I work at the intersection of ML and symbolic reasoning: extracting structured knowledge from LLMs and verifying it against a symbolic knowledge base rather than trusting raw model output.

My LiveKnowledge project closes this loop end-to-end: an LLM proposes candidate facts from unstructured text, an Answer Set Programming solver (Clingo) checks the merged program for logical consistency, only verified facts get written to the knowledge base, and a human reviews any gaps or contradictions the solver flags. This is neuro-symbolic AI in practice — using the LLM for language understanding and the symbolic layer for correctness guarantees.

I later ported the same loop into the browser as WebILP, adding inductive rule learning on top: it runs Popper's ILP encoding via Clingo WASM and tests candidate rules with Trealla Prolog, so you can go from plain text to a working symbolic ruleset with no install, entirely client-side.

If you're building AI systems where hallucination and unverifiable outputs are a dealbreaker (agents, knowledge bases, decision support), this hybrid approach is worth a look.