Menu

Polymath AI Case Study: How a Multi-Expert, White-Label AI Platform Is Engineered

Table of Contents

Calling a language model is the easy part of running an AI product. The harder work sits around the call: deciding who may use which assistant, what each reply costs, how an answer looks on a phone, and what stops a health or legal assistant from saying something reckless. Polymath AI is worth studying because it makes those surrounding concerns the center of the product.
It ships as a Laravel application with a Livewire web dashboard, a Filament operator console, and a Flutter project that compiles into iOS and Android apps. Fifteen assistants, each with a distinct personality and price tag, sit behind one shared backend.
Source basis. Zipprr publishes this analysis of Polymath AI, a product made and sold by someone else on CodeCanyon. Zipprr did not create, sell, or operate it. The paragraphs headed “Engineering reading” and the “Operator takeaway” lines are Zipprr’s own analysis, not statements about how the product’s code is written.

The stack at a glance

LayerWhat it uses
ServerLaravel on PHP 8.3 or later
DataMySQL 8 or MariaDB 10.6
Web appLivewire dashboard
Operator consoleFilament
Phone appsFlutter 3.32 or later, with one codebase serving both platforms
AI vendorsOpenAI, Google Gemini, Anthropic, and Stability, chosen per workflow
PaymentsStripe
Optional extrasFirebase push, SMTP or a similar email service, a support chat widget

A lens for reading it: five control planes

Instead of walking through thirty-odd features one at a time, it helps to sort them by the question each one answers. We call the result the five control planes.
PlaneQuestion it answersWhere it lives in Polymath AI
PersonaWhat is this assistant?Expert records: instructions, model, temperature, look, roster
MeterWho may use it, and what does a reply cost?Plans, credits, referral rewards, Stripe
IdentityWho is this person on this device?Accounts, web sessions, phone tokens, email codes
DeliveryHow does the answer reach them?Live streaming, response cards, voice, uploads, images, history
GovernanceHow does the operator steer all of it?Filament console, branding, announcements, safety layer, installer
The notable design choice is that the operator steers every plane through records and settings rather than code changes. The rest of this case study follows from that choice.
Clients
Phone app (Flutter)Web app (Livewire)
One Laravel backend
IdentityMeterPersonaDelivery
Governance
Filament consoleWeb installerBranding & announcements
Infrastructure
MySQL / MariaDBAI vendorsStripeFirebaseEmail

Customizing the server side is a job for Laravel development skills, and the phone apps call for Flutter app development experience.

Persona plane: an assistant is a row in a table

What the product provides. The roster holds fifteen specialists, including health information, legal information, tutoring, trip planning, data analysis, cooking, coding, and writing. Each record stores the assistant’s job description, its instructions, which vendor and model answer for it, the temperature, a price in credits, its colors, and optionally the plan a user needs. Conversations never mix across experts.

Engineering reading. When an assistant is data, the roster can change on a Tuesday afternoon with no deployment. Routing is the second benefit: because the vendor and model sit on the record, a quick-lookup assistant can run on a small, cheap model while a data analyst runs on a stronger one. The cost of this flexibility is that a prompt edit changes product behavior instantly, for every user, with no review step.

Operator takeaway. Keep a changelog of who edited which prompt, price, or model, and when.

Meter plane: credits connect plans to vendor bills

What the product provides. Administrators create plans and sync them to Stripe. Some experts can be reserved for certain plans, and every expert carries its own credit price. Renewals top up credits. The platform applies charges so that parallel requests cannot draw on the same credits twice, and a reply that fails is refunded without any manual step. Before a Stripe event is allowed to alter anything, its signature is verified, and the platform also handles a grace period that begins once a plan’s paid-through date has passed. Referral rewards stay off until an operator turns them on. Once active, a reward can fire when someone registers, verifies their email, or makes a first purchase; a monthly ceiling limits abuse, and the database itself rejects duplicate rewards.

Engineering reading. Three hazards sit in this plane: double spending, charging for failures, and replayed payment notifications.

Operator takeaway. Once a month, compare credits issued and consumed against the invoices from each AI vendor, so a mispriced expert shows up early.

Identity plane: one account, two kinds of login

What the product provides. People can sign up, sign in, and reset passwords on their own. New accounts are confirmed with a one-time code sent by email. Browsers use web sessions, while the phone apps hold tokens in the operating system’s protected storage: Keychain on iOS and Keystore on Android. A password reset signs the person out of every other session on web and phone.

Engineering reading. Browsers and native apps handle login differently, so forcing one mechanism on both usually weakens one of them. Giving each platform its own credential type, both resolving to the same account row, lets the plan and credit logic run exactly once. The sign-out-everywhere behavior after a reset matters most when a phone is lost.

Operator takeaway. Rate-limit sign-in and email-code attempts, and give phone tokens a finite lifetime.

Delivery plane: getting the answer onto the screen

What the product provides. Replies appear as they are generated, on web and on phone. Certain experts show a custom card in place of plain text: a cooking reply becomes a checklist with counters, a data analysis reply can draw a chart, a translation shows two languages side by side, and a travel plan becomes day cards on a timeline. The phone app adds spoken input, read-aloud playback, and file uploads in supported workflows. Both clients include an image studio and a saved gallery, and image generation is also available from the phone. Transcripts are saved to the account, stay in sync across devices, and can be exported. The web dashboard also covers profile settings, notifications, invitations, language, and theme, and the interface is available in Arabic (right to left), English, French, German, Portuguese, and Spanish.

Engineering reading. Streaming improves how fast the product feels, and cards only work if the client can predict the shape of what arrives. That points to a strict schema per card type, checked on the server, with plain text displayed instead whenever a reply does not fit.

Operator takeaway. Test on a slow connection and with a deliberately interrupted reply before launch.

Governance plane: where the operator steers

What the product provides. The Filament console covers the expert roster, plans, credit balances, user accounts, referrals, announcements, legal pages, vendor keys, and branding. The product name, slogan, logo, icon, and brand color appear across the web app and the public landing page, and the phone app downloads them from the server each time it opens, so routine rebranding avoids a new store build. Announcements can go to everyone, to one plan, to users of one expert, or to named people, over web, phone, push, or email. The public landing page lists whichever experts are currently enabled. Setup runs through a guided installer in the browser, which audits the server, wires up the database, handles activation, and creates the first admin account.

Sensitive experts, such as health and legal, can carry disclaimers and safety instructions that the server appends to the prompt. Changing the prompt text shown in the console cannot remove that layer.

Engineering reading. Appending safety text on the server keeps the guardrail outside the content an operator edits day to day. It is still a guardrail and not a guarantee, so it should sit alongside on-screen disclaimers and adversarial testing. Anyone building a dedicated AI legal assistant meets this same question. If you also need a review trail for changes an AI proposes, our write-up on traceable AI change approval covers one approach.

Operator takeaway. Treat disclaimers as part of the interface, not only as text inside the model’s instructions.

One request, followed across the planes

1
Identity: The person confirms their account with an emailed code.
2
Meter: Their plan grants credits.
3
Persona: They open an expert, and the platform pulls up that expert's record.
4
Meter: The platform checks plan eligibility and sets aside the expert's price.
5
Delivery: They send text, speech, or a supported file.
6
Persona: The vendor and model named on the record produce the answer.
7
Delivery: The reply streams in, possibly drawn as a card.
8
Meter: The charge stands, or it is reversed if the reply failed.
9
Delivery: The transcript is saved and appears on both devices.
10
Governance: Later, the operator adjusts prompts, prices, or branding from the console.

Trade-offs worth weighing

DecisionWhat you gainWhat to watch
Per-expert credit pricesPrice tracks each vendor's costMore numbers to maintain, and users may wonder why replies differ in price
Rules enforced in one backendWeb and phone cannot disagreeThe backend becomes the one place that must stay up
Separate web and phone credentialsEach platform uses its native secure storageTwo kinds of session to audit
Safety text added on the serverOperators cannot delete it by accidentDoes not replace testing or human review
Stripe-hosted checkout and billing portalLess payment interface to buildLess control over the exact flow
Experts as editable recordsFast iterationMistakes go live immediately
Self-hostingYou choose where data lives and what it costsYou also own patching, backups, monitoring, and uptime

Worked example: pricing a three-expert lineup

These figures are made up purely to show the reasoning, and they are not Polymath AI’s prices.
Writing helper (light model)Travel planner (mid model)Data analyst (heavy model)
Vendor cost per reply1 unit2 units5 units
Credit price1 credit2 credits3 credits
Revenue at 2 units per credit2 units4 units6 units
Margin per reply+1 unit+2 units+1 unit
If all three were priced at one credit, the data analyst would lose 3 units on every reply. Since pricing is set per assistant, the fix is a console edit instead of a software release.

Acceptance checks before launch

CheckPass condition
Send two requests at once when credits cover only oneOne goes through, the other is declined, and the balance never dips below zero
Kill a reply partway throughThe charge is reversed and the user sees a clear message
A subscriber on a cheaper plan opens an expert reserved for a pricier oneBlocked before any vendor is called
Replay the same Stripe eventCredits are added once
Send a webhook with a bad signatureRejected, balance unchanged
Reset a passwordOther web and phone sessions stop working
Rewrite the prompt shown in the console for a health expertThe server-added safety instructions are still sent to the model
Submit the same referral twiceOne reward only
Change the branding in the consoleThe phone app picks up the new look the next time it opens, with no new build
Return a malformed card payloadThe screen falls back to plain text
Worth tracking continuously: how often each expert answers successfully, how long replies take, what vendors charge compared with credits collected, how many refunds go out, and any gap between the credit ledger and customer balances, which should always be zero.

Where this architecture fits, and where it needs more work

A good fit for agencies running branded AI memberships, education or travel services, and any operator who wants one backend behind both a website and phone apps. If AI is only one module of a bigger product, a super app platform may be the better frame, and learning-focused businesses can compare learning management platforms.
Extra work is wise in three places. Health and legal experts deserve a professional review of their disclaimers. Heavy traffic calls for load testing before launch. And because the mobile app arrives as source code, the operator handles store accounts, signing, and publishing.

Next step: weigh this against your own AI product plan

Weighing up an AI assistant or a subscription business? For an outside view on model routing, credit design, or launch readiness, book a free demo with Zipprr, or see Zipprr’s AI chatbot solution.

Frequently asked questions

What is Polymath AI?

Polymath AI is an AI product you host yourself. Instead of a single general chatbot, it offers fifteen specialist assistants, a Laravel and Livewire website, a Filament operator console, and a Flutter project for iOS and Android. One backend runs plans, credits, and Stripe payments for all of it.
An expert is a saved record with its own instructions, model, temperature, look, and credit price. Opening one loads that record, so each specialist behaves consistently and keeps its own conversation history. Operators can reshape the roster from the console without waiting for a software release.
It can call OpenAI, Google Gemini, Anthropic, and Stability, depending on the workflow. The operator keeps the vendor keys, so customers never open vendor accounts. One key is the minimum for live AI replies, and each vendor invoices the operator for the usage.
A plan decides which experts a customer may open, and each renewal refills their credits. Every assistant then deducts its own price per reply, so expensive models can cost customers more. Deductions are built so parallel requests cannot draw the same credits twice.
The credits for that reply are returned automatically. Operators should still test awkward cases during acceptance checks, such as a stream that stops halfway, because vendors can fail at different moments and the customer should always see a clear message.
Yes. You can change the product name, slogan, logo, icon, and brand color, and they show up across the web app and landing page. The phone app downloads branding from the server whenever it opens, so everyday changes need no rebuild. Publishing to the app stores still follows the usual store process.
The server needs PHP 8.3 or later, a MySQL 8 or MariaDB 10.6 database, and Composer. Node.js with npm compiles the web assets, and Flutter 3.32 or later builds the phone apps. A guided installer in the browser audits the server, sets up the database, handles activation, and creates the first admin account.
Yes. Both clients talk to one backend, so plans, credits, chat history, and saved images look the same on each. Browsers rely on web sessions while phones store tokens in the operating system’s protected storage. Resetting a password ends the person’s other sessions on every device.

Book Your Meeting

Let’s Talk! Book Your Meeting