# Vercel AI SDK Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/vercel-ai-sdk
> Markdown URL: https://aitinkerers.org/technologies/vercel-ai-sdk.md
> Technology record last updated: 2026-09-18T15:13:38Z
> Generated: 2026-09-21T22:45:27Z

The Vercel AI SDK is the open-source TypeScript toolkit: a unified API for building AI-powered applications with LLMs from providers like OpenAI, Anthropic, and Google.

This is the definitive TypeScript toolkit for building fast, reliable AI applications. The SDK provides a unified API, abstracting away provider-specific boilerplate so you can swap models—like OpenAI's GPT-4o or Anthropic's Claude 3.5 Sonnet—with a single line change. It features real-time response streaming, type-safe structured object generation via `generateObject`, and ready-to-use hooks such as `useChat` for popular frameworks: React, Next.js, and Svelte. Focus on agent logic and user experience; the SDK handles the complex, multi-provider integrations.

- Official technology site: https://ai-sdk.dev
- Public AI Tinkerers demos and talks: 32
- Result page: 1 of 2

## Recent Public Talks and Demos

### [Panel: Building an AI Comic Engine Around the Visual Grid](https://prague.aitinkerers.org/talks/rsvp_Jp8vWlSfz2g)

Panel turns static comics into animated stories with game mechanics. It is a grid-based and AI-assisted engine and editor. It transforms a static page into a structured system of panels, media layers, dialogue bubbles, animations, loops, and reader-triggered interactions. In the demo, I’ll show the current working prototype: the board/canvas view, panel inspector, visual grid editing, media layers, panel-level animation and interaction controls, the animated outline and dialogue system. I’ll also show the first AI layer: generating static images directly into the grid, with support for character/style consistency and reference-based generation. The important part is the UX and how the generated asset lands inside an editable structure. I’ll explain how the system is built around a DOM/React-based visual grid, how GSAP animation and interaction are layered on top of it, and how this structure can evolve into a more advanced AI-native workflow: video generation, layout/story assistance, and eventually an agentic layer that can understand user intent and orchestrate editing steps.

- Event context: AI Tinkerers Prague June Meetup — 2026-06-12 — Prague
- Public talk page: https://prague.aitinkerers.org/talks/rsvp_Jp8vWlSfz2g

### [Deterministic tools, Non-Deterministic orchestration — Building a production AI Agent for auto parts lookup](https://poland.aitinkerers.org/talks/rsvp_8D7UFMN443Y)

An AI agent that takes a car's registration number and returns purchasable replacement parts, solving a problem of a fragmented Polish aftermarket where no single API covers all brands. The system splits the problem into deterministic source selection (which API to call, in what order, per brand) and an LLM-orchestrated agent asks clarifying questions when data is ambiguous, and combines partial results from multiple sources.

- Event context: AI Tinkerers Poland - Meetup in Gdańsk #1 — 2026-04-23 — Poland
- Public talk page: https://poland.aitinkerers.org/talks/rsvp_8D7UFMN443Y

### [From Tool Call to React Component: Building Generative UI for Agentic Workflows](https://san-diego.aitinkerers.org/talks/rsvp_d-XLZk6YEb8)

Live demo of a production pattern where AI agents render interactive React components — forms, profile cards, video embeds — directly inline in chat instead of returning plain text. I'll trigger real agent workflows, show the components appearing in the conversation stream, and walk through the completion callback loop where user interaction feeds back into the agent. I'll close with how we codify these multi-step integration patterns as reusable Agent Skills to ship new tools fast.

- Event context: AI Tinkerers San Diego: February Meetup at Google — 2026-02-27 — San Diego
- Public talk page: https://san-diego.aitinkerers.org/talks/rsvp_d-XLZk6YEb8

### [Building an agentic text creation and editing harness](https://vienna.aitinkerers.org/talks/rsvp_TRbR4hcClQE)

Cursor but for grant applications: It's our mission at StartMatch to make writing grant applications as easy as possible. While generic AI tools are quite capable at generating text, it still takes significant time and expertise to produce texts that meet the strict requirements of grant applications. That's why we built an agentic text creation and editing harness that is able to self- review and correct. Plus, we made the whole thing collaborative and super easy to use.

- Event context: AI Tinkerers Vienna: 2026 — 2026-02-19 — Vienna
- Public talk page: https://vienna.aitinkerers.org/talks/rsvp_TRbR4hcClQE

### [Building intuition around the AI game of telephone](https://manchester-nh.aitinkerers.org/talks/rsvp_rjwp2PakMKg)

An experiment to partially measure semantic and image drift through repeated AI generation and summary cycles. Ship-shape runs multi-round pipelines where AI models repeatedly transform content — writing essays then summarizing them, or captioning images then regenerating them — and measures how meaning and visual fidelity drift at each step.

- Event context: AI Tinkerers – Manchester/Bedford Second Meetup · February 18, 2026 — 2026-02-18 — Manchester NH
- Public talk page: https://manchester-nh.aitinkerers.org/talks/rsvp_rjwp2PakMKg

### [Synthesizing World Aware Agentic Assistants](https://manchester-nh.aitinkerers.org/talks/rsvp_U5CAYIo0BNg)

In this demo, I will be showing an implementation for just-in-time assembly of agentic assistants that is designed for dynamic real-world situations and IoT. The goal is to get contextual agreement by default between the agentic assistant and the user relative to the physical environment (manufacturing station, healthcare clinic room, home office, hotel room, etc.). These environments are dynamic, fragmented, and data/controls for a single physical object tend to exist across multiple systems, so I end up modeling each object then assembling the right agent from these individual object maps and the overall relationships between objects in the physical environment. The focus will be on the resulting assembly manifests that can be executed using any agentic framework with many different orchestration architectures, the condensed physical-digital graph, and the other elements that is passed in through the system prompt to create the contextual agreement by default. Generating the manifest is a longer talk, so the core of what will be shown is the visualizer to make it understandable and the visualizer code for how the system prompt is constructed (surprisingly simple).

- Event context: AI Tinkerers – Manchester/Bedford Inaugural Meetup · January 20, 2026 — 2026-01-20 — Manchester NH
- Public talk page: https://manchester-nh.aitinkerers.org/talks/rsvp_U5CAYIo0BNg

### [The End of Fake News: How AI is restoring trust in media](https://bogota.aitinkerers.org/talks/rsvp_iCS09eSf7ts)

VeloBlanco is the AI-powered co-pilot for fact-checkers and media professionals. We automate the manual bottlenecks of verification—extracting claims, conducting multi-source research, and scoring source credibility—in seconds rather than hours. Equipped with advanced detection for AI-generated audio and video, VeloBlanco handles the heavy lifting of investigation so you can focus on critical analysis and the final editorial decision.

- Event context: AI Tinkerers Bogotá - Nov 2025 — 2025-11-27 — Bogotá
- Public talk page: https://bogota.aitinkerers.org/talks/rsvp_iCS09eSf7ts

### [Torque: Like React, but for Synthetic Datasets](https://sf.aitinkerers.org/talks/rsvp_ZSKhh9gzUN8)

I'll show how with Torque in a matter of minutes you can create a high quality non trivial dataset for building and testing your initial evals: - Building a basic Q&A dataset - Adding tool-calling patterns - Modeling async operations (web searches, long-running tasks) - Scaling from 10 to 1000 examples with one parameter change

- Event context: Evals are Over - Evals are Back - AI in Prod @ Okta — 2025-10-31 — San Francisco
- Public talk page: https://sf.aitinkerers.org/talks/rsvp_ZSKhh9gzUN8

### [Fun Hacks for Chat Interfaces](https://london.aitinkerers.org/talks/rsvp_Xxzvhkvwrfc)

I have been building the chat out at codewords.ai and there are a lot of little hacks i have had to do to make a nice UX - Optimise Vercel's resumable streams to collapse chat chunks and reduce resume load on the FE - We inject extra context into user messages while maintaining cacheability in anthropics APIs - Clean partial conversation data back to a state where it can be persisted and resumed

- Event context: AI Tinkerers London Meetup - 29th October 2025 — 2025-10-29 — London
- Public talk page: https://london.aitinkerers.org/talks/rsvp_Xxzvhkvwrfc

### [Agentic Paper Verifier: Citing, Fact-Checking, and Contradicting Drafts with Hybrid Agents](https://poland.aitinkerers.org/talks/rsvp_SKAYDVw8sLY)

I'll demo a an agent that scans a user-drafted paper for uncited statements, inserts "Citation needed" comments, and spins up background jobs for verification. The TS Vercel AI SDK agent does quick scanning and RAG on user library; unsolved claims trigger Python LangGraph workflows for PubMed searches, relevance scoring, and factuality check. Outputs: inline citations, "couldn't find," or "contradicting evidence [citation]."

- Event context: AI Tinkerers Poland & GDG #1 - Meetup in Wrocław — 2025-10-14 — Poland
- Public talk page: https://poland.aitinkerers.org/talks/rsvp_SKAYDVw8sLY

### [Blocks: Turning Natural Language into Production-Grade React Components](https://london.aitinkerers.org/talks/rsvp_rrlKrwRqEm0)

In this talk, I’ll share how I built Blocks, a system inside Paid.ai that transforms natural language into production-ready React components. Blocks are not just code snippets, each one is a self-contained module (React + CSS) that: * Fetches and normalizes real data * Handles loading, error states, and null-safety * Applies consistent conversions (currency, dates, booleans) * Stays composable within a larger design system I’ll walk through the architectural migration from a brittle, endpoint-based pipeline to an agentic, orchestrated system powered by modular tools, Zod schemas, and real-time telemetry. The session will cover the orchestrator design, workflow composition (ClickHouse queries, SQL persistence, and component generation), and the guarantees that make Blocks safe, predictable, and fast to ship. Attendees will also see examples of Blocks in action: from a natural language prompt, through schema-validated toolchains, to a drop-in UI component.

- Event context: AI Tinkerers x MongoDB - Demopalooza — 2025-10-07 — London
- Public talk page: https://london.aitinkerers.org/talks/rsvp_rrlKrwRqEm0

### [Funny but Scary way to do generative ui](https://dublin.aitinkerers.org/talks/rsvp_xfI0fDHhQ1g)

How to generate ui components on demand with ai agents and module federation and why it might be a bad idea. (Hint: no sandbox)

- Event context: AI Tinkerers Dublin Anniversary Meetup — September 10, 2025 — 2025-09-10 — Dublin
- Public talk page: https://dublin.aitinkerers.org/talks/rsvp_xfI0fDHhQ1g

### [From Chat to Action: Building Reliable AI Agents with Tool Orchestration](https://hong-kong.aitinkerers.org/talks/rsvp_5n3nkrEYh0E)

I’ll demo how I built an AI-native Sales CRM that goes beyond chat. The assistant doesn’t just respond, it can plan, call multiple tools, orchestrate actions in sequence to perform real CRM tasks that free salespeople from busy work, and provide unique insights to close more deals. After a brief live demo, I’ll walk through the technical implementation and underlying architecture: defining tool schemas, managing orchestrated multi-step workflows, and handling of tricky edge cases such as ambiguity (like multiple matching records) and ensuring idempotency when actions are repeated. The focus will be on practical architecture decisions like showing how a conversational interface can drive real database updates, pipeline management/insights, and interaction logging, all without relying on traditional menu-based UIs.

- Event context: AI Tinkerers - Hong Kong Meetup (August) - Meetup with Baidu PaddlePaddle — 2025-08-22 — Hong Kong
- Public talk page: https://hong-kong.aitinkerers.org/talks/rsvp_5n3nkrEYh0E

### [Using Gemini's Large Context and Schema-Enforced Function Calling to Generate Production Kubernetes & Sveltos YAML Configurations](https://sf.aitinkerers.org/talks/rsvp_HXnmVafjK_w)

This demo is a technical deep dive into an AI assistant that generates production-grade Kubernetes and Sveltos configurations from natural language. I'll show how I built it to tackle the complexity of Kubernetes schemas and answer advanced technical questions about Sveltos, a multi-cluster management platform that extends Kubernetes with advanced GitOps capabilities. The core technical challenge is that Kubernetes and Sveltos schemas are massive and complex, requiring large context windows and strict structural enforcement to generate valid YAML. I'll walk through the code and demonstrate how I use a multi-stage agentic workflow with Gemini 2.5 Flash. The system first classifies user intent (Q&A, new config, modification, or out-of-topic) and then routes to a specialized agent. The key innovation I'll showcase is the use of Gemini's function calling capabilities to enforce the complex, nested schemas required for valid Kubernetes resources. This unique approach, combined with Gemini's large context window, makes it possible to reliably generate intricate K8s or Sveltos configurations directly from a simple prompt. For the Q&A portion, the assistant uses a RAG pipeline with Google's embedding models and Google Search Grounding to provide accurate, cited answers from technical documentation.

- Event context: Mastering the Google AI Stack — 2025-08-15 — San Francisco
- Public talk page: https://sf.aitinkerers.org/talks/rsvp_HXnmVafjK_w

### [Learnings from building AI agents](https://london.aitinkerers.org/talks/rsvp_uF6QxcY_VTE)

Quick Summary: Learnings from Building AI Agents for Code Review We built an AI code review agent for cubic that initially produced too many noisy, low-value comments—making code reviews harder instead of easier. After several architecture revisions, we cut false positives by 51% without losing important feedback. I'll explain how

- Event context: AI Tinkerers London Meetup - July 16th, 2025 — 2025-07-16 — London
- Public talk page: https://london.aitinkerers.org/talks/rsvp_uF6QxcY_VTE

### [CTF-Agent: how to hack your site with reasoning models](https://amsterdam.aitinkerers.org/talks/rsvp_BI82WZQHbPs)

An AI agent that completes CTF (capture-the-flag) challenges using reasoning models, an agent loop and only a single tool: HTTP requests. It successfully does SQL injection and several cookie spoofing attacks of increasing complexity.

- Event context: AI Tinkerers x AdvancedJS Amsterdam — 2025-06-25 — Amsterdam
- Public talk page: https://amsterdam.aitinkerers.org/talks/rsvp_BI82WZQHbPs

### [Debugging impossible problems with LLMs](https://la.aitinkerers.org/talks/rsvp_5xUO4Kpzjtk)

I've been prototyping an app which integrates LLMs with code debuggers (like pdb or lldb). These are tools developers use to step through complicated programs line-by-line. The integration makes interactive debugging with an agent in the loop possible.

- Event context: June 9th - LA AI Tinkerers Meetup & Demos — 2025-06-10 — Los Angeles
- Public talk page: https://la.aitinkerers.org/talks/rsvp_5xUO4Kpzjtk

### [Agents, Tools, Chaos: creating emergent multi-agent playgrounds!](https://berlin.aitinkerers.org/talks/rsvp_h-3bMmqdADU)

Building multi-agent simulations is chaotic, rewarding, and unexpectedly fun. I'll demonstrate how I integrated state-of-the-art large language models (LLMs) like GPT-4o and Claude, diverse agent toolkits (reasoning, navigation, communication, web browsing, coding, art generation), and a robust game engine into Squad – a dynamic AI playground. Squad enables autonomous agents to collaborate, compete, and occasionally spiral into fascinating emergent behaviors. The talk covers: Technical Architecture: - Frontend: Next.js, Phaser game engine, Zustand state management - Backend: Cloudflare serverless functions, Vercel AI SDK, Supabase realtime DB Agent Design Patterns: - Practical approaches for integrating diverse AI tools with agents - Strategies for overcoming challenges like agent coordination, memory consistency, and state synchronization Emergent Behaviors: - Real-world examples of complex interactions among tool-rich agents - Live demos showcasing dynamic agent interactions, unexpected collaborations, and humorous breakdowns

- Event context: AI Tinkerers Berlin - June 3rd 2025 — 2025-06-03 — Berlin
- Public talk page: https://berlin.aitinkerers.org/talks/rsvp_h-3bMmqdADU

### [Vibe(?) coded this AI podcast tool with 10K lines of code in one month](https://seattle.aitinkerers.org/talks/rsvp_29nHWDn9rzQ)

1.What I learned in this vibe coding journey: - To build a product not demo: vibe coding is not there yet but could definitely save time - Best model and tool: Gemini 2.5 pro and Claude 3.5+, I used Sublime + Chatbot, not Cursor or Windsurf because of token limit - Some useful tips: Using templates from Vercel, AI SDK is great to add a built-in chatbot (3 files and maybe ~100 lines of code) 2.About this AI podcast tool: - designed for long, conversational podcasts - key features: Search episodes by name or paste URL Generate speaker-separated transcripts from audio Edit speaker names and filter by speaker Auto-scroll and highlight transcript while playing Click any timestamp to jump and play AI chatbot with full podcast context to answer any question

- Event context: Friday - AI Tinkerers Seattle - May Meetup — 2025-05-31 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_29nHWDn9rzQ

### [AI that edits your resume](https://seattle.aitinkerers.org/talks/rsvp_HGAE1liMtbc)

I'll be presenting a resume builder with human-in-the-loop editing that works like Cursor for your resume. The demo walks through the complete workflow from job scraping and resume import to targeted edits and user approval. I'll show how conflict-free AI suggestions get generated with streaming tool-calls and updates the UI. The interface gives users close control with "preview before apply" edits so you can finalize your draft exactly how you want.

- Event context: Friday - AI Tinkerers Seattle - May Meetup — 2025-05-31 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_HGAE1liMtbc

### [Deploying AI agents without the fuss](https://miami.aitinkerers.org/talks/rsvp_bDB_iD7qv4Q)

Prototype agents are easy—running them 24/7 without blowing up your cloud bill isn’t. In this live-coding session we’ll turn a few minimal examples using Crew/Pydantic/Vercel AI SDK into a production-grade service: Pydantic for typed prompts & I/O, the Vercel AI SDK for streaming, and Agentuity’s “Heroku-for-agents” runtime for single-command deploys, autoscaling, logging, and cost controls. Expect real code, edge deploys, and a hard look at things like cross-agent comms, concurrency, and monitoring—no marketing slides, just the engineering path from laptop to live endpoint.

- Event context: AI Tinkerers - Miami: May Meetup at The LAB — 2025-05-29 — Miami
- Public talk page: https://miami.aitinkerers.org/talks/rsvp_bDB_iD7qv4Q

### [Building a lovable and v0 alternative for building streamlit apps](https://montreal.aitinkerers.org/talks/rsvp_UsgEW91TRws)

I'll demonstrate my experimental tool that enables developers to generate and deploy Streamlit applications from single prompts. Similar to how lovable.app, v0, and bolt streamline Next.js app creation, my tool focuses specifically on Python-based Streamlit apps. The demo will walk through the underlying architecture that transforms natural language descriptions into fully functional data applications The technical implementation leverages a custom prompt engineering pipeline with integrated code validation and automatic dependency management. I'll showcase the system's core components: the prompt decomposition engine that breaks complex app requests into modular Python functions, the Streamlit-specific code generator that ensures API compatibility, and the automated deployment system that packages everything into a ready-to-share web application. Rather than focusing on business use cases, this presentation is all about exposing the messy internals of building an AI code generation system specifically optimized for data visualization and interactive analytics

- Event context: AI Tinkerers Montreal - May 2025 Meetup at LightSpeed — 2025-05-07 — Montreal
- Public talk page: https://montreal.aitinkerers.org/talks/rsvp_UsgEW91TRws

### [Ambient Commerce / Why I Learned to Love Agentic Amazon](https://nyc.aitinkerers.org/talks/rsvp_96PPWvC2xvo)

I'll be presenting an MVP of an 'ambient commerce' experience, where a seller agent negotiates with a 'buyer agent' and runs a completely agentic e-comm transaction... onchain.

- Event context: Agents are Taking Over NYC this April with Kamiwaza! — 2025-04-16 — New York City
- Public talk page: https://nyc.aitinkerers.org/talks/rsvp_96PPWvC2xvo

### [PowBlocs - Creating personal desktop software with AI](https://bogota.aitinkerers.org/talks/rsvp_CeLKllT4i10)

PowBlocs is an open-source desktop app that lets you create your own software tools (mini-apps) with AI. These tools have desktop capabilities such as file access and can import packages from npm, JSR, or other external sources. It uses Tauri as the desktop framework and the Deno Runtime for running the backend tasks of the tools in a sandbox. This is a demo of the Tauri/Deno integration I published: https://github.com/carloslfu/tauri-deno-example. The AI part uses Claude and the Vercel AI SDK. I used my knowledge of AI code gen for backends, tex2sql, UI generation, and what I learned from open work such as Aider and articles that reverse engineer V0 and Claude Artifacts to inform the building of PowBlocs. The project is still a work in progress, and I will make the repository public during the presentation.

- Event context: AI Tinkerers Bogotá - January Meetup — 2025-01-30 — Bogotá
- Public talk page: https://bogota.aitinkerers.org/talks/rsvp_CeLKllT4i10

## Related Technologies

- [Next](https://aitinkerers.org/technologies/next) ([Markdown](https://aitinkerers.org/technologies/next.md)) — 186 public demos
- [Claude](https://aitinkerers.org/technologies/claude) ([Markdown](https://aitinkerers.org/technologies/claude.md)) — 174 public demos
- [OpenAI API](https://aitinkerers.org/technologies/openai-api) ([Markdown](https://aitinkerers.org/technologies/openai-api.md)) — 520 public demos
- [React](https://aitinkerers.org/technologies/react) ([Markdown](https://aitinkerers.org/technologies/react.md)) — 220 public demos
- [LangChain](https://aitinkerers.org/technologies/langchain) ([Markdown](https://aitinkerers.org/technologies/langchain.md)) — 445 public demos
- [Langfuse](https://aitinkerers.org/technologies/langfuse) ([Markdown](https://aitinkerers.org/technologies/langfuse.md)) — 13 public demos
- [Supabase](https://aitinkerers.org/technologies/supabase) ([Markdown](https://aitinkerers.org/technologies/supabase.md)) — 90 public demos
- [TypeScript](https://aitinkerers.org/technologies/typescript) ([Markdown](https://aitinkerers.org/technologies/typescript.md)) — 205 public demos
- [Anthropic](https://aitinkerers.org/technologies/anthropic) ([Markdown](https://aitinkerers.org/technologies/anthropic.md)) — 36 public demos
- [ChatGPT](https://aitinkerers.org/technologies/chatgpt) ([Markdown](https://aitinkerers.org/technologies/chatgpt.md)) — 83 public demos
- [Drizzle ORM](https://aitinkerers.org/technologies/drizzle-orm) ([Markdown](https://aitinkerers.org/technologies/drizzle-orm.md)) — 3 public demos
- [Ollama](https://aitinkerers.org/technologies/ollama) ([Markdown](https://aitinkerers.org/technologies/ollama.md)) — 77 public demos
- [Playwright](https://aitinkerers.org/technologies/playwright) ([Markdown](https://aitinkerers.org/technologies/playwright.md)) — 32 public demos
- [Tauri](https://aitinkerers.org/technologies/tauri) ([Markdown](https://aitinkerers.org/technologies/tauri.md)) — 8 public demos
- [Agentuity](https://aitinkerers.org/technologies/agentuity) ([Markdown](https://aitinkerers.org/technologies/agentuity.md)) — 1 public demo
- [Anthropic API](https://aitinkerers.org/technologies/anthropic-api) ([Markdown](https://aitinkerers.org/technologies/anthropic-api.md)) — 66 public demos
- [Artifacts UI](https://aitinkerers.org/technologies/artifacts-ui) ([Markdown](https://aitinkerers.org/technologies/artifacts-ui.md)) — 1 public demo
- [AWS Lightsail](https://aitinkerers.org/technologies/aws-lightsail) ([Markdown](https://aitinkerers.org/technologies/aws-lightsail.md)) — 1 public demo

## More Results

- Next: https://aitinkerers.org/technologies/vercel-ai-sdk.md?page=2