Sukku: Building a Robot That Knows How You Feel
Most robots understand commands. Sukku is my attempt at a robot that understands people — physically, emotionally, and mentally.
One emotion, three signals
Humans don't broadcast emotion through a single channel, so a single-channel model will always be wrong in exactly the situations that matter. Sukku runs a custom multimodal emotion-recognition model that fuses three streams into one unified emotional state estimate:
- Vision — facial micro-expressions, not just "smile detected"
- Audio — vocal prosody: pitch, pace, tension in the voice
- Context — conversation memory and environment
The fusion is the point. A flat voice with a smile means something different than a flat voice with a flat face. Any system that scores modalities independently and averages them misses the person entirely.
Memory is what makes it a relationship
The feature that changes how people react to Sukku isn't the emotion model — it's the persistent per-user memory. Sukku recognises returning humans and remembers them across sessions.
A robot that reads your mood is a demo. A robot that remembers you were stressed last week and checks in — that's the thing people don't expect from a machine.
Why edge, why a Pi
Everything runs on edge hardware — a Raspberry Pi 5 with an accelerator. Two reasons:
- Latency — emotional responsiveness dies if the reaction takes a cloud round-trip. Human conversation operates in hundreds of milliseconds.
- Privacy — a camera and microphone that continuously read your emotional state should not stream that data anywhere. On Sukku, the raw signal never leaves the device.
Constraint breeds efficiency: fitting vision + audio + fusion models into edge compute forces you to be honest about what each model actually needs to see.
Sukku sits in the same thesis as everything I build — Veris proves what a camera saw, Kimi remembers what I did, and Sukku understands who it's looking at. Machines should meet reality where it is.
// RELATED_TRANSMISSIONS
- PHANTOM: I Built a Benchmark Where an RL Attacker Gaslights an LLM DefenderMost cybersecurity benchmarks test what an agent knows. PHANTOM tests whether it can still think when a co-evolving adversary is lying to it in real time.
- Veris: Building a Camera That Cannot LieA Raspberry Pi camera with a hardware-derived ed25519 identity, on-chain signature verification on Solana, and Filecoin storage. Deepfakes die at capture time.
- Training a Masked Diffusion Language Model on a MacBookGPT generates left to right. My model starts from a fully masked sequence and denoises it into a story — trained from scratch on TinyStories on an M4 Pro.