Sarab Ur Rehman

AI Engineer Intern

IT SOLERA • Temps plein

Jun 2026 - Present • 3 mos

Designed and proposed an AI-powered financial reconciliation system that automates root-cause analysis of ledger discrepancies across multi-system environments (ERP, bank feeds, CRM/billing, payment processors). Unlike conventional reconciliation tools that only flag mismatches, the system pinpoints the exact transaction responsible for each gap and explains it in plain language — e.g., tracing a $4,200 discrepancy to a duplicate refund that never synced back to the ERP. Architected a two-layer pipeline separating deterministic logic from generative AI: a transaction-matching and anomaly-detection engine normalizes multi-source logs into a common schema and identifies unmatched, duplicated, mistimed, or partially-matched transactions, producing a verifiable evidence trail; a constrained LLM narrative layer then converts each evidence trail into an auditor-ready root-cause explanation, with every claim traceable to a specific source transaction — addressing a known failure mode where generic LLMs generate plausible but ungrounded explanations for financial discrepancies. Key components: Multi-system transaction ingestion connectors normalizing ERP, banking, CRM, and payment data into a unified schema Deterministic discrepancy detection engine for duplicates, timing gaps, rounding differences, and unsynced transactions Grounded narrative generation layer producing plain-language, evidence-linked reports suitable for month-end close and audit workpapers Synthetic benchmark dataset with injected discrepancies and ground-truth root causes for evaluating explanation accuracy Outcome: a validated, deployable close-cycle tool aimed at reducing the manual hours finance teams spend tracing reconciliation gaps, backed by quantitative benchmarking of narrative grounding accuracy against manual and ungrounded-LLM baselines.