Sidney Primas

Sidney Primas is co-founder and chief technology officer of LemonSlice, which builds interactive video characters that converse, move, and respond in real time. His defining challenge is the avatar Turing test: creating digital characters whose appearance, emotional timing, and behavior approach the experience of speaking with another person.

Primas studied engineering at Duke University and earned a master’s degree from MIT, where he developed computer-vision algorithms for automated microscopic urinalysis. At Jawbone, he led technical work on wearable bioimpedance sensors and helped bring heart-rate algorithms from clinical validation into mass production; his work contributed to five signal-processing patents and products manufactured at a scale exceeding 500,000 units.

From early 2020, Primas helped build Edge Analytics, applying machine learning and algorithm development to healthcare, wearables, consumer electronics, and connected devices. He subsequently founded Infinity AI with Lina Colucci and Andrew Weitz. The company joined Y Combinator’s Winter 2024 batch, evolved into LemonSlice, and introduced live conversational avatars generated from a single image in 2025. Its collaboration with Microsoft produced an interactive Theodore Roosevelt experience, allowing visitors to converse with an expressive historical character.

• Expressiveness depends on more than lip synchronization. Primas emphasizes microexpressions, gestures, emotional timing, physical movement, and interactions with an avatar’s surroundings. Audio representations must capture emotional nuance, not simply encode spoken words.

• Real-time video needs causal generation and long-horizon stability. LemonSlice’s video diffusion models generate characters using only available conversational inputs, accelerate synthesis by reducing denoising, and address accumulated visual errors that can degrade extended interactions.

• The model harness determines whether avatars work outside the lab. Continuous video requires coordinating speech recognition, language models, synthesized audio, CPU and GPU workloads, buffering, interruptions, and streaming. LemonSlice’s open-source integration examples show how its visual infrastructure connects with external voice systems and frameworks including LiveKit and Pipecat.

• Emotional intelligence and reasoning are distinct technical layers. Primas envisions an emotionally intelligent audiovisual model that interprets speech and video, maintains an internal emotional state, and generates expressive responses, while a separate reasoning system supplies knowledge, tool use, and complex decisions.