AI Tinkerers - "One-Shot"

What if building complex AI agents felt as natural as composing React components—and they could even rewrite their own code? 🤯

In this episode of One Shot / AI Tinkerers, host Joe sits down with Evan Boyle, founder of GenSX, to explore a radically new way to design, run, and ship long-running agent workflows:

🔑 Key takeaways

- React-inspired component model for agents – why JSX-style, type-safe functions beat static graphs for scalability and code reuse.

- Traces, telemetry & evals baked-in – see every prompt, variable, and LLM call in real time.

- $4 self-modifying coding agent – Evan demos an agent that checks out its own repo, refactors 3 K lines, runs tests, and pushes to GitHub… iteratively.

- Real-world production use cases – from million-document legal discovery to inbox-wide entity extraction and analytics.

- Durable execution & infra shift – why 5-second latencies and massive parallelism are forcing a rethink of serverless, queues, and caching.

- Developer experience first – faster dev loops with component-level caching, cursor rules, and LLM “rubber-duck” debugging tricks.

🛠️ Tools & frameworks mentioned

GenSX, React/JSX, OpenAI & Anthropic models, Temporal, Pulumi, Cursor, LangChain, LlamaIndex, Crew AI…and more.

🔗 Try GenSX → https://www.gensx.com

💬 Join the community → https://github.com/gensx-inc/gensx

🐦 Follow Evan on X/Twitter → https://x.com/_Evan_Boyle

🙌 Enjoyed the conversation?

👍 Like, 🔔 subscribe, and drop your questions or aha moments in the comments. It helps more builders discover the pod!

📍 Chapters

00:00 Intro & Evan’s background

04:28 Why existing agent frameworks break at scale

12:55 Inside the React-style component model

23:10 Live demo: Hacker News Analyzer (1,000 LLM calls in parallel)

32:45 Tracing, telemetry, and evals

38:20 The self-modifying code agent ($4/iteration)

50:40 Real production agent use cases

59:05 Dev-tooling tips: caching, logging-only debug loops

1:08:30 The future of AI infrastructure & closing thoughts

#GenSX #AIAgents #DeveloperExperience #React #SelfModifyingCode #AIWorkflow #OneShotPodcast

While most people are still asking ChatGPT to write code snippets, Kevin Leneway is building full-stack products using nothing but prompts.

In this One-Shot episode, he reveals the exact system he’s used to launch over 40 startups at Pioneer Square Labs.

We break down:

- How he writes BRDs and PRDs that don’t suck

- Why vibe coding fails and how to actually use AI agents

- The markdown checklist that replaces a product team

- How to go from idea to working app with zero context switching

- His open-source starter kit that makes Cursor and Claude 3.5 feel like magic

If you’re a builder, this will change how you work.

No gimmicks. Just a ruthless focus on speed, clarity, and shipping.

Watch now. Learn the system. Steal it.

Check out more information like this at Kevin's YouTube: https://www.youtube.com/@kevinleneway2290

Also, check out Kevin's Github:

http://github.com/kleneway/next-ai-starter

NextAI Starter Repo: https://github.com/kleneway/next-ai-starter

BRD/PRD/Checklist prompts: https://chatgpt.com/share/67c5ee78-0b94-800d-b467-ceecdbf6ce70

Agent Tasklist example: https://gist.github.com/kleneway/07432638aeaf6210316ebbc32dfbe643

Storybook Link: https://storybook.js.org/

UX Rubric Example: https://github.com/kleneway/pastemax/blob/main/docs/ux-rubric.md

PasteMax Open Source Repo: https://github.com/kleneway/pastemax/

While most people are still asking ChatGPT to write code snippets, Kevin Leneway is building full-stack products using nothing but prompts.

In this One-Shot episode, he reveals the exact system he’s used to launch over 40 startups at Pioneer Square Labs.

We break down:

- How he writes BRDs and PRDs that don’t suck

- Why vibe coding fails and how to actually use AI agents

- The markdown checklist that replaces a product team

- How to go from idea to working app with zero context switching

- His open-source starter kit that makes Cursor and Claude 3.5 feel like magic

If you’re a builder, this will change how you work.

No gimmicks. Just a ruthless focus on speed, clarity, and shipping.

Watch now. Learn the system. Steal it.

Check out more information like this at Kevin's YouTube: https://www.youtube.com/@kevinleneway2290

Also, check out Kevin's Github:

http://github.com/kleneway/next-ai-starter

NextAI Starter Repo: https://github.com/kleneway/next-ai-starter

BRD/PRD/Checklist prompts: https://chatgpt.com/share/67c5ee78-0b94-800d-b467-ceecdbf6ce70

Agent Tasklist example: https://gist.github.com/kleneway/07432638aeaf6210316ebbc32dfbe643

Storybook Link: https://storybook.js.org/

UX Rubric Example: https://github.com/kleneway/pastemax/blob/main/docs/ux-rubric.md

PasteMax Open Source Repo: https://github.com/kleneway/pastemax/

In this episode of the “New AI Tinkerers Podcast,” Joe sits down with Tanya Verma—co-founder of Tinfoil and fresh admit to Y Combinator—to explore how confidential computing is changing the game for AI privacy and security.

Tanya explains the concept of “secure enclaves,” hardware-based encryption that keeps your data safe even from the cloud provider.

Tanya also discuss real-world use cases for private AI, how open-source models like DeepSpeed Chat stack up, and why privacy is crucial for large-scale adoption of generative AI.

Whether you’re building a high-stakes internal chatbot or just want to keep your personal data secure, this conversation will help you understand how Tinfoil’s technology ensures no one can peek under the hood—not even Tinfoil itself.

Don’t miss their deep dive into the future of confidential computing, agent-based automation, and the challenges of scaling secure AI solutions.

Highlights & Takeaways

• What confidential computing really is—and why it matters

• How Tinfoil is bringing secure enclaves to generative AI

• The trade-offs between on-prem vs. cloud-based AI solutions

• Why open-source models and privacy go hand in hand

• Practical steps for deploying private AI applications at scale

Subscribe for more interviews with builders and visionaries in AI—and get an insider’s look at the next generation of secure and private computing!

Vik from Moondream AI joins Joe to demo a vision-language model that runs locally—on your laptop, your phone, even a Raspberry Pi.

From visual question answering to gaze detection and UI automation, Vik shows how Moondream is redefining edge computer vision—no cloud required.

Whether you're into robotics, home automation, or lightweight AI, this “One-Shot” is packed with insights for builders.

Try it yourself at moondream.ai 🚀

00:00 - Intro to Moondream’s compression tech for 2B parameter models

00:22 - Joe welcomes Vik from Moondream

01:53 - Shift from traditional CV to promptable vision-language models

03:23 - Playground demo: Visual Question Answering (VQA)

04:42 - VQA demo results: speed, structure, and accuracy

05:03 - Object detection, pointing, and captioning demos

07:57 - Prompts that push reasoning: uniform detection

10:07 - Cross-task benefits: gaze detection improves directional reasoning

11:02 - Comparing Moondream’s VQA to Quinn’s visual reasoning model

13:21 - Why edge deployment still matters in vision

15:21 - 0.5B model runs on Raspberry Pi using 816MB with int4

16:15 - HAL 2000 setup: Moondream + Tiny LLaMA + Coqui TTS

20:55 - Texas rancher uses drone and Moondream for cow detection

21:53 - Commercial use: air-gapped environments like retail, safety

23:09 - UI automation and button detection with pointing feature

29:41 - Vision for ambient agents and local inference

30:27 - Power efficiency: 10x less energy than 7B/20B cloud models

31:01 - Moondream API & Hugging Face transformers integration

36:11 - Vik’s background: From AWS to machine learning

40:56 - Discovering AI Tinkerers and global meetups

#AITinkerers #MoondreamAI #EdgeAI #ComputerVision #LLM #OpenSource #OneShot

AI agents are evolving fast, but what if they could seamlessly integrate with real-world tools, automate workflows, and take action—without breaking security or reliability?

In this One-Shot episode, we sit down with Karan from Composio, a cutting-edge AI framework designed to connect LLMs with APIs, streamline tool orchestration, and unlock new levels of agent autonomy.

With over 250+ API integrations, Composio is solving one of the biggest hurdles in AI automation: getting things done—securely, efficiently, and at scale.

Three mind-blowing takeaways from this episode:

⚡ LLMs Alone Won’t Cut It:

"So essentially, if you think about it, agents are kind of LLMs plus tools, right?" – Karan, Composio

💡 From Browsers to APIs: The Smarter Path to AI Execution:

"Like you don’t want everything to be controlled by a browser, right? Like these LLMs are stochastic models. They can go wrong." – Karan, Composio

🚀 The Future of Autonomous AI Assistants:

From AI-powered SDRs and personal assistants to automated data management and software development, Composio is helping AI agents go from passive responders to proactive problem solvers.

Karan breaks down how Composio handles authentication, authorization, and multi-tenant support, making it a must-have tool for anyone building AI-powered automation.

But here’s the kicker…

The future of AI isn’t just about answering questions—it’s about doing real work. Whether you’re building AI-powered personal assistants, automated sales workflows, or next-gen productivity tools, Composio is paving the way for AI agents that truly get things done.

🔗 Watch this One-Shot episode and discover how Composio is reshaping AI automation.

Welcome to the future of AI-driven action. Welcome to AI Tinkerers.

What if AI could talk, react, and even mirror human emotions in real-time? In this One-Shot episode, we explore how Tavus is pushing the boundaries of AI-driven avatars—building an experience so fast and fluid, it feels like a real conversation.

Brian from Tavus has spent months refining the latency, speech processing, and personality layers of AI avatars, bringing us closer to seamless, interactive digital humans. Meet Brad—an avatar that doesn’t just listen, but truly engages.

Here’s what makes this episode unmissable:

1. AI Conversations Without the Lag – Tavus has cracked the sub-second response time, making interactions feel instant and natural.

2. Building Personalized AI Personas – With Tavus’ API, developers can customize avatars, define personalities, and integrate them into real-world applications.

3. The Future of Digital Assistants – From virtual customer service to education and beyond, AI avatars are redefining how we interact with machines.

Brian takes us under the hood, explaining how AI speech gating, low-latency LLMs, and streaming architecture come together to make these avatars feel real.

But this is just the beginning…

AI avatars are evolving fast. Whether it’s sales, education, personal assistants, or even healthcare, real-time AI-driven conversations are about to become mainstream.

Watch now and see how Tavus is building the future of human-AI interaction.

Welcome to the new era of AI avatars. Welcome to AI Tinkerers.

Highlights:

00:00 - Introduction and Welcome

00:43 - Meet Brian and His AI-Powered Avatar

01:17 - Introducing Brad, the Conversational Video Interface

02:03 - How the Avatar Processes Audio and Video

03:06 - Latency Challenges in AI Conversations

04:07 - Optimizing Speech Processing for Real-Time Response

05:47 - The Role of ASR and Low-Latency Models

06:45 - API Setup for Custom AI Conversations

07:46 - Integrating Real-Time AI in Everyday Use Cases

08:45 - Unexpected Delight in Fast AI Responses

09:47 - The Future of Virtual Avatars and Trust

10:27 - Configuring LLM and API for Custom AI Interaction

11:43 - Balancing Function Calls and Response Speed

12:53 - Designing AI Conversations for User Experience

13:41 - Emerging AI Business Use Cases

15:30 - AI in Sales and Marketing

16:48 - The Role of AI in Education and Healthcare

17:54 - AI as a Virtual Assistant for Professionals

18:11 - Brad on Mimicking Human Conversations

19:15 - Bridging the Uncanny Valley in AI Avatars

20:14 - Pipeline Behind AI Conversations and Visuals

21:34 - How AI Processes Audio and Generates Expressions

22:08 - Critical Factors for Natural AI Conversations

23:10 - The Future of AI in Professional and Social Spaces

25:27 - Ensuring AI Safety and Preventing Misuse

26:22 - Brian’s Drone Project for Illegal Dumping Detection

27:56 - How AI-Powered Drones Help San Francisco

28:47 - Automating City Clean-Ups with AI

29:37 - Hacking Systems for Social Good

30:10 - Future of AI and Open Source Innovations

In this explosive “One-Shot” episode, we’ll dismantle the complexities of building a flying Linux computer—GPU, 5G, real-time inference, and all—and see how voice commands can orchestrate everything from speed adjustments to target tracking.

Maxwell Wang, SpaceX software engineer, created Dispatcher, an end-to-end aerial AI system that’s turning sci-fi ambitions into practical, weekend-ready tech.

Imagine telling a drone, “Back up 50 meters,” and watching it respond instantly—no clunky menus, no joysticks. That’s the magic of Dispatcher: a system that blends local compute for real-time tasks with cloud-powered LLMs for semantic understanding.

The result?

A frictionless, voice-driven aerial experience that scouts parking lots, follows targets, and returns home—all in one fluid operation.

Three jaw-dropping takeaways that’ll make you hit “Play” on the episode:

1. Real-Time AI on the Edge:

Maxwell’s drone runs powerful vision models locally for live object tracking—say goodbye to lag from iffy connections.

2. Voice Commands, Simplified:

Want the camera to tilt up or the drone to land? Just say it. Dispatcher’s function-calling infrastructure translates speech into drone actions.

3. Hybrid Autonomy is Here:

By splitting computation between on-board GPU and the cloud, Maxwell shows how voice-driven drones can go far beyond line-of-sight or simple waypoint flying.

The technical depth in this episode will blow your mind. Maxwell lays out how he integrates hardware, IoT protocols, and language models into a seamless drone orchestration layer—all without needing an army of developers.

But here’s the best part:

Dispatcher isn’t just a demo—it’s a sign of things to come. Voice-driven drones are poised to revolutionize everything from disaster relief to autonomous inspections.

Watch this “One-Shot” episode and discover a future where talking to machines is as easy—and thrilling—as it sounds.

Welcome to the dawn of drone-based AI. Welcome to AI Tinkerers.

Highlights:

00:00 - Introduction and Welcome

00:27 - Meet Maxwell and His AI-Powered Drone

01:17 - Drone Dispatcher Demo

02:47 - Breaking Down the Drone's Architecture

04:50 - Hardware Components and Challenges

06:48 - Software and Communication Protocols

10:10 - Vision Models and Real-Time Processing

12:55 - Project Challenges and Iterations

20:08 - Getting Started with Drone Hacking

20:35 - Maxwell's Background and Future Plans

22:10 - Maxwell’s Background & Inspiration

24:40 - Wrapping up

25:27 - Call for a co-founder

🚀 What If Your AI Could Talk Back in Real-Time? Inside Daily’s Voice Revolution!

Join Kwindla Kramer, CEO of Daily, as he unveils the next frontier of conversational AI that’s about to transform how we interact with technology.

In this exclusive AI Tinkerers “One-Shot”, discover how real-time voice interfaces are breaking down barriers across industries—from medical support to customer service.

Key Breakthroughs:

✨ Ultra-Low-Latency Conversational Agents: Rapid voice-to-voice pipelines that feel truly human, optimizing for sub-second response times. Sub-800 millisecond AI responses that feel truly human

✨ Live Orchestration and Context-Aware Tools: Dynamic pipelines that integrate speech-to-text, AI reasoning, text-to-speech, and tool calls—enabling complex tasks like scheduling appointments or handling call center workflows

✨ Flexible, Open-Source Ecosystem: Leveraging Pipecat’s modular design, developers can easily build, modify, and scale voice-based experiences, tapping into open innovation across industries from telehealth to global enterprise support

Kwindla's vision expands the boundaries of what’s possible with real-time AI voice interfaces—unifying natural conversation, contextual awareness, and robust tooling into a single, lightning-fast layer of communication.

From MIT Media Lab to pioneering real-time interaction technologies, Quinn demonstrates how generative AI is creating intelligent, context-aware conversational experiences that adapt to human communication in real-time.

Witness live demos of voice agents powered by Claude and GPT-4, and explore use cases that will make you reimagine the potential of AI-driven interactions.

Watch more at: https://one-shot.aitinkerers.org

00:00 Introduction and Customer Feedback

00:51 Welcome to AI Tinkerers One-Shot

01:19 Introducing Kwindla and Daily's Journey

02:38 Daily's Innovations

03:19 Impact of COVID and Telehealth

04:40 Generative AI and Real-Time APIs

06:28 Voice to Voice Interfaces and Pipecat

09:04 Technical Deep Dive into Pipecat

12:04 Live Demonstration of Pipecat

16:13 Future of Multimodal AI and Pipecat

26:27 Exploring Innovative Use Cases

29:17 Voice Interfaces Revolutionizing Call Centers

31:40 Voice-Driven Drones and Hybrid Systems

32:51 The Future of Conversational AI

39:31 Challenges and Opportunities in Scaling AI

42:14 Voice Cloning and Security Concerns

46:20 Emerging Gen AI Startups

52:00 Building Fluid User Interfaces with AI

53:21 Reflections and Future Directions

#VoiceAI #ConversationalAI #AIDevelopers #NextGenTech

Imagine an AI companion that understands your body better than your annual check-up ever could. At the latest AI Tinkerers "One-Shot" session, Varun Pant—AWS engineering manager and robotics innovator—discusses his AI Tinkerers Hackathon Grand Prize winning project, SmartNourish, a groundbreaking human-in-the-loop health monitoring system targeting the 1 in 3 Americans at risk of diabetes.

Key Breakthroughs:

✨ Real-time Metabolic Intelligence: Using continuous glucose monitoring and advanced AI, the system provides personalized, predictive health insights

✨ Proactive Health Companion: An AI agent that learns your unique metabolic responses, suggesting precise lifestyle and nutrition recommendations

✨ Dynamic Human-in-the-Loop Design: Intelligently escalates critical health signals, potentially involving nutritionists or clinicians

Pant's innovation demonstrates how AI can transform healthcare from reactive to predictive, leveraging technologies like LangChain, LangGraph, and Anthropic's Claude models to create a truly personalized health monitoring experience.

🔬 Developed by an engineer with a background in space robotics and machine learning, SmartNourish represents the cutting edge of AI-powered personalized medicine.

Watch more at: https://one-shot.aitinkerers.org

0:00 Welcome to AI Tinkerers One-Shot

0:28 Intro to Human-in-the-Loop Agents

1:34 Varun's Background in Robotics

2:51 Defining AI Agents and Workflows

4:00 Human-in-the-Loop: Key Concepts

6:26 SmartNourish Project Demo

8:36 CGM Data Dashboard Walkthrough

10:22 Understanding Glucose Spike Patterns

12:06 Tech Stack Overview [code walkthrough]

13:14 Agentic Workflow Design [code walkthrough]

17:16 LLM Model Selection Insights

18:14 Agent Framework Exploration

25:40 Hackathon Development Challenges

26:07 AI-Assisted Coding Techniques

29:31 Future Vision: Voice and Proactive Health

30:59 Continuing the SmartNourish Project

31:48 Productization Challenges

33:22 Building AI Guardrails

34:12 Real-World Impact of AI in Healthcare

34:45 Closing Remarks and Project Details

#aitinkerers #healthtech #langchain #anthropic