🎓 MS CS (AI) @ USC · Graduating May 2027 | 🤖 Robotics Research @ USC ISI | ⚛️ Nuclear M&S
I build machine learning systems.
My work spans much of the modern stack. I've shipped a production retrieval-augmented generation engine serving thousands of queries a day with vector search and reranking, built a self-hosted tool-calling agent, trained hierarchical reinforcement learning agents, classified 96-channel time series with Transformers, and built Gaussian-process surrogates for Monte Carlo physics simulations. The domains have ranged from search and documents to fuel cell fleets, modular robots, and nuclear reactors.
@ USC Information Sciences Institute — Polymorphic Robotics Lab (Sep 2025 – present)
- Developing a distributed inverse kinematics algorithm for modular self-reconfigurable robots: each module runs its own local solver and coordinates by message passing over the physical connection graph, applying damped least squares to its local Jacobian column. No single controller needs the full kinematic chain
- Moved constraint enforcement into individual modules, making the solver morphology-agnostic. The same algorithm runs on arbitrary assembled topologies without re-deriving kinematics
- Write the physics simulation stack in C; tuned constraints in Unreal Engine and MuJoCo (including unit-scale inertia corrections for critical damping) and built the sim-to-real pipelines validating on physical SuperBot hardware
@ Bloom Energy — Failure Analysis, Quality & Reliability (Summer 2026)
- Scaled a telemetry-based failure detector for DC-DC power converters to the full confirmed-failure population: recall 82.2% → 92.5%, zero new false positives, ~27-day median detection lead
- Owned a failure investigation end to end, reliability modeling through physical flow-bench testing, and presented it at the company-wide review
- Proposed and built a locally-hosted conversational agent (Ollama, tool-use architecture) over the fleet analytics, plus a config-driven CLI, so the team's PhD specialists could run analyses without me
- Researched single-event effects device physics and neutron detection instrumentation for a radiation-effects test campaign
@ LineSlip Solutions (2024 – 2026)
- Production RAG pipeline on Llama 3.1-8B with Elasticsearch retrieval and custom reranking, serving 10K+ queries/day; 35% accuracy improvement, 40% latency reduction via INT8 quantization
Applying AI and optimization where physics constrains the solution:
⚛️ Nuclear — reactor modeling & simulation, surrogate models, digital twins, fault diagnosis · 🔋 Energy Systems — degradation, predictive maintenance, fleet reliability · 🤖 Robotics — distributed control, modular systems, sim-to-real · 🧪 Physics-Informed ML — learning dynamics without discarding the governing physics
Monte Carlo neutron transport in OpenMC, with a surrogate measured against the stochastic uncertainty of the calculation it replaces
- 488-run transport campaign over a PWR pin-cell lattice, sweeping enrichment, fuel temperature, moderator density, and pitch
- Verified by Shannon-entropy source convergence, 1/√N scaling of σ, and spectrum shape; quoted uncertainty confirmed against seed-replicate scatter
- Gaussian-process surrogate for k-infinity: 25 pcm error against the true response surface (noise deconvolved in quadrature) vs 90 pcm transport σ, at ~10⁷× lower marginal cost
- Reactivity coefficients validated against transport runs the surrogate never saw; Doppler 0.3σ, enrichment 0.2σ. The moderation optimum found by sweeping density and by sweeping pitch agrees to 0.06%
- Tech: OpenMC, scikit-learn, ENDF/B-VII.1, Docker
Deep learning on time-series reactor sensor data for accident detection and diagnosis
- Semi-supervised anomaly detection across 96 operational parameters and 18 accident scenarios
- Multi-class classification with SHAP-based root-cause attribution, validated against known accident physics
- Data: NPPAD dataset (Nature Scientific Data), PCTRAN PWR simulator · Tech: PyTorch, scikit-learn
Real-time video style transfer at 6.45 FPS on 1080p
- RAFT optical flow for frame-to-frame temporal consistency; trained on 118K images with distributed data-parallel training across 4 GPUs
- Tech: PyTorch, DDP, RAFT
Fine-tuned Llama 3.2 coaching app — Session Winner, 55th Annual Senior Design Conference
- QLoRA PEFT for culturally-aware generation; full-stack web application
- Tech: Python, TypeScript, Llama, QLoRA, Firebase
Nuclear & Reactor Physics: OpenMC · Monte Carlo neutron transport · k-infinity eigenvalue calculations · cross-section libraries (ENDF/B) · reactivity coefficients & Doppler feedback · single-event effects (SEE/SEB) · JESD89A
Modeling, Simulation & UQ: Surrogate models (Gaussian processes) · digital twins · uncertainty quantification · Latin hypercube & Sobol sampling · design of experiments · Kaplan–Meier · Weibull hazard modeling · probability calibration · numerical solvers (damped least squares, null-space methods) · Unreal Engine · MuJoCo · sim-to-real
ML & Agents: PyTorch · scikit-learn · Transformers · distributed training (DDP) · anomaly detection · SHAP · LLM fine-tuning (QLoRA) · tool-use architectures · local model deployment (Ollama) · production RAG
Languages & Systems: Python · C · SQL · C++ · Java · Shell · Linux · Docker · Git · CI/CD · FastAPI · Postgres · Elasticsearch
Creative Collaborator: AI-facilitated UI for Creating Engaging and Insightful Memes — first author | AHFE International, 2024 | DOI: 10.54941/ahfe1005579
Distributed Inverse Kinematics for Modular Self-Reconfigurable Robots — in preparation, IROS 2027
LinkedIn · [email protected] · Los Angeles, CA