Steadfast: a life insurance needs plan, talked through

1. 1User interface & intuitivenessOne question at a time, chips for every answer, 16px type, and a plain explanation beside every insurance word.
2. 2Functional requirements & impactDependents, income, debts and coverage go in. A priced needs plan with its math comes out, plus term vs permanent life.
3. 3Solution design & innovationWe trained our own underwriting model, so every price on the page comes from this person's answers.
4. 4Demonstration and presentationReal screenshots of every screen on this page, and a timed four-minute runbook.
5. 5Does it work?The core flow runs signed out, end to end, with 450 automated tests behind it. Lincoln's own policies and advisors are built into the result.
6. 6Technology platforms employedNext.js 16, PyTorch, FastAPI, Neon Postgres, the OpenAI Responses API, Vercel, and open data from the CDC and the Society of Actuaries.
7. 7Security accommodationsDesigned around US insurance rules first, then hashed tokens, zod on every boundary, and a model that sees only a fact sheet.
8. 8Technical creativityThe model explains and the code calculates. Any figure the model invents is caught before it reaches the screen.
9. 9Architecture & methodologyA user flow through the model and the underwriting engine, a storyboarded conversation, docs as contracts, and a timestamped log of the night.
10. 10ComplexityAn underwriting engine we trained ourselves: a survival model fitted to 47,469 US adults and priced with Society of Actuaries tables, held to 0.99 observed-to-expected deaths. Plus one API route, no chart library, and 450 tests that run in three seconds.

• A guided conversationFamily, income, debts, coverage, then two health questions. Chips or their own words, one question at a time.
• A plan with its mathOne rounded figure, a reasonable range, and a ledger where every row states its assumption.
• Prices from our own modelA health class and term and permanent prices for this person, from the underwriting model we trained.
• Term vs permanent, with tradeoffsOnly the rows that differ for them, the closer fit marked, plus put-the-difference-aside and a two-policy option.
• Questions answered from the planAsk anything about it. Answers use only figures already on the page.
• Life CanvasSketch a different version of their life on the saved plan: a new baby, a raise, a paid-off mortgage, five years ahead. Every number and price redraws live, and nothing is saved until they say so.
• The handoffA one-page PDF and a request for a licensed Lincoln Financial advisor, with the plan attached.
• Written for first-time buyersA plain explanation beside every insurance word, 16px type, and a guide built from the same glossary.

1. steadfast/calculator

   

   01

   A conversation, one question at a time

   Family, income, debts, then coverage. Every question has answer chips, a line on why we ask, and room to type in their own words: “240k”, “$1,800 a month”, “my wife smokes, I don’t”. It works signed out.

2. steadfast/calculator

   

   02

   The plan builds while they talk

   One rounded figure with a reasonable range, and a ledger beside it. Every row states its assumption, like “10 years × $85,000”, and the rows always add up to the figure.

3. steadfast/calculator

   

   03

   Five charts, drawn by hand

   How the number adds up, the family timeline, income continuity, a cost explorer and term vs permanent life. Plain SVG with no chart library, so each one fits a 390px phone.

4. steadfast/calculator

   

   04

   A full plan and an essentials plan

   Plan A covers everything they told us about. Plan B keeps the essentials at a lower monthly cost. Both are priced side by side, each names the Lincoln policy it’s closest to, and they choose.

5. steadfast/calculator

   

   05

   Term or permanent life, in their dollars

   Two sentences of explanation, then only the rows that differ for them, with the closer fit marked and the reason. It also shows buying term and putting the difference aside, and a two-policy option when that costs less.

6. steadfast/calculator

   

   06

   Prices from our own underwriting model

   Two short health questions, about tobacco and four common conditions, set a health class with a plausible range. Build is assumed average, and the page says so. The class, the range and the prices come from the model we trained.

   How the model was trained

7. steadfast/calculator

   

   07

   Questions answered from their own plan

   After the plan, they can ask anything: “Why until my youngest is 22?” Answers use only the figures on their page.

8. steadfast/dashboard

   

   08

   Save it, and come back to a dashboard

   Saving needs only an email; a link to choose a password follows. The dashboard holds the plan, one-tap life changes, health toggles and a chat that reprices everything live.

   See the dashboard

9. steadfast/calculator

   

   09

   Something to take to the kitchen table

   A one-page PDF of the assessment, with the math, term vs permanent life and the pricing notes, and a request for a licensed Lincoln Financial advisor to call, with their plan attached.

10. steadfast/calculator

    

    10

    Written for people buying for the first time

    Insurance words come with an info button and a plain explanation that opens on tap, hover or keyboard. A plain-words guide answers the questions people ask before they buy, from the same glossary, and a test keeps every explanation to three sentences with no jargon. Reading text is 16px or larger, and there’s no red anywhere.

    Read the guide

11. steadfast/calculator

    

    11

    Just as calm on a phone

    Below tablet width the conversation and the plan become two tabs, so each gets the whole screen. Chips stay large enough to tap, and every chart redraws to fit.

1. steadfast/dashboard

   

   01

   The saved plan, as it opens

   The figure, its range and the ledger, with the planner’s welcome in the chat beside it. What-if controls sit right under the number: years of support, college, the mortgage.

2. steadfast/dashboard

   

   02

   Life Canvas: one-tap life changes

   The Life Canvas lets them sketch a different version of their life and watch the plan redraw. “A new baby”, “A 20% raise”, “Paid off the mortgage”, “Partner stops working”, “Five years from now”: each one is a patch of facts over the saved answers, so the whole plan recomputes and the chat says what moved. Nothing is saved until they say so.

3. steadfast/dashboard

   

   03

   Health changes reprice through the model

   Toggling “High blood pressure” sends the new answers to the underwriting engine. Plan A moves from $162 to $210 a month, the class is named, and every chart redraws. “Make this my plan” keeps it.

1. Step 1

   They answer

   A chip or a typed sentence.

2. Step 2

   Read locally

   A parser handles money shorthand, ages, households and health. Messages of 8 words or more, or ones it can’t read, also go to the model.

3. Step 3

   Needs math

   Plain TypeScript works out the plan and its ledger. The model never computes money.

4. Step 4

   Priced

   Our underwriting engine prices both plans. Requests are debounced and cached.

5. Step 5

   Fact sheet

   Every figure the model may say, already formatted, built from the plan and the engine’s prices.

6. Step 6

   Explained

   The model writes the summary and answers from the fact sheet only.

7. Step 7

   Checked

   Any dollar figure that isn’t on the fact sheet is logged, and that answer is never cached.

Trained with PyTorch, served with numpy

A Gompertz proportional-hazards survival network, trained five times on bootstrap resamples with the exact left-truncated, right-censored likelihood. The trained ensemble is exported to 112 KB of numpy weights, so the live engine serves without PyTorch and fits a Vercel Python function. One script rebuilds everything from the public sources in about 30 seconds after the download.

1. Step 1

   The cohort

   47,469 US adults from the CDC’s NHANES surveys, 1999 to 2018, linked to the National Death Index: 5,448 deaths over 467,880 person-years. Split 70/15/15, and the test set was never touched in training.

2. Step 2

   The model

   A proportional-hazards survival model on the attained-age scale with a Gompertz baseline per sex and a two-layer network (64 units, SiLU) over ten answers. The exact left-truncated, right-censored likelihood, weighted by survey weights, in PyTorch.

3. Step 3

   The range

   Five models trained on bootstrap resamples. Their spread is the best-to-worst plausible class, which is where the “ranges, not verdicts” rule gets its numbers.

4. Step 4

   The classes and prices

   Relative risk against an insured-like pool sets Preferred Plus through Table 8 at the industry class mix, with carrier-style knock-out rules. Society of Actuaries 2015 VBT mortality and an actuarial premium formula set the price, calibrated to 152 published rates with 115 held out.

• 47,469

  US adults in the training cohort (NHANES 1999–2018, linked to the National Death Index)

• 0.837

  10-year C-index on held-out people (age and sex alone: 0.792; a Cox model: 0.832)

• 0.99

  Observed over expected deaths on the test set

• 26 / 26

  Pricing and underwriting invariants pass, like “longer terms never cost less”

• Estimates, never quotes

  NAIC Unfair Trade Practices Act (Model 880) · Advertisements of Life Insurance Model Regulation (Model 570)

  Misrepresenting a policy’s terms, benefits or price is an unfair trade practice in every state. So every price on every screen says “planning estimates, not quotes”, the health class says the insurer decides after its review, and the PDF says it is not a quote, application or underwriting commitment. The model is instructed never to say someone qualifies or to quote a Lincoln price.

• Guaranteed and not-guaranteed kept apart

  NAIC Life Insurance Illustrations Model Regulation (Model 582) · Standard Nonforfeiture Law (Model 808)

  An illustration may not show a non-guaranteed value as if it were guaranteed. Permanent life cash values here are the guaranteed minimums under the Standard Nonforfeiture Law on the 2017 CSO table, labeled “guaranteed minimum”. The “put the difference aside” figure at 5% sits next to it labeled “not guaranteed”, and the chart says which is which.

• A licensed advisor makes the recommendation

  NAIC Producer Licensing Model Act (Model 218)

  Only a licensed producer may sell, solicit or negotiate insurance. Steadfast shows the tradeoffs in the person’s own numbers and explains why one fits closer; the application and any recommendation come from a licensed Lincoln Financial advisor, who gets the worked plan with the request. The two kinds are named the way Lincoln names them, term life and permanent life, and Lincoln’s permanent product is named truthfully as universal life, not whole life.

• Underwriting factors a regulator would recognize

  State unfair-discrimination statutes (Model 880, §4) · Colorado SB 21-169 and Regulation 10-1-1 · NAIC Model Bulletin on the Use of AI Systems by Insurers (2023)

  Life insurers may classify only by factors tied to expectation of life, and newer rules ask that algorithms be governed, tested and explainable. The engine uses age, sex, build, tobacco, four doctor-diagnosed conditions and self-rated health: the questions on a paper application. No credit score, ZIP code, occupation or purchased consumer data. Every quote returns the factors that moved the class with their multipliers, a model-card endpoint publishes the metrics, and the validation report is regenerated by a script.

• Financial and health information, kept to a minimum

  Gramm-Leach-Bliley Act, Title V · NAIC Privacy of Consumer Financial and Health Information Regulation (Model 672) · Insurance Data Security Model Law (Model 668) · NYDFS Part 500

  Health answers collected for insurance are protected information. The planner works without an account, so most people never store anything. A saved plan holds the five health answers and the facts behind the ledger, and nothing else. The model never receives a name, email or account, and its storage is off. There are no third-party analytics or advertising pixels, and every email is one the person asked for.

• Plain language and reachable controls

  NAIC Policy Language Simplification Model Act (Model 575) · ADA Title III and WCAG 2.1 AA

  States require policies to pass a reading-ease test, and courts apply the ADA to websites. The chat is held to three sentences a turn, every insurance word has a glossary entry that a test keeps jargon-free, reading text is 16px or larger, every control works by keyboard with a visible focus ring, nothing is conveyed by color alone, and motion respects the reduced-motion setting.

• Only hashes of tokens are stored

  Session cookies and password links are 32 random bytes. The database keeps only their SHA-256, so a copy of the tables can’t be used to sign in.

• Passwords: bcrypt, cost 12

  At least 8 characters and at most 72 bytes, since bcrypt ignores anything longer. A wrong email is checked against a dummy hash, so it takes as long as a wrong password.

• Every input is checked on the server

  Each server action and the one API route validate their input with zod. Amounts are worked out again on the server, never trusted from the browser. Model and engine replies are parsed against a schema too.

• Links are single-use and short-lived

  A password link lasts 24 hours and is claimed in one transaction, so it can’t be used twice. Setting a password signs out every other session. One link a minute per account, and the same reply whether or not an email has an account.

• No personal data in URLs

  Plans live in the database, tied to the account, never in the address bar. The only query value anywhere is a random password-link token. Cookies are httpOnly, SameSite=Lax, and secure in production.

• The model sees only the fact sheet

  It never gets a name, email or account. It receives their answer or question, the plan’s fact sheet and the chat so far, with storage turned off (store: false). Its output is structured, and the app applies only known fields.

• The core flow needs no account

  Anyone can open the planner and get a priced, explained plan in a few minutes. Sign-in exists only to come back later; it is never on the path to the result.

• Lincoln’s products are in the answer

  Each plan names the Lincoln policy it’s closest to: TermAccel under $1M with the quick application, LifeElements for larger amounts or older ages, and WealthPreserve 2 IUL for permanent coverage. Terms snap to Lincoln’s 10, 15, 20 or 30 years, and the age and amount limits are checked.

• It hands a qualified lead to an advisor

  “Talk to an advisor” records a request with the person’s email, a preferred time and the plan attached, so the advisor opens the call with a worked needs analysis instead of a blank form.

• What it would take to ship

  Swap the engine’s calibrated public rate charts for Lincoln’s rate book and underwriting classes, add filed disclosures, and keep everything else. The engine is one service behind a typed contract, so that swap touches nothing in the app.

• 450

  Automated tests

  16 files, under three seconds, run with the typecheck and lint before any change lands.

• 351

  Unit

  The pure library: the answer parser, the needs math, the conversation order, every schema, pricing, the Lincoln policy match, the what-if answers, and every glossary explanation.

• 89

  Integration

  Server actions and page readers with the model, the engine and the database mocked: accounts and password links, the model’s three tasks, the engine’s tradeoffs, and what each page shows a visitor.

• 10

  Design rules

  A lint over the source for the design system: no hex colors, no font sizes off the scale, no round corners, no removed focus rings, no emoji.

1. 01

   The conversation was storyboarded first

   Family → Income → Debts → Coverage was fixed before any code, and every question was written with its chips, its “why we ask” and the fact it sets. The plan panel follows the same order, so the right side answers the left side.

2. 02

   A design system before any screen

   Tokens, nine components and the content rules (three sentences a turn, ranges not verdicts, no red, no exclamation marks about money) were written into docs/design and ported into typed components. A design lint test fails the suite when a screen breaks them.

3. 03

   Docs as contracts

   One project brief for every contributor, human or AI, with the architecture, the rules and the roadmap. A presenter’s runbook with a timed script. A night log with one entry per change: what, why, files, how it was verified.

4. 04

   A roadmap, worked in order

   Saturday 13:34 first commit, 14:55 design system, 15:07 planner, 18:52 landing, 23:16 dashboard, 00:53 engine deployed as a Vercel service. Then the agreed list: the engine’s tradeoffs in the UI, demo hardening with a seeded account and pre-warmed answers, deploy.

5. 05

   Verified the same way every time

   Typecheck, lint, build, 450 tests, then driving the real app in a browser at desktop width and at 390px. The screenshots on this page are captured by a script from the running app, as the demo family.

pnpm demo starts the app and the engine together, and pnpm demo:seed resets the demo account and caches the model’s answers for the demo family ahead of time, so the summary and the suggested questions answer instantly.