Experience

AI/ML Intern

Signal-integrity designs built in Ansys HFSS 3D Layout had to be redrawn by hand as Cadence Allegro boards, about an hour of CAD work per design. Via geometry was also tuned by hand.

• A four-stage Python/SKILL pipeline (PyAEDT, pyedb, Shapely) that converts HFSS 3D Layout designs into native Allegro boards. It automates geometry extraction, unit and layer remapping, and copper-void decomposition.
• Net-aware via classification and padstack generation (GND vs. signal; buried, microvia, backdrill) taken directly from extracted AEDT variables, producing a reusable drop-in Allegro module per design.
• A workaround for a partially broken Allegro SKILL API (void, circle and bulk-deletion functions): all geometry is precomputed in Python and the pipeline emits only SKILL calls confirmed to work.
• A multi-objective via optimizer (NSGA-II plus Bayesian optimization with a Gaussian Process surrogate) that tunes via size, length and placement against return-loss and TDR impedance targets, pre-screening candidates before full HFSS simulations.

• Per-design translation from about 1 hour to about 21 seconds
• 70% less manual SI-to-CAD effort
• 80% less manual via configuration

• Python
• SKILL
• PyAEDT
• pyedb
• Shapely
• Ansys HFSS
• Cadence Allegro

Undergraduate Researcher, with Prof. Ryan Eric Grant

Measure whether offloading data exchange to a SmartNIC beats host-based communication on HPC clusters, and what it costs the host.

• MPI microbenchmarks on NVIDIA BlueField-3 SmartNICs, measuring end-to-end latency across host-only and offloaded architectures under varying message sizes and workloads.
• Benchmarking and automation scripts in C++ and shell, with SLURM job scheduling across test runs and the test suite under Git.

• Up to 4× lower communication latency than host-based approaches
• Host overhead kept under 200 µs

• C++
• MPI
• SLURM
• Bash
• Linux
• BlueField-3

Machine Learning Engineer

Detect AI-generated video reliably, including footage from generators the model has not seen and video that has been degraded.

• Fine-tuned OpenAI's CLIP for binary real/fake classification on the GenVideo dataset (1M+ samples), adapting its vision-language embeddings and training procedure.
• Training and data pipeline components in C++ and CUDA.
• Validation with cross-generator and degraded-video tests.
• Co-authored a paper for the Canadian Undergraduate Conference on AI on temporal-spatial inconsistency detection in AI-generated video.

• Detection accuracy from 58% to 87%

• PyTorch
• CLIP
• C++
• CUDA
• Python

Software Developer

• Built an end-to-end image classification system (FastAPI, GPT-4o Vision) that labeled 10,000+ images with a multi-level taxonomy; it was presented to the Schulich School of Business.
• Designed an AWS S3 dual-bucket pipeline that cut manual data handling by 70%, with backend services sustaining 99% successful requests under concurrent load.

Machine Learning Engineer, with Prof. Chi-Guhn Lee

• Built an AI agent for Nestlé sales forecasting and insight generation using LangChain, LangGraph, RAG and SQL.
• Curated and labeled 1,000+ images with YOLOv8 for Hydro One, with automated filtering that removed manual review of non-transmission-line photos.