How I built a switchable AI provider system in Next.js (Gemini GPT-4 Claude, one workspace setting)

Кирилл Несмелов — Wed, 10 Jun 2026 10:38:52 +0000

I had Gemini hardcoded into every AI function in my codebase. Then a beta user told me her clients have data policies about Google. I rebuilt the whole AI layer in two days. Here's the architecture.

The wrong way (what I had)

import { GoogleGenerativeAI } from "@google/generative-ai";
const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY!);

export async function summarizeProject(projectId: string) {
const model = genAI.getGenerativeModel({ model: "gemini-1.5-flash" });
const result = await model.generateContent(prompt);
return result.response.text();
}
Works fine until you need to support another provider. Then you're touching every function.

The pattern: one factory, one interface

I defined a single AIClient interface that every provider implements:

export interface AIClient {
complete: (prompt: string, systemPrompt?: string) => Promise;
}
One factory function reads the workspace setting and returns the right client:

export async function getAIClient(workspaceId: string): Promise {
const { aiProvider } = await getWorkspaceSettings(workspaceId);

switch (aiProvider) {
case "openai": return createOpenAIClient();
case "anthropic": return createAnthropicClient();
default: return createGeminiClient();
}
}
Each factory wraps the provider SDK behind the same interface:

export async function summarizeProject(projectId: string, workspaceId: string) {
const ai = await getAIClient(workspaceId);
const data = await getProjectContext(projectId);
return ai.complete(buildPrompt(data), "You are a project analyst.");
}
Three decisions worth explaining

Workspace-level, not user-level. Project data belongs to the workspace. The whole team should have the same answer to "what AI provider processes our data?" — not "depends who's logged in."

No silent fallbacks. If the selected provider fails, the error propagates. I don't retry with another provider. If someone chose Claude for compliance reasons, silently falling back to Gemini is worse than an explicit error.

Per-request factory call, not module-level cache. Workspace settings can change. Module-level caching means the old provider keeps running until the function cold-starts. The extra DB read per AI call is negligible next to the actual inference cost.

The one thing the pattern doesn't solve

Different models behave differently on the same prompt. Claude follows system prompts very tightly. GPT-4 tends verbose without explicit length instructions. I handle this with thin per-provider prompt adjustments in the factory. It's not elegant but it works.

If you've solved the "same prompt, different model behavior" problem more cleanly — I'm curious how.

Building Melororium — pay-once project management for freelancers. Launching July 30.

Day 7: 51 people answered my question. I wasn't ready for what they said.

Кирилл Несмелов — Mon, 08 Jun 2026 13:18:22 +0000

Day 7: 51 people answered my question. I wasn't ready for what they said.
Last week I asked: "What convinced you to switch away from a subscription tool — or what's keeping you locked in?"
51 people answered. I read every single one.
I expected arguments about SaaS economics. Or "subscriptions are fine, you get what you pay for."
Instead, almost every answer was some version of the same thing:
"I already left. I use free Notion + Trello + a spreadsheet I built myself. I stopped paying because I refused to keep feeding the machine."
Nobody is sitting around waiting for a better subscription-free tool. They've already quit — and they're living with the consequences of their DIY stack. Things that don't connect. Manual exports every Friday. Three browser tabs open at once just to see where a project stands.
That one insight shifted how I think about who I'm building https://melororium.com for.
It's not someone who's currently paying $50/month for ClickUp and angry about it. It's someone who already said "no" — months ago — and built their own janky workaround. I'm building for the person who quit before I showed up.
So let me show you what I've actually built.
This is the part where I want to be specific, because "almost done" is the oldest lie in indie building.
I recorded a short walkthrough this week — nothing polished, just the real product as it exists right now:
https://youtu.be/Sa35DwMXAOQ
Here's what you'll see working:
Kanban board — full drag-and-drop, columns, task cards, assignees. Not a prototype. The thing I actually use every day to run my own work. You can move a task from "In Progress" to "Done" and it updates across every view instantly.
Tasks — complete system. Priorities, deadlines, subtasks, descriptions, comments. The boring stuff that actually matters when you're juggling 6 clients and can't afford to forget anything.
Clients — bring your clients in, attach projects to them, see everything related to one client in a single view. No more searching through Notion pages or scrolling through Slack to find what you promised someone.
Team members — add the people you work with. Assign tasks to them. See who has too much on their plate before they send you a panic message at 11 PM.
Work hours analytics — full time tracking with real reports. Not just "hours logged" — actual breakdown by client, by project, by task. The numbers a freelancer or small agency actually needs to answer: am I making money on this client or just staying busy?
Calendar integration — deadlines and tasks sync so your week looks like one thing, not three disconnected apps all demanding attention at the same time.
We're on the finishing straight. The core is done. Right now I'm working through the last small fixes before the founding release — and building one more thing.
What I'm building next:
A direct AI assistant integration. Not "AI features" as a marketing checkbox — a way to create tasks, log work, and get a weekly project summary just by talking to it.
The goal: you finish a client call, you tell the assistant what was discussed, the workspace updates itself. No context switching. No manual note-taking after every conversation.
I'm building this because I tracked my own time one week and found I was spending 5+ hours just on admin — logging work, reorganizing tasks, writing status updates that nobody reads. That time should not exist.
The wishlist is open right now.
While the product isn't publicly released yet — founding access opens July 30 — the waitlist is live at https://melororium.com
If you reserve a spot now, you lock in the founding price: $199 one-time. No subscription. Ever.
After launch the price goes up permanently. I'm not doing artificial scarcity — I just can't keep the founding price open indefinitely once real support costs kick in.
We also built a referral system this weekend: bring 1 paying friend → 5% off your price. Bring 5 → 10%. Bring 30 → 50%. All the discount stacks get applied at launch in July. Full pricing breakdown at https://melororium.com/pricing
Day 7 real numbers:
2 clicks from Google. First organic visitors ever. The site is already appearing at position 5 on some queries after 7 days — which genuinely surprised me
Product Hunt launch: Tuesday June 9. Auto-scheduled. No PR, no budget, no influencer outreach. Just the story and whatever this community thinks of it
0 paying customers. Still intentional. I want founding members who understand what they're getting into, not impulse purchases
40+ published articles on the site. Probably only 8 of them are indexed by Google so far. The rest are sitting there waiting to be discovered
The scary part about being this close to done: "done" isn't really a thing in software. You ship, people use it, something breaks, something's missing, something that seemed obvious to you makes no sense to anyone else.
The real work starts July 30.
But I know this: the people who've been running DIY stacks for two years — free Notion plus a spreadsheet plus three browser tabs — are going to open this and immediately understand it. No pitch needed. No onboarding email sequence. Just: oh, this is what I was trying to build myself.
That's who I built it for.
Question for this week: If you're running a DIY productivity stack right now — what's the one thing about it that drives you the most crazy? The thing you've accepted as "just how it is" but secretly hate?

DEV Community: Кирилл Несмелов

How I built a switchable AI provider system in Next.js (Gemini GPT-4 Claude, one workspace setting)

Day 7: 51 people answered my question. I wasn't ready for what they said.