Resume

Simon Fraser University

Python, LLM APIs, ROUGE, pandas, matplotlib

• --Introduced a new method of evaluating LLM model unlearning based on the RWKU paper.
• --Built a Wikipedia data pipeline for same-name entity collisions: collected 73k+ disambiguation pages and extracted candidate entities + lead text for filtering.
• --Ran 24k+ probe queries over 200 high-profile entities (GPT-4.1) and wrote evaluation scripts to score target vs. neighbor behavior using ROUGE-L recall and aggregate reports.
• --Using prompt-only in-context unlearning, reduced target-answer match from 0.96 to 0.08–0.11 ROUGE-L while keeping neighbor probes high (0.97–0.98); documented failure modes and limitations.

Rust

• --Implemented a decoder-only LLaMA-2 inference engine in Rust: RMSNorm, RoPE, multi-head attention, SwiGLU FFN, and KV-cache autoregressive decoding.
• --Built weight loading, tokenization, and sampling (temperature, top-p) with a focus on correctness and clean Rust design.

Python, NumPy, PyTorch

• --Implemented a decoder-only Transformer (3 layers, 4 heads, d_model=768, ~164k params, context=256) with token/positional embeddings, causal multi-head attention, and pre-LayerNorm residual blocks for next-token prediction.
• --Built autoregressive generation with context-window cropping and configurable sampling strategies (temperature, top-k) to control output diversity and quality.
• --Validated model training with loss convergence curves and qualitative text generation samples; achieved stable training behavior across 50k+ training steps.

Python, NumPy

• --Implemented a NumPy-based autograd engine enabling neural networks to compute gradients automatically and learn via backpropagation.
• --Supported core tensor ops and edge cases (broadcasting, reductions), enabling stable training behavior comparable to common DL frameworks.
• --Built a small nn layer/optimizer stack (Linear, normalization layers, SGD/Adam) and training utilities (gradient clipping, weight decay). Added tests and demo training runs to verify gradient correctness and convergence.

Python, SQLite, SQL

• --Refined a 10-entity conceptual model into an 8-table BCNF-compliant SQLite schema, reducing invalid inserts to zero by enforcing FK/PK constraints and validating reviewer eligibility before writing.
• --Developed a Python-based, on-demand reviewer-assignment workflow (CLI) that automated conflict-of-interest checks and enforced reviewer quotas (≤3/proposal), reducing manual review-assignment effort.
• --Designed and implemented 6 SQL-driven analytical workflows processing grant applications with complex business logic including multi-constraint eligibility validation across 8 normalized tables.