Zikun Fu

A chat assistant that answers business questions over a simulated optometry clinic's data with structural anti-hallucination guarantees: a read-only DuckDB engine, a sqlglot-parsed SQL allowlist, charts built from query results (never model-typed numbers), and exact SQL provenance on every answer. An oracle-backed scoring harness measures 100% grounding accuracy and 0% hallucination; 59 offline unit tests plus a live bypass-safety suite verify the guarantees hold. Built on the Hermes agent platform with Open WebUI as the front-end.

An automated pipeline for a local optometry clinic that turns its scattered online footprint — Google Business Profile, reviews, PageSpeed, 25-point geo-grid rankings, backlinks, on-page health — into a graded, client-ready report with a 30/60/90-day action plan. Built-in cost gates (estimated spend approved before any paid API call; a full run is under $1) and human-review gates for health-content compliance (every clinical claim is flagged for clinician sign-off; review text is never reproduced verbatim). LLM use is interpretation-only: the AI writes the narrative, never the numbers.

Auto-grades scanned handwritten final exams for a 4th-year compilers course by sending per-question page images to Claude (via OpenRouter) with model solutions and strict grading rules. Produces per-sub-question scores with confidence and legibility ratings, a human-review flag list, and a full audit trail; prompt caching keeps costs low — ~70 students graded for $10.34 total (~$0.15/student). Flagged items go through a scripted human-review adjustment workflow.