Wyvern – Real-time machine learning platform for marketplaces
Hello HN. I’m Shu, here with my best friend Suchintan. We’re the creators of Wyvern (https://github.com/Wyvern-AI/wyvern), a real-time machine learning framework to help marketplaces adopt ML earlier in their life. Wyvern is an end-to-end platform to help marketplaces build and scale custom machine learning pipelines, with the goal of enabling data scientists to own the full machine learning stack. It offers a bundle of: 1. Feature store to store and serve features 2. Model service to serve ML models 3. Feature + Model observability platform 4. Orchestration framework for…
In plain words
Wyvern is an open-source machine learning platform designed for marketplaces to build and deploy custom ML pipelines at scale. It provides a feature store, model serving, observability tools, and orchestration capabilities, with pipelines optimized to run in under 200 milliseconds. The platform enables data scientists to manage the complete ML stack independently, with built-in safeguards against train-test skew. Wyvern targets marketplace companies seeking to implement machine learning earlier in their growth stages.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
Hello HN. I’m Shu, here with my best friend Suchintan. We’re the creators of Wyvern (https://github.com/Wyvern-AI/wyvern), a real-time machine learning framework to help marketplaces adopt ML earlier in their life. Wyvern is an end-to-end platform to help marketplaces build and scale custom machine learning pipelines, with the goal of enabling data scientists to own the full machine learning stack. It offers a bundle of: 1. Feature store to store and serve features 2. Model service to serve ML models 3. Feature + Model observability platform 4. Orchestration framework for ML Pipelines While being mindful of the following constraints: 1. Pipeline evaluation in < 200ms to optimize user experience [1] 2. Minimizing train / test skew for model training Suchintan built the ML Platform at Faire and Gopuff to improve their Search and Discovery experience. At both places, the platform became an engine that empowered the data team to independently deliver new models to production, generating over $100M of impact. Small marketplaces usually buy solutions like Algolia and AWS Personalize for Search and Discovery. Once a marketplace grows to a certain size (usually >$100M GMV), they hit limitations around how many custom (”long-tail”) insights they can incorporate into these solutions. Long-tail insights may come in the form of “features” that help move model accuracy, or changes in model objectives that help optimize for underlying business objectives: - B2B Marketplaces like Faire are able to segment users based on what they’re selling through on their storefront by asking them to categorize themselves on sign-up, and feed that into machine learning models to cold-start personalization for new users. - Margin-sensitive companies like Gopuff are able to build a function to predict the expected revenue of a product being shown to the user, and ordered the results accordingly. This was composed by several ML models coming together trying to predict the probability the user would purchase a product if we showed it to them, and the expected margin of that transaction. Cart-state as a signal is also very helpful when ranking complementary products for users, ie specific types of chips would be favored if you had some coke in your cart, versus sprite We’ve talked to a bunch of other marketplaces and we’ve learned they have their own long-tail insights that may improve the user experience within their own models: - Recipe marketplaces like Cookpad may associate each recipe with a flavour profile, and leverage that data to map recipe flavors to predicted user profiles (ie weight Savoury recipes higher for users that enjoy Savoury food) - Device reselling marketplaces like Valyuu may use the brand of the users’ device they’re viewing the website on to predict which type of product they were most likely to purchase (iPhone users buying another iPhone) The question we asked ourselves is: How much engineering involvement is actually required here, and how much can we generalize? We built Wyvern to abstract ML engineering work mentioned above away from data scientists. As a result, they can just focus on defining the request/response of the API, the model, the features for the model, the business logic, and finally training the models with the feedback data generated by the ML pipeline. We would love your opinions and hot-takes. Please follow our Quickstart (https://github.com/Wyvern-AI/wyvern#quickstart) to give it a try. [1] https://iarapakis.github.io/papers/TOIS17.pdf
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
