About Luminade
We're building the conversational voice layer for how people work. People have found dictation useful for input and text-to-speech useful for output. We're going further: a single conversation you never have to step out of, where you get responses back and actually get things done.
We started with email, the place where most work actually lives. Now we're expanding to calendar, documents, and more, so our AI has complete context on what matters in a person's day. The goal is simple: let people work without being hostage to a screen.
We're building first for people with vision impairment, ADHD, and dyslexia, communities where this kind of interface isn't a nice-to-have. But this is not a niche product. The same way curb cuts were designed for wheelchair users and ended up benefiting everyone, we're using accessibility as the design constraint that forces us to build something genuinely better. The endgame is hundreds of millions of sighted users who want to stay productive during their commute, away from a screen, or simply in flow.
The Team
Our CEO, Sriram, is an engineer and entrepreneur (1 exit) who started this company because of his own experience with vision loss. He knows this problem from the inside. Our CTO, Mikhail, is a former Google engineer and four-time world champion in competitive skydiving. He brings the same precision and intensity to systems architecture that he brings to everything else.
We're backed by South Park Commons, have raised $2M, and work in-person in San Francisco.
What You'll Own
The quality of our AI agent IS the product. You'll own the intelligence layer — this isn't a research role. This is the person who makes the agent reliable, fast, and smart enough that users trust it with their actual work.
The voice agent pipeline. Designing and optimizing the full model orchestration chain for real-time, low-latency conversational interactions. You'll decide which models to use, how to stitch them together, and how to make the whole thing feel instant.
Prompt engineering and evals. Crafting, testing, and iterating on prompts across the product. Building eval frameworks that catch regressions before users do.
Agentic workflow orchestration. Building the multi-step reasoning and action-taking capabilities that let users manage email, calendar, and documents through natural conversation. Context management, memory, and knowing when the agent should act versus ask.
LLM reliability and regression testing. LLMs are nondeterministic. You'll build the systems that ensure consistent, high-quality responses across thousands of user interactions.
Model selection and optimization. Evaluating and integrating the right LLMs, speech-to-text, text-to-speech, or speech-to-speech models. Making hard tradeoffs between quality, latency, and cost at every layer of the stack.
What We're Looking For
You've worked deeply with AI/ML in production and felt the pain of making these systems reliable at scale. You've shipped LLM-powered applications where model output goes directly to humans, not dashboards. Ideally you've built real-time or voice AI systems — but more than any specific experience, you take initiative, you ship, and you think about the person on the other side of every interaction.
