Generative Computing: An Evening with IBM Research

Welcome to an evening with IBM Research where we discuss Generative Computing, Mellea, and the practice of writing programs that call models the way they call functions.

An event hosted by IBM Research — PyData Ireland is excited to be a community partner.

This is the next Open Source Science Dublin meetup, where cutting-edge scientific research meets open-source technology, AI, and cloud-native infrastructure.

Expect technical deep dives, cross-disciplinary conversations, and (of course) pizzas 🍕!

For this event, we're exploring Generative Computing: the practice of writing programs that call models the way they call functions, with the same expectations of structure, testability, and governance we apply to the rest of our software.

LLM calls today are too often ad-hoc prompts glued into application code — brittle, hard to test, and impossible to audit. Generative Computing reframes them as first-class program elements: typed, composable, verifiable, and governable.

You'll see how this plays out in practice through Mellea, an open-source Python library for building predictable generative programs; FactReasoner, a real-world port that gained reliability and speedups by moving onto Mellea; and the Mellea Skills Compiler, which turns plain-language skill specs into governed, certified pipelines.

Talks

Talk 1: TBD

Talk 2: Generative Computing with Mellea: Programs That Call Models Like Functions

• Speaker: Rahul Nair, IBM Research
• Description: An introduction to Mellea (mellea.ai) an open-source Python library for building predictable generative programs. We'll cover the core abstractions — instructions, requirements, sampling strategies, and verifiers — and show how treating model calls as structured, testable operations changes what you can build. Live coding, real examples, and a look at how Mellea composes with the rest of the Python ecosystem.

Talk 3: FactReasoner: Porting a Probabilistic Factuality Assessor to Mellea

• Speaker: Radu Marinescu, IBM Research
• Description: FactReasoner (paper , code) is IBM's probabilistic long-form factuality assessor — it decomposes generated text into atomic claims and reasons over their support against retrieved evidence. This talk walks through the port from a bespoke implementation to Mellea, the reliability improvements and speedups that came out of it, and what porting an existing system teaches you about where Generative Computing pays off in practice.

Talk 4: The Mellea Skills Compiler: From Plain-Language Specs to Governed Pipelines

• Speaker: [TBD], IBM Research
• Description: The Mellea Skills Compiler takes a plain-language skill specification and compiles it into a governed, certified generative pipeline — with evaluation harnesses, requirement checks, and provenance baked in. We'll show the compiler end-to-end on a worked example and discuss what "certification" means for LLM-backed skills, and how this closes the loop between rapid iteration and auditable production deployment.