MLOps Community

Coding Agents Conference. March 3rd at the Computer History Museum. luma.com/codingagents

open.spotify.com/episode/32k745… Would love to hear how others are thinking about this shift. A lot of the patterns Hamza described sounded very close to ML orchestration, even though the market is packaging it differently.

3⃣ The semantics debate was surprisingly useful. They spent time unpacking what durability can and cannot guarantee, especially around the external state. A lot of people seem to assume these systems can magically recover everything after failure, which is not really the case.

@htahir111 from ZenML was on the latest MLOps Community episode with Demetrios talking through durable execution, agent harnesses, and why a lot of “long-running agents” are basically while loops with state recovery glued around them.

No setup required — every attendee gets a ready-to-use VM. Seats are limited. 📍 NYC — 307 West 38th Street, Studio 1505 📅 June 4 | 5 PM – 9 PM EDT Register here: luma.com/nyc-i97h

Topics include: ✔️ Fine-tuning retrieval models ✔️ Sparse + dense + hybrid search ✔️ Relevance evaluation metrics ✔️ Similarity vs relevance ✔️ Recommendation systems + RAG workflows ✔️ Production retrieval stacks

Generic embedding models get you “kind of similar.” But product search in production? That’s where things break. Wrong variants. Missed attributes. Semantically correct but commercially useless results.

The “Tinder for food” experiment was also wild. Apparently, it’s working better than expected. go.mlops.community/pg49rh

3⃣ They also talked about recommendation systems needing controlled exploration. Recommending the same sushi forever is easy. Getting someone to try something new without making the recommendations feel random is where most of the work is happening.

Listened to the latest MLOps Community episode with the iFood team talking through what happens when recommender systems turn conversational.

home.mlops.community/public/videos/… Would love to hear how other teams are handling orchestration right now, especially after seeing how quickly tool sprawl shows up in production.

3⃣ The strongest use cases weren’t flashy. Most of the value came from boring operational friction: check-ins, vouchers, delayed flight claims, trip coordination in WhatsApp groups, and helping humans avoid opening 14 browser tabs to compare flights.

Been listening to the MLOps Community episode with Nicolás Alejandro Bogliolo from Despegar about building Sofia, their travel agent system. Travel sounds like an obvious place for agents until you look at the workflows properly.