CV · Jiaqi Gan

Education

Degrees, schools and dates
Degree	School	Date
Ph.D. Candidate in Physics, Biophysics Research	University of Pittsburgh	Expected Apr 2027
M.S. in Physics	University of Pittsburgh	2023
B.S. in Applied Physics	Shanghai University	2020

Graduate Researcher

Aug 2022 – Present

Framework development: Developed LogicDynamo in PyTorch to infer cellular dynamics from single-cell RNA and RNA-velocity data, implementing the model-fitting workflow that connects biological measurements to dynamical-system analysis.
Interpretable modeling: Designed a generalized additive ODE model that separates individual regulatory effects from pairwise interactions, using saturating dose-response functions and sparse regulation to make the learned dynamics interpretable.
Validation: Evaluated the framework across 2 simulated dynamical systems and 5+ real single-cell datasets; implemented ensemble uncertainty estimation and computational gene perturbations to assess noise sensitivity and produce regulator-effect reports for experimental follow-up.

Software development: Co-developed GraphVelo, an open-source Python package for estimating the direction and speed of cell-state changes; implemented local tangent-space projection to refine noisy RNA-velocity inputs and support mapping between RNA and chromatin-accessibility measurements.
Benchmarking: Performed ground-truth tests across 3 simulated trajectory types (linear, cyclic and branching), evaluating direction, magnitude and sign with cosine similarity, RMSE and sign accuracy; demonstrated retention of speed information lost by cosine-kernel projection.
Sensitivity analysis: Tested robustness to noise and dimensionality reduction, comparing velocity estimates with known simulated dynamics; found better speed agreement in PCA than UMAP, identifying limits on quantitative interpretation of cell-state visualizations.

ML Research Intern

Jul 2026 – Present

Data generation: Built a pipeline that produced 3,395 synthetic calls and 15,102 training targets. Enforced prefix-only reply inputs and separate dialogue and summary reviews.
Post-training: Fine-tuned six 0.6–4B Qwen3, Llama and Gemma models with one rank-16 LoRA adapter per model for replies and summaries. Balanced both tasks in each update and applied loss only to target tokens.
Evaluation: Evaluated 4,608 outputs on 192 held-out calls with a custom five-criterion LLM judge. Improved reply qualification by 43–74 percentage points and summary qualification by 43–68 points across all six models.

Benchmark design: Built CallScreenBench end to end to evaluate call screening and owner notes across 22 call archetypes; implemented stateful caller simulations, 5 quality-measure groups and 3 guardedness diagnostics, evaluating 6 locally run language models.
Metric validation: Stress-tested evaluation metrics with 6 scripted policies and estimated uncertainty with paired bootstrap intervals; exposed a scoring loophole when a hangup-and-copy agent achieved 100% caller-detail recall without meaningful dialogue, demonstrating the need for counter-metrics.
Predictive modeling: Analyzed 10,211 real inbound calls and 330,956 transcribed turns to predict later sensitive-information requests; a bag-of-words baseline reached 0.87 ROC-AUC from the first 8 caller lines on 1,276 calls held out by phone number.