About Me

Equifax · Atlanta, GA

• Fine-tuned and evaluated Gemma 3 4B and Gemma 4 E4B via QLoRA on Vertex AI Workbench to classify ~8B financial transactions across 30 primary and ~300 total categories.
• Adapted Gemma 4 E4B for legacy infrastructure by modifying model internals to support older library versions.
• Tuned quantization and dtype configurations for NVIDIA T4 GPUs, raising test accuracy from 75% to 99%.
• Used contrastive learning to fine-tune ModernBERT and fit a linear classifier, achieving the same results with ~100x speedup.

Georgia Tech Research Institute · Atlanta, GA

• Extended a physics-informed neural network surrogate modeling rocket and flare infrared intensity sequences for military use.
• Designed and A/B-tested an intermediary-value loss module anchoring the physics loss to mid-network intermediate predictions.
• Implemented push-forward training — feeding model predictions back as inputs during training — improving test MSE by over 12%.
• Contributed to a YOLO-based object-detection pipeline for moving vehicles, using GPS coordinates to assign global position labels.

Georgia Tech VIP · Atlanta, GA

• Applied multi-objective genetic programming with LLMs to evolve a CNN image classifier; benchmarked island configurations, finding multi-LLM setups yielded greater early diversity, with certain combinations degrading gene validity over time.
• Designed a multi-LLM island architecture supporting multi-population evolution, migration, and checkpointing.
• Built a reproducible HPC pipeline on PACE-ICE using SLURM sbatch, H100/A100 allocation, and shared model storage.

MessageGears · Atlanta, GA

• Standardized Angular/JSP actions UX with confirmation modals and shipped multiple UI fixes.
• Rewrote Groovy/Spock API tests and split suites by endpoint, cutting runtime from minutes to seconds and accelerating builds.

Spectral Text Encoder

• Built a spectral text encoder mapping tokens to sine triples summed into a fixed-length waveform embedding.
• Raised in-batch contrastive validation accuracy from 56% to 78% using frequency-separation and per-channel InfoNCE losses.

Wave Scalar Field Sequence Generator

• Built a PyTorch video transformer learning 2D wave dynamics from height frames via block-causal attention and RoPE.
• Reduced rollout error via a pushforward horizon schedule, velocity-supervision aux head, and residual conv refinement.

Agent-Powered Creator-Game

• Built an open-ended game converting spoken ideas into explorable 3D scenes using an agentic Gemma 4 loop with custom tools.
• Ran fully local inference on Apple Silicon using quantized models, FLUX for image generation, and TRELLIS for 3D mesh.

Zero External Data LLM Reinforcement Learning (IP)

• Designed a creator/solver self-play curriculum with verifiable rewards and a difficulty-calibration reward, using GRPO.
• Extended it to multi-turn RL (20 Questions) with a custom per-turn reward, then A/B tested periodic role-reversal.