Autonomous Agentic Maker & Tutorial Studio

An autonomous multi-agent tutorial that analyzes raw hardware build media to generate kid-friendly scripts, real anime reaction stickers, and 44.1kHz audio cues.

Comment

Inspiration

Millions of students, educators, and hobbyists build incredible robotics, Arduino gadgets, and electronics projects every day. However, sharing these creations with the world is challenging. Turning a raw, silent circuit recording into an engaging, broadcast-ready YouTube or TikTok tutorial is overwhelming:

  • Writing engaging scripts that explain technical electronics concepts using intuitive, kid-friendly analogies takes hours.
  • Hunting for visual assets and expressive character reaction overlays matching specific themes is tedious.
  • Designing matching sound effects synchronized with physical hardware movements requires professional audio editing tools.

We were inspired to build the Autonomous Agentic Maker & Tutorial Studio — an intelligent multi-agent AI co-pilot that eliminates this friction. By feeding raw hardware build recordings into our agentic pipeline, creators instantly receive a complete, synchronized production package tailored to any creative prompt.

What it does

Our system orchestrates a team of specialized, autonomous AI agents to automate the end-to-end tutorial asset creation workflow:

  1. Multimodal Circuit Analysis: Ingests raw video clips and photos. Using Google Gemini 3.5 Flash, it autonomously detects electronic components (e.g., microcontrollers, ultrasonic distance sensors, servo motors), formulates intuitive analogies, and pinpoints 3–5 dramatic key moments across the video timeline.
  2. Kid-Friendly Scriptwriting: Composes a structured teleprompter dialogue with synchronized timecodes and embedded inline cue tags: [SFX: ...] and [REACTION: ...].
  3. Dynamic Reaction Asset Engine: Rather than using static placeholders, the agent dynamically interprets the user's creative prompt (e.g., Attack on Titan, Jujutsu Kaisen, Spy x Family, Retro Arcade, or Cyberpunk) to formulate targeted web queries, retrieve real character reaction frames, and process them into die-cut vinyl stickers with emotion tags.
  4. 44.1 kHz Sound Design: Searches CC-licensed audio or procedurally synthesizes 16-bit 44.1 kHz WAV sound effects (radar sonar pings, power-up boots, obstacle alarms, and victory fanfares).
  5. Human-in-the-Loop Director Review: Gives creators full oversight to inspect assets, listen to audio cues, submit natural-language revision notes, and approve the project.
  6. Production Package Exporter: Generates a timestamped export directory containing structured script.json, teleprompter_script.md, high-res transparent PNG overlays, broadcast WAV files, and a standalone interactive HTML web player.

Mathematical Formulation

For hardware projects involving ultrasonic distance sensing (such as the Arduino sonar radar), the multimodal agent decomposes physical wave propagation into kid-friendly analogies using the time-of-flight equation:

$$d = \frac{v_{\text{sound}} \cdot \Delta t}{2}$$

where $v_{\text{sound}} \approx 343 \text{ m/s}$ in ambient air and $\Delta t$ is the round-trip echo pulse latency detected by the sensor. To produce stylized sticker overlays, our processing engine applies mathematical morphological dilation on the alpha mask:

$$M_{\text{outer}}(x,y) = \max_{(dx, dy) \in B_r} M_{\text{alpha}}(x+dx, y+dy)$$

where $B_r$ represents a structuring ball of radius $r = 16\text{px}$, creating a clean outer border followed by a Gaussian drop-shadow kernel.

How we built it

  • Core AI Intelligence: Google Gemini 3.5 Flash (gemini-3.5-flash) via the official Google GenAI SDK (google-genai) for multimodal vision understanding, JSON schema compliance, and dynamic prompt formulation.
  • Agent Orchestration: Google Agent Development Kit (ADK) state machine pattern managing sequential and parallel agent execution (ANALYZING $\rightarrow$ SCRIPTING $\rightarrow$ GENERATING_ASSETS $\rightarrow$ AWAITING_APPROVAL $\rightarrow$ APPROVED).
  • Backend Framework: Python 3.11 with FastAPI development server and asynchronous background task execution.
  • Automated Verification: pytest test suite with 100% passing unit and integration tests (7/7 passed).
  • Containerization: Fully Dockerized with Dockerfile and service.yaml ready for Google Cloud Run.

Challenges we ran into

  • Dynamic Agentic Search vs. Hardcoding: Ensuring the agent had complete creative autonomy to handle any franchise or style requested in the prompt without relying on hardcoded lists required careful system prompt design and structured JSON output constraints.
  • Asset Processing Reliability: Building a robust die-cut image processing engine that handles varying aspect ratios, transparency channels, and text scaling without distortion.
  • Infrastructure & Local Execution Note:
    The entire multi-agent pipeline was developed and demonstrated 100% locally via FastAPI and connected directly to Google's official Gemini 3.5 Flash foundation model. We did not host a live public instance on Google Cloud Run because the Devpost free Google Cloud credit coupon distribution had concluded by the time of our registration, and regional international card restrictions prevented setting up a GCP billing account. The project includes a complete Dockerfile and is 100% Cloud Run ready for one-click deployment.

Accomplishments that we're proud of

  • True Zero-Hardcoding Agentic Intelligence: Adapts dynamically to any theme, style, or anime franchise requested by the user.
  • Sub-15 Second Execution: Generates an entire production package (scripts, image assets, audio cues, and viewer) in seconds.
  • Robust Quality: 100% passing automated test suite (7/7 passed in pytest).
  • Complete Deliverable Bundle: Delivers ready-to-use assets for video editors, teleprompter readers, and interactive web viewers.

What we learned

  • How to harness Gemini 3.5 Flash multimodal vision to extract technical component lists directly from video footage.
  • The power of combining autonomous AI planning with Human-in-the-Loop review to ensure creators maintain creative control over generated media.
  • Implementing asynchronous multi-agent coordination with the Google Agent Development Kit.

What's next for Autonomous Agentic Maker & Tutorial Studio

  • Automated Video Timeline Assembly: Direct integration with video editing tools (e.g. OpenTimelineIO / FFMPEG / Premiere XML) to automatically drop the generated stickers, audio waveforms, and teleprompter captions directly onto the creator's video timeline.
  • Multilingual Localization: Leveraging Gemini's multilingual capabilities to automatically translate scripts and synthesize audio voiceovers in 30+ languages for global STEM education.
  • Interactive AR Overlays: Converting extracted circuit pins and component bounding boxes into animated augmented reality wiring guides for interactive student learning.

Built With

  • agent
  • agentic-ai
  • ai-agents
  • cloud-run
  • computer-vision
  • development
  • docker
  • fastapi
  • firestore
  • gemini-3.5-flash
  • genai
  • google
  • google-adk
  • google-gemini
  • google-genai
  • html5
  • javascript
  • multi-agent-systems
  • multimodal-ai
  • pillow
  • pydantic
  • pytest
  • python
  • rest-api
  • uvicorn
Share this project:

Updates

Submission history