Tama

Giving visual intelligence a body

I built Tama as an experiment in embodiment. I connected visual understanding, conversation, expression, and physical form so the intelligence felt like one character rather than a collection of features.

Presence cannot be pinned to a corner

I made Tama physical so its face could meet your gaze, its voice could come from a place in the room, and it could stay present when the screen changed. That spatial continuity made the exchange feel more personal than talking to another panel inside the computer.

Every state needed an answer

I designed each expression in Figma and mapped it to a system state. I used an open-source robotics face framework as the animation foundation, then used AI-assisted coding to integrate the state logic into the working software.

Latency is part of the personality

I made Tama acknowledge the wait instead of going silent. I constrained responses to stay concise, personal, and expressive, using BMO from Adventure Time as the personality reference: playful and emotionally present, but still useful.

Measured wait

I timed the first successful Gemini call at 9.97 seconds of model time. I set thinking to minimal on the same frame to cut that to 4.05 seconds — about 59% faster thinking, with thinner scene detail as the tradeoff.

I did not make the cloud faster by making Tama physical. I used trigger-only vision, a downscaled screenshot, local STT, and a face plus fillers so an 8–10 second wait felt like a creature thinking instead of a spinner.

API footprint

I kept the cloud cheap so Tama could stay a prototype. 40 requests over 28 days, every one succeeded, 9 cents on the chart.

Nine passes to make the form work

Designing from the inside out

I modeled every physical component at its real dimensions before printing so the display, Pico, speaker, controls, and wiring space became constraints inside the CAD model. I checked the real display against its digital stand-in before committing to the shell.

I turned the rear hatch into a removable access panel and a sticky-note holder so a maintenance requirement became part of the desk experience.

From model to first fit

I printed the enclosure, learned to solder at the CCA Hybrid Lab, assembled the electronics, and installed everything in the shell. The first full print fit correctly, which validated the component measurements and internal layout before final assembly.