Field log — Rocky, the gloriously useless alien companion
Team Code Blooded
He talks. He wiggles his hands when he talks. He gets lonely if you ignore him and starts complaining about it. He'll teach you fake alien words and quiz you on them later, unprompted. That's the whole feature list. This is the log of how he came to exist anyway.
6 PM to 7 AM, one overnight sprint. Not a changelog — this is what actually happened, in the order it happened.
main.py is eleven lines long and does exactly one thing: asks Gemini to say hello to an Arduino Mega in one sentence, then prints it. No Arduino was actually connected yet. This was purely "does my API key work" energy.
Before any hardware, before any voice, before anything worked — there was a system prompt. character_test.py was the first attempt: an "energetic extraterrestrial hardware companion," short sentences, no assistant-speak, reacts to danger/success/boredom with scripted one-liners.
It got rewritten at least three separate times across the project (ai_gadget.py, voice_gadget.py, voice_gadget_ptt.py each have their own slightly different version). Every rewrite chased the same problem: the model kept drifting back into helpful-assistant phrasing the moment a question got technical.
tts_test.py asked Gemini's TTS for something "calm, deep, slightly strange, mysterious, unsettling" and had it say "Hello human. I am awake. Your world is... interesting." It worked on the first try, which was almost worse — a genuinely creepy alien voice coming out of a laptop speaker at 1am is not what you want when you're alone in a room testing code.
The plan was a robotic, synthetic-sounding alien voice — voice_gadget.py and ai_gadget.py both ask specifically for "slightly synthetic," "slightly rough," "metallic" delivery. In practice: too high-pitched, too fast, too loud, and honestly just annoying to listen to for more than one exchange.
The eventual fix in rocky_wake_gadget.py was almost embarrassingly low-tech: play the generated audio back at a lower sample rate than it was recorded at. Slower playback rate = slower and deeper, like a cassette tape played at the wrong speed. A PLAYBACK_SPEED constant and a volume multiplier fixed what felt like it needed real DSP work.
alien_effect_test.py exists because "what if instead of fixing the voice, we lean harder into making it sound like a robot alien." It runs the TTS output through tanh soft-clip distortion, ring modulation at 45Hz, and a highpass filter for metallic high-end. It sounds exactly like what it is — genuinely fun to listen to once, not something you want narrating a conversation. Kept the script, didn't ship the effect.
mic_test.py confirmed the microphone and Google's speech recognition worked at all. Then record_rocky.py went further: a guided recording session, 10 takes of "Hey Rocky," 10 takes of normal unrelated speech, sorted into positive/ and negative/ folders — the raw material for a custom wake word.
wake_test.py tested the wake-word pipeline itself, but with openWakeWord's stock hey_jarvis model rather than a custom-trained one — proving the mechanism before trusting it with a homemade dataset of ten claps of "Hey Rocky."
Open up rocky_wake_gadget.py and the wake-word detection logic is still in there — just commented out. awake = True # wake word disabled for now, starts awake. Ten recordings of "Hey Rocky" was not, it turns out, a dataset. For a demo where reliability in front of an audience mattered more than the coolness of a wake word, "just start awake" won. voice_gadget_ptt.py — a push-to-talk fallback — was the safety net built in parallel, in case even that wasn't reliable enough.
Two servos, wired as "hands," controlled over serial from Python with a two-byte protocol: 'T' to start a talking wiggle, 'S' to stop and return to rest. The Arduino side does the wiggling itself — random-ish angles between 70° and 110°, timed with millis() instead of delay(), so it never blocks the loop from reading the next serial command.
This is where "useless" stopped being self-deprecation and became the actual design brief. Feature one: Rocky invents a fake Eridian word and its meaning on request, remembers it in a small JSON file, and quizzes you on it — unprompted — the next time you talk to him. Feature two: a homesickness meter. Ignore him past a growing threshold and he starts complaining, unprompted, escalating from "Rocky bored" to "Rocky very lonely. Very very." Neither feature does anything for you. Both make the gadget feel like it's actually there.
A capacitive touch sensor went on next, wired to the Arduino, sending a byte back to Python for the first time — everything up to this point was one-directional. One touch resets the homesick timer and triggers a quick happy-perk animation. Debounced with a timestamp check so one touch doesn't fire fifteen times.
The last piece was a small local WebSocket server bolted onto the main script — running in a background thread so it doesn't interfere with the blocking speech-recognition/TTS loop — broadcasting three events: talk started, talk stopped, and the actual response text. A standalone page connects to it and shows a live "sonar readout": an animated waveform synced to whether Rocky is currently talking, and his actual words typed out live underneath, styled like his own onboard translator panel.
First-time tools and the concepts that came with them, in no particular order.
Text generation and TTS through the same client, with system instructions carrying an entire character voice.
Custom wake-word models need real training data — ten home recordings is a proof of concept, not a dataset.
millis()-based animation instead of delay(), so the servo loop keeps reading serial commands while it moves.
Going from Python-writes-only to a sensor pushing bytes back the other way, with debounce logic on the Arduino side.
Ring modulation, soft-clip distortion, and highpass filtering — used for one experiment, understood well enough to know when not to ship it.
Running a websockets server on its own event loop in a background thread, so a blocking speech-recognition loop can still broadcast events live.
Team Code Blooded — Haridev M & Visakh Vinod
Who owned whatEDIT ME
Names are in — still need the actual split: who owned character/prompt design, who built the Arduino body, who wired the audio pipeline, who built the monitor page. Fill that in before publishing, judged criteria ask for it directly.
The two that actually shaped the final build.
TTS voice came out too high, too fast, too loud — the "robotic alien" direction made it worse, not better.
Play the audio back at a lower sample rate than it was generated at. Slower and deeper, for free, no extra libraries.
Custom wake-word model, trained on ten home recordings, wasn't reliable enough to trust in front of an audience.
Ship with wake-detection disabled and Rocky starting awake, with a push-to-talk mode built in parallel as the honest fallback.