Founding Engineer, Applied AI

Team
We are an a16z-backed pre-seed team: CEO (ex H.I.G. Capital, J.P. Morgan, Harvard Business School) and CTO (ex Salesforce Field Sales, Optimism / OP Labs, U.C. Berkeley CS).

We're looking for a founding engineer to help us build this rocketship.

Thesis
Kosha is building the retail execution layer for physical commerce.

Software has spent two decades optimizing for knowledge workers: people at desks, with laptops, working in structured interfaces. But most of commerce doesn't happen at a desk. It happens in store aisles, on loading docks, in face-to-face conversations. The systems of record for that world were built for the office and bolted onto the field, and it shows. Reps hate them, data quality is terrible, and most of what actually happens on the ground is never captured at all.

The new age of voice and vision models change that. For someone standing in an aisle, the natural interface isn't a screen. It's talking and looking. For the first time, models are good enough to turn a rep's spoken debrief and a photo of a shelf into structured, trustworthy data.

Our wedge is CPG, starting with beverages (the #1 consumed item on the planet). Every day, hundreds of thousands of reps walk into stores, have conversations, and make judgment calls, and almost none of that intelligence survives the visit. We intend to capture it. Voice debriefs and shelf images today, ambient capture and wearables next, converted into structured CRM records, competitive intelligence, and trade spend optimization.

Visit by visit, this becomes a proprietary dataset of what's actually happening on the ground, the true "eyes" and "ears" of what's happening. A real-time knowledge graph that incumbents like Nielsen and Circana can't see, an entirely new dataset of market intelligence for brands, distributors, and investors alike.

What you'll do

• Work on building out agent harnesses for various customer outcomes
• Own the voice stack end to end, from capture on a phone in a noisy store to clean, structured intelligence, and every latency, accuracy, and architecture tradeoff in between
• Build with the modern voice and AI ecosystem (Deepgram, Vapi, ElevenLabs, and whatever ships next) and decide what to buy, wrap, or build ourselves
• Deal with data structuring for agentic tooling: entity resolution, deduplication, and reconciling messy data across sources and integrations into one coherent picture

You might be a fit if you

• Are a builder first: you've shipped real products independently, you default to action, and ambiguity energizes you rather than blocks you
• Have worked with large codebases or architected large services at the scale of an enterprise CRM or ERP (data models, agent infrastructure, and the intelligence layer on top), even if you haven't built one yet
• Have built with LLMs beyond the demo stage: agents, structured extraction, context management, and evals that catch regressions before customers do
• Have wrangled real-world data: entity resolution, multi-source reconciliation, or the unglamorous work of making messy inputs trustworthy
• Have experience with multimodal AI, voice or vision or both, and understand why AI that feels like magic is brutally hard underneath
• Have mobile experience (React Native or similar) or the range to pick it up fast
• Want to own systems, not tickets: you think in architectures and data models, not features

You won't be starting from a blank repo. There's a working product with real users and disciplined foundations underneath it, but everything from here is yours to shape.

The best AI products feel like magic and are brutally hard underneath. If that gap is the kind of problem you want to live in, let's talk. This will be a fun ride, and we will learn a lot.