The idea here is to look at the current options LLM give us to arrive at determinism when we are trying to parse LLM outputs. From JSON Mode to structured outputs to Pydantic Models there are a plethora of these techniques to get the controlled output, but what if there was a much more cleaner and effective way. BAMLby folks at BoundaryML is a solution in that direction which is a much needed pill for everyone and anyone wanting to make LLM applications where structuring outputs is no more a chance game but a definite game.
The proposal is to give a talk on how to adopt BAML and what capabilities it has and how it can be used to augment software building in a more confident and deterministic way. Its ability to sit in the IDE and help debug the calls and directly see the prompts and the raw and controlled outputs puts away confusion and brings more focus.
That would be:
1. Structured Determinism with LLM outputs
2. Streaming Responses
3. Adding function calling to non supported LLM's
4. Opens up a lot of opportunities to build concrete , scalable and predictable systems