The SF Systems Club met on September 24, 2026 at LatchBio, 185 Berry Street in San Francisco, for an evening titled Research to Practice, from Inference Engines to Multimodal Coding. Three hundred and sixty people registered.
Shadaj Laddad opened it. The club's goal, he said, is a community of people excited about systems broadly — compilers, distributed systems, programming languages, embedded systems — fields of computer science and engineering that share a great many principles. The focus is kept on core engineering: no fluff, no marketing, here to learn things. This meetup in particular set out to connect the Bay Area's engineering and research communities, which he said do not talk to each other often enough, with a night of talks showing research making its way into the real world. He named the organisers — Chris, Conor, Jeff, Neil and himself — invited anyone with a large space to host a future meetup, and thanked LatchBio for the space and the food.
LatchBio used their two minutes to make a case for biology as a domain worth working in, and for a specific technical claim: that bespoke scientific workflows can be decomposed and made verifiable, chunked up with rewards constructed from each chunk's domain and context. On that approach they have published benchmarks across drug development, molecular measurement and biosecurity. The recruiting pitch was a small team of engineers embedded with a large team of scientists, and four years of biology infrastructure — FUSE drivers, container orchestration, distributed file systems — now repurposed to evaluate agents.
Shreya Shankar gave the main talk, on serving LLMs for database-style operations. She built DocETL, an open-source system for processing unstructured data, during her EECS PhD at UC Berkeley, and joins CMU as an assistant professor in 2027. The talk starts from what AI has become good at: the basic text operations NLP has worked on for decades — extraction, classification, summarization — which models have handled passably for a while but only recently handle zero-shot, so nobody has to collect training data first. From there into the inference question that follows once those operations are run at database scale. Parker Zeigler also spoke, on interactive multimodal programming systems; that talk was not recorded.
Alexy Khrabrov interviewed two of the club's organisers, Shadaj Laddad and Chris Wensel. The history runs deep: Shadaj gave his first talk as a child at the Scala meetup Alexy had just started in the Bay Area in 2011, and both his parents were long-time speakers there too, which the two of them agree makes it a family business. Shadaj works at the interface of functional programming and distributed systems, and now AI, and leads the Hydro team at AWS. Chris Wensel is the creator of Cascading.
Alexy's notes from the floor are a struct.fm episode. The recordings are below.