Software Engineer, Multimodal Storage Infrastructure

# **About Eventual** Every breakthrough Physical AI system — humanoid robots, autonomous vehicles, video generation models — is trained on petabytes of video, lidar, radar, and sensor data. But today's data platforms (Databricks, Snowflake) were built for spreadsheet-like analytics. They don't know how to index a clip by content, co-locate sensors on the same row as video, version multimodal datasets, or push predicates down to a corpus of MP4s. Robotics and video-AI teams build the missing layer themselves: stitching together five to eight tools, organizing disorganized video and sensor data, building schemas and versioning that don't exist. *"It was rebuilding what Databricks built 15 years ago for analytics — just for AI data."* Eventual was founded in 2022 to ship that layer once. Our open-source engine, [Daft](https://daft.ai/), is the distributed data engine purpose-built for multimodal AI — already running 2 PB/day at Amazon, 60-100 PB at another FAANG company, and in production at Mobileye, TogetherAI, and CloudKitchens. We are building a multimodal warehouse on top of our engine for Physical AI: video, sensors, and sim outputs co-indexed on the same row, aligned on timecode, and versioned — with a content-aware query layer on top. We're building this in partnership with the top PhysicalAI labs and public AI infrastructure companies today. We have raised $30M from Felicis, CRV, Microsoft M12, Citi, Essence, Y Combinator, Caffeinated Capital, http://Array.vc, and angels from the co-founders of Databricks and Perplexity. We've assembled a world-class team from AWS, Render, Pinecone and Tesla. We have spent our careers powering the last generation of PhysicalAI in self-driving, and are excited to now do this for the next. Join our small (but powerful!) team working together 4 days/week in our SF Mission district office. # **Your Role** As a Storage Infrastructure Engineer, you'll take everything we know about modern databases and apply it to the world of Physical AI. Our warehouse co-indexes video, sensors, embeddings, and sim outputs on the same row, versioned, with a third query layer (not row/column, not vector/semantic) — content-aware queries over what's *inside* clips. Your job is to make that layer fast: the right indices for petabyte-scale video, predicate pushdowns that elide whole files, file formats that respect random access into clips, and a query path that turns *"left-arm grasp failures on deformable objects"* into the smallest possible read. You should believe, in your bones, that **the best read is the read elided.** ## **Key Responsibilities** - Design and build the storage and indexing layer: row groups, column chunks, secondary indices, vector indices, and the metadata that lets queries skip everything that doesn't matter. - Push the query engine harder — predicate pushdown, projection pushdown, late materialization — across multimodal columns including video, embeddings, and sensor streams. - Choose, extend, or build on top of modern open formats (Parquet, Iceberg, Delta etc) and build our own/contribute upstream where it makes sense. - Build versioning and schema evolution for multimodal datasets so customer data stays reproducible across months of experimentation. - Partner with the Dataloading team on the format-to-loader boundary so an `iceberg.scan(...)` translates into the absolute minimum of bytes hitting NVMe. - Partner with the Visual Understanding team to land model outputs in the index without an external glue layer. ## **What we look for** - You love thinking about indices. B+ trees, LSM trees, bitmap indices, vector indices, learned indices — you have favorites and you have grudges. - You love thinking about query engines. *Predicate pushdown* makes you happy. *Late materialization* makes you happier. - Strong familiarity with the storage hierarchy: cloud object stores, NVMe, block storage, spinning disk, RAM, GPU memory — and the latency and cost of moving between them. - Strong opinion…