OpenAI Whisper, React, and FastAPI: how to build an AI meeting summarizer with no budget

Developers no longer need a paid stack to build a useful AI application: a detailed breakdown of a meeting summarizer built with React and FastAPI has been released, relying only on free models and services. As an example, the author takes a service that accepts audio, transcribes the recording, extracts key points, and displays action items in a simple web interface.

Why Zero Budget Is Real

A couple of years ago, such a project almost automatically meant expenses for OpenAI API, GPU rental, and paid deployment. The situation is different now: open-source models and generous free tiers cover most basic scenarios, especially if the task is narrow and well-defined. For meeting summaries, this is particularly notable because here you don't need abstract "superpowers" of a model, but normal transcription handling, extracting decisions, and a list of next steps.

The breakdown specifically emphasizes that the developer is no longer tied to a single vendor. You can take a cloud-based free model or switch to local execution if privacy, data control, and predictable costs matter. For this, local tools like Ollama and LM Studio are provided, and for the cloud option — free quotas from API providers. Essentially, building an MVP is now possible without infrastructure procurement and lengthy budget approvals.

What the Stack Consists Of

The project is built on a maximally simple set of tools. The logic here is not in the "trendiest" stack, but in making it possible for any developer to quickly replicate the solution, get a working result, and if desired, replace individual components without rebuilding the entire system.

• OpenAI Whisper — for converting audio to text locally and without paying for requests
• GLM-4.7-Flash from Zhipu AI — as a free cloud option for summaries
• LFM2-2.6B-Transcript from Liquid AI — as a local model for meeting notes
• FastAPI — for API file upload, processing, and result storage
• React + SQLite — for the interface and storing transcripts, summaries, and action items

In addition, the material recommends free AI development tools, primarily Codeium and Continue, to write and edit code faster. But the architecture of the project itself remains very straightforward: minimum layers, minimum dependencies, and a clear data flow from audio file to a card with the final summary. For an educational project or internal prototype, this is more important than an "ideal" engineering scheme.

How the Project Is Built

The pipeline is arranged without unnecessary complexity. A user uploads a meeting recording, lecture, or voice note, FastAPI receives the file and passes it to Whisper, which builds a transcript. Next, the text is sent to a summarization model, which returns a short description of the discussion and a list of action items. After that, the result is saved in SQLite, and the React interface displays the transcript, summary, and tasks on one screen. The complete code for this pipeline is indeed provided directly in the material by the author.

In the example, Whisper runs in tiny configuration to speed up processing on CPU, and for the LLM two modes are offered: a free cloud option through the Zhipu AI API and completely local through the Liquid AI model, which requires less than 3 GB of RAM. This makes the project flexible: you can start with the cloud and then switch to a local scenario without breaking the rest of the architecture.

"If one model doesn't fit, you can switch to another without changing

the infrastructure."

After local assembly, the project is offered to be deployed for free on Vercel and Render. This option is suitable for a demo, internal tool, or early user testing, but the author honestly warns about the limitations: Whisper and transformers take up significant disk space, and free tiers quickly hit memory and startup time limits. Therefore, for production, you'll almost certainly need to either move some tasks to a cloud API or prepare separate infrastructure for local models.

What This Means

The practical conclusion is simple: an AI prototype no longer requires a separate team and model budget from the start. For indie developers, small studios, and product teams, this reduces the cost of experimentation — an idea like a meeting summarizer, internal copilot, or voice-to-notes service can now be tested in a couple of evenings and only then decide if it's worth paying for scaling.