Chenyu Fang, Ph.D.

Personal project · 2026

A live service that forecasts tomorrow's electricity load for New York's 11 grid zones and the statewide total, then scores itself against what actually happened.

• On par with NYISO's own day-ahead forecast: 4.80% vs. 4.85% average hourly error (MAPE), no significant difference under a paired bootstrap.
• 53% lower error than the best naive baseline over a 12-month backtest.
• Conformal calibration raised how often actual load falls inside the 80% forecast range from 57% to 77%.
• A Temporal Fusion Transformer served as ONNX through FastAPI on GCP, with daily ingestion and a LightGBM fallback; deployed to Kubernetes (kind in CI on every push, and a 72-hour GKE Autopilot trial).

• PyTorch
• TFT
• LightGBM
• ONNX
• FastAPI
• GCP
• Kubernetes

Personal project · 2026

A voice mock-interview app for ML and AI engineering interviews, grounded in rubrics and tailored to each user's résumé and target job.

• A LoRA-tuned Qwen3-4B grader agreed with Claude's grades at QWK 0.94 (1.0 = perfect agreement) across five seeds, against 0.79 for the best baseline.
• vLLM serves a grading call in 30 ms (p95).
• Hosted grading runs on DeepSeek at about 1/200th of Claude's cost, with 94% of grades within one point of Claude.
• Hybrid BM25 + BGE retrieval raised Recall@5 from 72% to 84%.

• LoRA / PEFT
• Qwen3-4B
• vLLM
• LLM evaluation
• RAG
• FastAPI

Team project with a second engineer · Techlent Machine Learning Engineer Fellowship (training program) · 2026

An LLM bid-review system where models only read, a deterministic rule engine decides, and every value is cited to its page. I owned the FastAPI backend, the LLM gateway and the bid checks.

• Every extracted value is verified by a text-layer match or an independent second read; anything unverified goes to a person.
• Chose a Postgres job queue over LangGraph after a 12-scenario fault-injection harness: about 5 s recovery and 0 lost jobs.
• On a synthetic tender, rule drafting matched 15/15 checklist items with no unverified rules, for $0.02.
• Deployed on Azure Container Apps with Entra ID sign-in and OIDC deploys that store no secrets; the demo needs no sign-in.

All bid numbers are on synthetic data.

• Document AI
• FastAPI
• PostgreSQL
• Guardrails
• Azure

Research Assistant, University of Oklahoma · 2026 to present

• Compared XGBoost, a re-implemented LorentzNet and a Particle Transformer under rules fixed in advance, three seeds and paired bootstraps; ML beat the cut-based selection by 9‑21% in significance.
• LorentzNet matched the transformer at a tenth of its size; probability weights in place of random accept/reject gave a 9‑19x larger effective sample.

Ph.D. research, University of Oklahoma · 2018 to 2025

• Python/C++ pipelines and XGBoost classifiers that separate a rare signal from backgrounds about 10⁷ times larger.
• The final model recovers 41.2% of signal at a 7.7% false-positive rate, a projected 6.2σ result; interpreted with SHAP and packaged as a tested, versioned CLI.