审计。 | Native iOS app.原生 iOS。 |
| AI & dataAI 与数据 | Three-layer detector; weekly retrain (hand-set → logistic regression → GBM); Plaid sandbox.三层检测器；每周重训（手设 → 逻辑回归 → GBM）；Plaid 沙盒。 | Rewrite when verification fails; per-user tuning.核验不过自动重写；按人调优。 | Real bank data (Plaid production).真实银行数据（Plaid 生产环境）。 |
| Business商业 | Free; validating behavior change.免费，验证行为改变。 | 30-day trial → $6.99–9.99 a month.30 天试用 → 每月 $6.99–9.99。 | App Store. |
| Gate门槛 | Precision ≥ 70%; hallucination ≤ 5%.精确率 ≥ 70%；幻觉率 ≤ 5%。 | Money not spent on confirmed patterns.已确认模式上少花的钱。 | Trial-to-paid conversion.试用转付费率。 |
Three promises.三个承诺。
- As Dev, who orders delivery late at night, I want to see that order the next morning with the reason it was flagged, so that I can judge whether it was worth it.作为经常深夜点外卖的 Dev，我想第二天早上看到这笔消费和它被标记的原因，好判断值不值。
- As Maya, who quit budgeting apps, I want a Sunday recap with numbers I can trace, so that I trust it without checking my statement.作为放弃过记账 app 的 Maya，我想要一份数字可追溯的周日周报，这样不用对账单也能信它。
- As a new user, I want WhimTell to show me patterns from my last 12 months on day one, so that I get value before I've answered anything.作为新用户，我想第一天就看到自己过去 12 个月的规律，这样在回答任何问题之前就先拿到价值。
What "done" means."做完"是什么样。
| Story故事 | Criteria标准 |
|---|---|
| S1 · Morning card | FunctionalityCard shows amount, time, a one-line reason and a confidence level ("maybe" / "likely"). UsabilityAnswered with one swipe; two cards in under 30 seconds. Volume0–3 cards a day; at most one push a day; a quiet day shows a check mark.功能：卡片含金额、时间、一句理由、把握度（"也许"/"很可能"）。易用：一次滑动完成；两张卡 30 秒内。数量：每天 0–3 张；推送最多 1 次；没事的日子显示一个勾。 |
| S2 · Sunday recap | ReliabilityEvery factual statement carries the IDs of the transactions behind it; unsupported statements ≤ 5% on a fixed 100-report set. PerformanceAll numbers computed in SQL before the model writes; the model never does arithmetic.可靠：每条事实陈述带上背后交易的编号；固定 100 份报告上无支撑的陈述 ≤ 5%。性能：数字先用 SQL 算好，模型只负责写句子。 |
| S3 · Day one | FunctionalityLook-back runs as soon as the bank is connected; first session asks ≤ 8 confirmation cards. SupportabilityRead-only access through Plaid; WhimTell can never move money.功能：连上银行立刻回看；首次最多问 8 张确认卡。可支持：通过 Plaid 只读接入，永远动不了钱。 |
| System | PerformanceDetector precision ≥ 70% on cards the model hasn't seen, measured weekly; skip rate ≤ 30%. ReliabilityThe daily 9:00 pipeline (pull → detect → cards → push) runs unattended. QualityAn LLM-only version runs alongside as the control; the three-layer system must beat it.性能：模型没见过的卡上精确率 ≥ 70%，每周测；跳过率 ≤ 30%。可靠：每天 9:00 的流水线无人值守运行。质量：只用大模型的版本并行做对照，三层系统必须赢过它。 |
Trust breaks faster than it builds.信任崩得比建得快。
The first week is when a wrong card costs the most — and when the model knows the least.
第一周是错一张卡最伤信任的时候，也恰恰是模型懂得最少的时候。
| Risk风险 | Type类型 | Mitigation应对 |
|---|---|---|
| Cold start: no labels on day one.冷启动：第一天没有标签。 | Technical | Hand-set weights for week one; the most conservative cards first; precision measured weekly.第一周手设权重；最保守的卡先出；精确率按周测。 |
| Two wrong cards and the user leaves.错两张卡，用户就走。 | Adoption | Precision as the north star with a recall guardrail; confidence on every card; never block, never shame.精确率做北极星、配召回护栏；每张卡标把握度；不封卡、不羞辱。 |
| The LLM makes up numbers in the recap.大模型在周报里编数字。 | Technical | SQL computes, RAG supplies the records, a JSON template forces traceable statements; hallucination tracked on a fixed set.SQL 算数、RAG 给记录、JSON 模板强制每句可追溯；在固定集上跟踪幻觉率。 |
| Card fatigue.卡片疲劳。 | Operational | 0–3 cards a day, one push a day; a quiet day shows a check mark instead of a card.每天 0–3 张、推送 1 次；没事的日子显示一个勾。 |
| Access to real bank data and financial-app rules.真实银行数据准入与金融类 app 规则。 | Compliance | Sandbox first; read-only; "financial education, not advice"; production access in v3.先沙盒；只读；"是财务教育，不是建议"；生产权限放 v3。 |
Lean canvas.精益画布。
Problem问题
- Emotional spending is invisible until the statement.情绪化消费到对账单才看见。
- Budget tools treat it as math: categories, limits, a red number.记账工具把它当数学题。
- Tools that react shame, block or lecture.会反应的工具要么羞辱、封锁、说教。
Solution解决方案
- A three-layer detector that learns your normal.学会你"正常水平"的三层检测器。
- A card or two each morning; one swipe each.每天早上一两张卡，各滑一下。
- A Sunday recap where every number is traceable.每个数字都可追溯的周日周报。
Key metrics关键指标
- v1: detector precision (recall guardrail).v1：检测精确率（召回护栏）。
- Launch: money not spent on confirmed patterns.上线后：已确认模式上少花的钱。
- Paywall: trial-to-paid conversion.收费后：试用转付费率。
Unique value proposition独特价值主张
Budget apps tell you what you spent. WhimTell tells you why — and helps you change it, 30 seconds a day.
记账 app 告诉你花了什么。WhimTell 告诉你为什么——并帮你改变它，每天 30 秒。
Concept: a coach, not a parent.一句话：是教练，不是家长。Unfair advantage不公平优势
- A per-user baseline from 12 months of history.12 个月历史算出的个人基线。
- Every swipe is a training label.每次滑动都是训练标签。
- No lending, no blocking — the things alternatives monetize.不借贷、不封锁——替代品靠这些赚钱。
Channels渠道
- Reddit personal-finance communities → early-access list.Reddit 理财社区 → early access 名单。
- Web app first; App Store in v3.先网页版；v3 上 App Store。
- LinkedIn company page.LinkedIn 公司页。
Customer segments客户细分
- Budgeting-app dropouts.记账 app 的放弃者。
- Pattern-blind emotional spenders.看不见规律的情绪化消费者。
Cost structure成本结构
LLM API calls (labeling a thousand transactions costs well under a dollar) · hosting (Vercel, Railway, Supabase) · Plaid · people.大模型接口（给一千笔交易打标签不到一美元）· 托管 · Plaid · 人力。Revenue streams收入来源
30-day free trial → $6.99–9.99 a month. Anchors: Cleo from $5.99; Rocket Money $7–14; YNAB, Copilot, Monarch about $8–9. No ads, no lending, no selling data.30 天免费试用 → 每月 $6.99–9.99。锚点：Cleo $5.99 起；Rocket Money $7–14；YNAB、Copilot、Monarch 约 $8–9。不接广告、不借贷、不卖数据。Three layers. Different jobs.三层，各司其职。
Rules are cheap but blind. The per-user scorer is personal but can't explain itself. The LLM explains well but needs your baseline computed for it. Every swipe becomes a training label; the scorer retrains weekly.
规则便宜但盲目；按人打分器个性化但不会解释；大模型会解释，但得先把你的基线算给它。每次滑动都变成训练标签，打分器每周重训。
Detect. Confirm. Coach.发现、确认、再建议。
One loop, nothing else. Two tests decided the scope: does the loop fail without it, and can we build it in 12 weeks? Press play.
只做一个闭环。两条标准定取舍：没有它闭环转不转得起来；12 周内做不做得出来。点播放。
In v1v1 做
- Read-only bank connection (Plaid); day-one look-back over 12 months.只读连接银行（Plaid）；首日回看 12 个月。
- Three-layer detector; 0–3 cards a day; one swipe each.三层检测器；每天 0–3 张卡；各滑一下。
- Sunday recap with every number traceable to a transaction.周日周报，每个数字可追溯到交易。
- Weekly retraining on swipes; precision tracked on unseen cards.每周用滑动重训；在没见过的卡上跟踪精确率。
Not in v1v1 不做
- Budgets, goals, net worth — that's Monarch's job.预算、目标、净资产——那是 Monarch 的活。
- Chat with the coach.和教练聊天。
- Blocking cards or enforcing limits. Ever.封卡或强制限额。永远不做。
- Native app; real bank data — both in v3.原生 app、真实银行数据——都在 v3。
Status: bootcamp demo (1st place) → sandbox rebuild, precision 37% → 70% → web launch on whimtell.ai, Nov 2026.
进度：训练营演示版（第一名）→ 沙盒重建，精确率 37% → 70% → 2026 年 11 月在 whimtell.ai 上线网页版。