Resume

Entity-Aware LLM Unlearning Evaluation · Simon Fraser University

Python, LLM APIs, pandas, matplotlib, BM25, ROUGE-L

• --Extended the RWKU evaluation framework to measure whether LLMs can suppress knowledge about specific people while preserving answers about similarly named individuals.
• --Built a Wikipedia data pipeline processing 73K+ disambiguation pages and released a reproducible benchmark of 1,000 ambiguity groups, 3,000 entities, and 6,000 evaluation questions with source provenance.
• --Conducted a 54K-response evaluation across three LLMs, comparing baseline, prompt-only suppression, and retrieval-guided approaches; automated scoring and reporting for target forgetting and neighbor preservation.
• --Developed and evaluated identity-aware routing, reducing false links on 800 held-out neighbor questions from 19 to 9 (53%) while maintaining 99.25% target-linking recall.

Python, FastAPI, React, TypeScript, ESP32-C5, BLE, NumPy

• --Built a wearable movement-analysis system that streams 6-axis IMU data from an ESP32-C5 at 50 Hz over custom 18-byte BLE packets to a FastAPI backend and live React dashboard.
• --Developed mounting-independent PCA calibration, deterministic state-machine rep segmentation, and per-rep and set-level metrics for movement consistency, range, smoothness, and tempo.
• --Designed a bounded, fail-soft coaching pipeline using deterministic deviation detection before Jev, Gemini, and ElevenLabs; added automatic provider fallbacks, CI, and 99 backend tests.

C, Python, STM32F411, ESP32-C5, MPU6050, BLE, I2C, UART

• --Built a gesture-controlled lock prototype combining an STM32F411, MPU6050 IMU, two XIAO ESP32-C5 boards, and servo actuation.
• --Implemented on-device inference over 75-sample motion windows, extracting 27 acceleration features and running a binary linear classifier on the STM32.
• --Integrated an I2C-to-UART-to-BLE control path and a Python data pipeline for collecting labeled IMU windows, training the classifier, and exporting model parameters to firmware.

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.