Inspiration
The Trojan horse got through the gates by looking harmless. We hid a solar tracker inside a cardboard horse for the same reason: people walk right up to it. Small field panels lose energy when they point the wrong way or jam, and nobody notices until a battery dies. Odyssey notices, fixes what it can, and tells you the rest in plain English.
What it does
• Chases the light: a servo sweeps the solar tail from 15° to 90°, parks at the peak, then nudges ±4° every 1.5 s to keep it. • Reads people: an ultrasonic sensor tracks distance and approach speed. Slow visitors get a greeting and can touch it. Anything faster than 60 cm/s inside 50 cm makes it tuck its tail away. • Catches its own faults: if the tail's real angle is 15° off for 800 ms, it retries once, then cuts the servo and asks for help. • Sleeps when power stays under 3 mW, and checks for light every 15 s. • Tells you why: ask a question and it answers aloud from its own logs, with real times and numbers.
How we built it
An ESP32 runs every reflex from a pure C++ logic class that we tested on a laptop first. It streams JSON telemetry at 10 Hz to a FastAPI server, which logs to Tiger Data and posts highlights to Notion. xAI's Grok answers from a snapshot of those records, and any action it proposes needs a person to approve it. ElevenLabs handles the voice both ways. A Node.js Companion Studio lets us record states and design new movements.
Challenges we ran into
At 3 am our first truthfulness run with real keys scored 0/20: the Gemini model we started on had been retired (we've since moved to Grok). The next run hit the free tier's 5 requests a minute. Sending a records snapshot with each question cut it to one request, and the score climbed to 15/20. Two of the five failures were timeouts; the other three were bugs in our own checker, like reading "4:06 AM" as an invented number. The HC-SR04 was its own fight: a person stepping in from the side looked like a 1,000 cm/s lunge until we only computed speed inside 150 cm.
Accomplishments that we're proud of
Every cute behavior does a real tracker's job. It won't invent facts. And it keeps working with no internet, no AI quota and no voice credits.
What we learned
Keep the LLM out of the reflex path, ground it in your own data, and build the fallback before you need it.
What's next
The same firmware on real field panels, units that compare output with their neighbours, longer tracking-versus-fixed tests, and an open-source STEM kit.
Built With
- arduino
- c++
- elevenlabs
- esp32
- fastapi
- grok
- node.js
- notion
- postgresql
- python
- tiger-data
- web-serial
