Building a Private AI Service That Charges Per Query

#ai #privacy #api #ethereum

Building flat.cash/ask: An AI Service Built on Zero Logs and Flat-Rate Micro-Billing

When we set out to build flat.cash/ask, we had a straightforward thesis: Using an AI assistant shouldn't mean trading away your digital privacy.

Most consumer AI platforms today operate on a data-harvesting model. Your queries, your uploaded documents, your code snippets, and your personal dilemmas are logged, analyzed, tied to your identity, and often used to train future model iterations.

For casual queries, that might feel like an acceptable tradeoff. But what happens when you’re pasting medical test results, reviewing a sensitive legal contract, or debugging proprietary financial logic?

Here is how we built flat.cash/ask—an AI service designed from the ground up for strict confidentiality, featuring zero retention, a predictable flat-rate micro-payment model, and an ultra-resilient multi-model backend.

Why Privacy Matters for AI Queries

Developers, researchers, and professionals are increasingly using LLMs for high-stakes tasks. Consider the categories of data leaking into mainstream AI systems daily:

Medical Questions: Pasting lab results, symptoms, or medication interactions to get a preliminary breakdown. This is sensitive Protected Health Information (PHI).
Legal Queries: Reviewing Non-Disclosure Agreements (NDAs), contract clauses, or employment terms.
Financial Queries: Analyzing cap tables, parsing bank exports, or debugging proprietary tax strategies.

When you send these prompts to standard providers, they hit servers that log requests for safety filters, telemetry, and model training. Even with enterprise "opt-outs," managing compliance across third-party dashboards is tedious.

We wanted a tool where privacy is absolute by architecture, not by policy.

The Architecture: Zero Logs, Zero Retention

To achieve true privacy, you can’t just promise not to look at the data—you have to design a system that literally cannot save it.

Stateless Handlers: Our backend API routes are completely stateless. When a request hits flat.cash/ask, it is parsed in memory, handed to the LLM provider, streamed back to the client, and immediately expunged.
No Database Logging: We do not maintain chat history databases, vector stores of user inputs, or request audit logs tied to user profiles. Once the HTTP response finishes streaming, the data ceases to exist on our servers.
Ephemeral Context: If you refresh the page, the state is gone. There is no "sidebar history" tracking your deepest questions over months of usage.

The AI Backend: Mistral-Primary with Forge Fallback

Reliability is just as critical as privacy. A privacy-focused tool that goes down when you need it is useless. We designed a resilient, multi-tiered AI routing pipeline:

Primary Route (Mistral): We route initial requests through Mistral’s high-performance models. They offer incredible reasoning capabilities, handle dense technical text gracefully, and align well with our performance and efficiency requirements.
Fallback Route (Forge): Infrastructure fails. API rates spike, providers go down, and networks timeout. If our primary Mistral route experiences latency spikes or errors, our middleware instantly fails over to a secondary Forge-backed endpoint without dropping the user's connection or interrupting the stream.

Because both layers operate under strict zero-retention guidelines, switching backends mid-stream never compromises user data security.

How Payment Works: 1 FLAT per Query via FlatID

Subscription fatigue is real. Nobody wants to enter a credit card for a $20/month recurring subscription just to ask three AI questions a week. At the same time, traditional credit card processing fees ($0.30 + 2.9%) make micro-transactions economically unviable for single-digit-cent API calls.

We solved this using FlatID balances.

The Pricing: It’s refreshingly simple. 1 FLAT per query. No hidden token math, no fluctuating prompt-token multipliers, and no surprise bills at the end of the month.
FlatID Balance Deduction: Users fund a FlatID balance ahead of time. When a query is successfully dispatched, our billing service performs an atomic, instant deduction of 1 FLAT from the user's balance ledger.
Frictionless Micro-Billing: By using an internal balance ledger rather than hitting a payment gateway per query, we bypass network overhead and transaction fees, making a flat-rate model sustainable.

Under the Hood: A Quick Look at the Flow

Here is the high-level flow of what happens when a user hits flat.cash/ask:

[Client UI] 
    │ (Sends Prompt + FlatID Token)
    ▼
[Auth & Ledger Check] 
    │ (Verifies balance >= 1 FLAT, deducts atomically)
    ▼
[API Router] 
    ├── Tries [Mistral Primary] 
    └── On Failure ──► Falls back to [Forge Fallback]
    │ 
    ▼ (Zero logs written to disk)
[Stream Response to Client]
    │ 
    ▼
[Connection Closed - Memory Cleared]

Try It Out

Building an AI service that respects user privacy doesn't require over-engineered enterprise compliance layers—it requires ruthless minimalism in data handling and a transparent business model.

If you want to test out an AI assistant that doesn't hoard your prompts, check out flat.cash/ask. Load up your FlatID balance, ask your hardest questions, and rest easy knowing your data vanishes the second the stream ends.

DEV Community