Hamid Vakilzadeh

Selected Work — AI, Evaluation & Open Source

• Architected a multi-agent LLM system — lead orchestrator, parallel specialist sub-agents, research planner, and an LLM-as-judge verifier — with tiered model routing for ~40% cost savings and an evaluator-optimizer loop that revises until every figure is verified.
• Built an anti-hallucination pipeline pairing an LLM verifier with a deterministic, pure-code citation-grounding audit that re-matches every cited number against source SEC/XBRL facts at any unit scale — making each figure traceable to its filing.
• Engineered a custom LLM evaluation harness with a numeric hallucination guard, enforced by an automated test suite with CI gating on every change.
• Established LLM observability with Langfuse + OpenTelemetry span trees (per-run, per-tool, per-generation token / cost tracking) on a serverless Cloudflare Workers + React 19 stack.

• Designed a correctness-monitoring framework for an agentic AI workflow, separating model-owned input quality from downstream system behavior.
• Defined SLIs, a multi-label error taxonomy, human labeling workflow, and an LLM-as-judge calibration process for high-risk outputs.
• Built dashboard-ready metric definitions for evaluation coverage, review triage, and judge-vs-human alignment.

• Evolved from RAG-based literature Q&A into an MCP-enabled agentic system for source-grounded search, citation-network analysis, and full-text extraction across Semantic Scholar, arXiv, and Wiley.
• Scaled to 100,000+ queries/month across hosted and local workflows — Smithery.ai, Claude, Cursor, Claude Code, and npm installs.
• Published system architecture and evaluation results in the Journal of Information Systems (ABDC A).

• Built an AI teaching assistant integrated with Canvas LMS, automating student support and curriculum-aligned Q&A for university courses.
• Developed multi-turn conversational AI with context-aware retrieval across course materials, syllabi, and assignment databases.

• Created and maintain the official Python API wrapper for XBRL.US with 30,000+ downloads, enabling programmatic access to structured financial data.
• Designed clean API abstractions for complex SEC filing queries, cutting analyst data-extraction time from hours to minutes.

• Co-built an MCP server for querying as-filed SEC, FERC, and ESEF reports with structured XBRL context — labels, periods, units, dimensions, and provenance; ranks on page 1 of Smithery with 26,000+ connections.

Academic Appointments

• Lead AI, ML, and NLP research applied to financial reporting, audit quality, and corporate governance; published 9 peer-reviewed articles (600+ citations).
• Serve on the Chancellor's AI Advisory Committee as Business School Dean's Liaison; teach Accounting Information Systems (mean eval 4.48/5.00); mentor 10+ doctoral students.
• Delivered 20+ invited talks and keynotes on generative AI for the Michigan CPA Association, AAA, and XBRL.

Selected Publications