← All work

Applied AI

MSHN

A retrieval-augmented market-anomaly explainer that grounds its answers in SEC filings indexed before a price event.

MSHN explores how language models explain market anomalies using information available before the event. It combines a retrieval pipeline with an evaluation harness that measures factual grounding separately from causal reasoning.

What it does

  • Indexes SEC filings in PostgreSQL and pgvector for retrieval-grounded explanations.
  • Evaluates explanations across four axes with an LLM-judged harness.
  • Caches explanations in PostgreSQL so repeat queries do not require another model call.

Engineering challenges

  • Separating factual accuracy from the quality of causal reasoning during evaluation.