Abhishek

Tuteja

Three concurrent LLM agents assess model documentation across 20 SR 11-7 elements, feeding a judge and orchestrator layer that runs cross-agent consistency checks, confidence-scored retries, and scoped re-assessment of disputed findings. The pipeline is hardened against adversarial input with prompt-injection defenses and a contrarian judge stance that counters self-enhancement and position bias. It ships on an evaluation harness of Jest and Playwright tests plus a custom score-drift eval that fails CI when a prompt change moves scores more than ten points. Output is a weighted compliance score with per-element gap analysis, evidence citations, and a downloadable PDF report.

Ingests raw broker trade data, normalizes it into a consistent schema, and constructs positions with accurate realized and unrealized P&L. Built for the cases that break naive trackers: partial fills, multiple trades per asset, short positions, and multi-currency positions priced through real-time FX rates into a single base currency. A P&L accuracy testing framework guards calculation correctness, because the priority is deterministic, auditable accounting over flashy visualization.