Miju Labs
Background
Studied economics at the University of Chicago and worked in brand strategy before joining Exa as an early employee (AI Engineer bio). At Exa, public posts describe her as head of strategy and marketing and, at one point, chief of staff; Amplify credits her with leading growth and with Exa's distinctive billboard marketing. Latitud says she was in its 2024 fellowship after a previous company that 'didn't work out', and that Latitud wrote Taste Labs' first cheque.
She left Exa to found Taste Labs, which ran in stealth, reportedly already selling to labs and application companies, before launching publicly on 16 Jun 2026.
What they run now
Runs the company and the lab sales. She is positioning Taste Labs as a 'judgement company' that starts in design and expands to code style, editorial and brand. Two product lines: post-training data, rubrics and RL environments for labs, and the Brand API for application companies. She is hiring engineers and growth staff and scaling the Taste Makers designer network. At launch she said inbound demand had been high enough to delay the public launch (TBPN).
Career
- presentFounder & CEO, Taste Labs
- 2024Founder, Earlier startup (name not public) — Latitud: 'didn't work out'
- —Brand strategy (employer not public)
- —Early employee; growth lead / head of strategy & marketing (also referred to as chief of staff), Exa
- —University of Chicago, Economics
On the record
Next-token prediction regresses to the probability-weighted average, and RLHF on aggregated preferences strips out distinctive outliers, so post-training flattens creative output. source
Taste is pattern recognition built through exposure and selection, not an innate gift, so it can be captured as training signal. source
Subjective domains lack the verifiable rewards code and math have; the company's mission is 'making the unverifiable verifiable, starting with design'. source
Taste needs continuous retraining as trends shift, unlike one-off fixes such as six-fingered hands; a recurring rather than one-time data need. source
Solving taste needs both better post-training data for foundation models and inference-time tools for brand and context consistency. source
