Zikun Fu

Natural-language questions name database concepts indirectly, so text-to-SQL systems need a reliable grounding layer. I formalized Database Entity Recognition as sequence labeling over {O, Table, Column, Value}, curated a 1,000 human annotated benchmark from Spider and BIRD, and built a 15,026-example synthetic annotation pipeline that derives labels from paired SQL using AST parsing and integer linear programming. A T5-Large tagger reaches 80.8% F1 open-world; adding a MiniLM verifier that grounds spans against the target schema raises this to 93.2%. Published at IEEE IRI 2025; code is MIT license and the dataset is public under CC BY 4.0.

A Python/DuckDB analytics assistant that answers business questions over simulated optometry-clinic data. Four mechanisms constrain the model's data path: a sqlglot-parsed SQL allowlist, a presentation database physically separated from the hidden oracle tables, charts and tables built from query results rather than model generated numbers, and exact SQL provenance attached to every answer. In the published scoring harness, every answerable question matched the database, and questions the data could not support were declined rather than answered with invented values. Public code includes offline gate, executor, rendering, and live bypass safety tests. Built on the Hermes agent platform with Open WebUI as the front-end.

A legacy Symfony/Laravel research platform needed replacing without losing years of multilingual group data. I led the migration to a Dockerized LimeSurvey 6.x + MySQL stack, wrote the Python ETL that imported 3,400+ multilingual records across 60+ research groups through the LimeSurvey REST API with field-mapping and validation passes, built PHP plugin modules and Chart.js dashboards for the reporting views, and deployed the replacement on DigitalOcean behind an Apache reverse proxy with automated TLS. Scope covered the migration plan, the data layer, and the production cutover; the platform remains in use by the faculty.

An automated pipeline for a local optometry clinic that turns its scattered online footprint (Google Business Profile, reviews, PageSpeed, geo-grid rankings, backlinks, page health) into a graded 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 used for interpretation only.

A marking assistant for scanned handwritten assignments in a compilers course, built and run with the course instructor's approval. Question page images go to Claude (via OpenRouter) alongside model solutions and rubric rules, and the pipeline returns score suggestions with confidence and legibility ratings, a flag list, and a full audit trail. Flagged questions are read and rescored by the human marker through a scripted adjustment workflow before any mark is released. Prompt caching held model cost to roughly $0.15 per student.