B.Sc. Computer Science · University of Southern California, Expected May 2028
Undergraduate researcher in USC LiraLab (Prof. Erdem Bıyık) and with Prof. Jiapeng Zhang; Lead Systems Engineer on NASA L'SPACE MCA Team 19 and NPWEE Team 10; autonomy software for University Rover Challenge 2026. I work where ML, robotics, and systems meet real constraints: distribution shift, sparse preference feedback, exact verified reuse, and permanently shadowed lunar operations.
| Work | Status | Links |
|---|---|---|
| TARE — Tool Annotation Reuse, Exact. Black-box pin-and-replay across reference database releases; byte-identical to a full stock rerun or refuse. First author with Prof. Jiapeng Zhang. | Preprint · targeting USENIX OSDI '27 | tare · arXiv:2611.02182 · project site |
| LiraLab — Few-shot latent preference adaptation / PEBBLE Experiment 1 under noisy teachers | Active undergraduate research | liralab · liralab-experiment-1 |
| Visual Compounding Problem — Closed-loop vision stacks that pass isolated tests and fail when errors couple | Independent benchmark + manuscript prep | visual-robot-learning |
Ten independent, reproducible research frameworks — ~33,000 lines of Python, ~5,300 lines of tests, 310 pytest cases. Each module implements synthetic ground-truth DGPs, YAML-configured benchmarks, and structured evaluation reports. The unifying question: how do modern ML estimators degrade as their identifying assumptions are systematically violated?
| # | Module | Description | Repo |
|---|---|---|---|
| 01 | Causal ML | Meta-learners (S/T/X/R/Causal Forest), TMLE, NOTEARS, PCMCI, off-policy evaluation | causal-ml-decision-making |
| 02 | Foundation Model Adaptation | LoRA, adapter bottlenecks, prefix tuning, influence functions (LiSSA, TracIn), calibration | foundation-model-adaptation |
| 03 | Reinforcement Learning | CQL, IQL offline RL; CMDP Lagrangian safe RL; ensemble world models; CEM planning | reinforcement-learning-real-environments |
| 04 | Bayesian Deep Learning | Variational BNNs, MC Dropout, deep ensembles, conformal prediction (RAPS), calibration | probabilistic-bayesian-deep-learning |
| 05 | Generative Modeling | DDPM/DDIM, classifier-free guidance, flow matching, diffusion posterior sampling | generative-modeling-scientific |
| 06 | Graph ML | GCN, GAT, GIN, GraphSAGE, DiffPool, temporal GNNs, link prediction | graph-ml-relational-systems |
| 07 | Self-Supervised Learning | SimCLR, MoCo, BYOL, VICReg, Barlow Twins, MAE; linear probe + k-NN eval | self-supervised-representation-learning |
| 08 | Neurosymbolic Reasoning | Differentiable forward chaining, neural theorem prover, typed program synthesis | neurosymbolic-reasoning |
| 09 | Federated & Privacy ML | FedAvg/FedProx/FedNova, DP-SGD, Rényi DP accounting, pFedMe, secure aggregation | federated-privacy-preserving-ml |
| 10 | Scientific ML | PINNs, E(3)-equivariant GNNs, scVAE, FNO, Bayesian optimization for materials | ml-for-science |
| Project | Description | Repo |
|---|---|---|
| TARE | Exact incremental re-annotation wrapper for unmodified CLI tools (HMMER / Infernal bake-offs; refuse path for DIAMOND) | tare |
| SpaAIder | Local-first multi-agent runtime — TypeScript orchestration, Python FastAPI memory substrate, episodic–semantic–vector memory, dependency-aware parallel scheduling | SpaAIder |
| URC Autonomy Stack | ROS 2 (Humble) autonomy for University Rover Challenge 2026 — 7-target mission orchestrator, EKF fusion, GNSS waypoint nav, ArUco visual servoing, YOLO detection | urc-cv |
| ClinRad-Bench | Clinical ML evaluation under prevalence shift (2% / 12% / 35%), calibration (ECE), selective prediction, bootstrap CIs | ClinRad-Bench |
| KumonScan | Desk + QR attendance for Kumon centers (Neon Postgres, Math / Reading / Both session rules, timeapi.io clock) | KumonScan |
| Curriculum Vitae | LaTeX academic CV (auto-compiled) | curriculum-vitae |
Curated featured work. GitHub pins only six repos; this profile is the full directory.