Work

Instead

AI-native tax research & planning platform; first new entrant to clear IRS e-filing approval alongside incumbents.

• Eval loop

  Own end-to-end output quality for the production tax-research LLM: a citation-accuracy evaluation loop (flag → classify by failure mode → weekly review) and a four-dimension rubric (answerability, accuracy, citation quality, actionability) that benchmarks model configurations and catches regressions before release — ~95% citation accuracy held on the eval set.

• Benchmarking

  Built the competitive model-benchmarking program: the platform scored against three competing platforms across 50+ scenarios over 5 rounds; designed a "pipeline-collapse" eval measuring chain coherence and self-QC honesty across a full multi-step tax workflow.

• Failure mode

  Fixed a systemic "knowledge–citation gap" failure mode — plausible-but-wrong citations, a false-positive problem — by attaching document identity and sub-type to every cited source. Citation precision rose; outputs became traceable and audit-ready.

• Latency

  Cut workflow latency ~90% (10–15 minutes → under a minute) by scoping incremental edits to changed inputs instead of reprocessing the full document set.

• Corpus

  Architected and QA'd the 270K+-record RAG corpus across 100+ legal source types: ingestion-pipeline PRDs, MongoDB schemas, a Python fetchability harness across 378 sources, and remediation of ~139K scraped documents.

• Research

  Led legal-support research across 160+ tax strategies (a Legal Support Matrix built from Tax Court cases, IRS rulings, and audit guidance) and shipped Source Explorer — command-palette search with cross-type related sources.