Nicholas – Python, AWS, PyTorch

An Indigenous‑led AI initiative using modern language and machine learning tools to help protect Indigenous knowledge, strengthen community voices, and support long‑term cultural and economic resilience.

• Designed and documented the end‑to‑end technical architecture for a new generative AI product focused on low‑resource languages.
• Researched state‑of‑the‑art techniques for fine‑tuning large language models on low‑resource and culturally specific datasets and translated findings into concrete experiments.​
• Built the LLM fine‑tuning pipeline, including data preparation, training, and evaluation workflows, and delivered working prototypes.​
• Defined and maintained the technical roadmap, prioritizing model improvements, infrastructure needs, and experiment cycles.
• Led data infrastructure design and collection processes for sensitive and low‑resource language data, ensuring quality and responsible handling.

An AI‑powered feeding optimization platform for land‑based fish farms that uses real‑time video analytics to monitor fish behavior and uneaten feed, then automatically adjusts feeding to match actual appetite. By continuously tuning feed delivery, it improves growth and feed conversion while reducing waste, operational costs, and environmental impact of aquaculture operations.

• Led AI product scale‑up from a single prototype to large‑scale deployment across many customer sites in multiple countries.
• Architected and implemented the MLOps pipeline, including experiment tracking, model registry, and artifact lineage, to enable fast, reliable model iteration and deployment.
• Designed business‑aligned evaluation metrics, including a custom composite metric focused on worst‑case operational scenarios and real farm KPIs rather than academic benchmarks.
• Built a real‑time fleet monitoring suite to track model health, data drift, and system status across all installations, sharply reducing time to detect and fix production issues.
• Led an active learning program and team to design and roll out workflows that improved data labeling efficiency and lowered annotation costs.
• Drove the transition to data‑centric AI practices, prioritizing dataset quality and targeted edge‑case slices to improve performance without relying on ever‑larger models.

The first generation of the ReelAppetite solution, this system utilized on-premise servers and cloud computing to validate the market viability of autonomous feeding. It established the foundational control system architecture and computer vision capabilities that allowed the company to secure its initial customer base and validate the core technology.

• Resolved critical thermal constraints: Identified and mitigated thermal bottlenecks in the production hardware, preventing system throttling and ensuring reliability in high-temperature environments.
• Engineered custom power profile: Developed a specialized low-power operational mode for the edge hardware (NVIDIA Jetson), balancing compute resources to maintain performance targets within a reduced power budget.
• Conducted receptive field analysis to identify and remove redundant network parameters, reducing model size by 75% without compromising detection accuracy.
• Applied Post-Training Quantization (INT8) to double inference throughput, allowing sophisticated models to run at high frame rates on edge devices.
• Achieved real-time performance: Successfully optimized the full computer vision pipeline to meet strict frame-rate requirements (60 FPS) while adhering to rigorous thermal and power safety limits.
• Delivered a software-based optimization solution that resolved hardware limitations, preventing the need for costly physical retrofits or recalls.

This engineering initiative focused on optimizing the ReelAppetite Edge AI pipeline to operate reliably within strict hardware and environmental constraints. The project involved deep optimization of deep learning models and system resources to ensure high-performance object detection could run continuously on passively cooled edge hardware in harsh environments.

• Resolved critical thermal constraints by identifying and mitigating thermal bottlenecks in production hardware, preventing throttling and ensuring reliability in high‑temperature environments.
• Engineered a custom low‑power operational profile for NVIDIA Jetson–class edge hardware, balancing compute resources to stay within a reduced power budget while maintaining performance targets
• Performed receptive field–driven model pruning to remove redundant network parameters, shrinking model size by about 75% with no loss in detection accuracy.
• Implemented post‑training INT8 quantization to approximately double inference throughput and enable high‑frame‑rate execution of complex models on edge devices.
• Optimized the full computer‑vision pipeline to sustain real‑time performance at 60 FPS under strict thermal and power safety limits
• Delivered a software‑defined optimization solution that compensated for hardware constraints, avoiding costly physical retrofits or field recalls.​

A research project on neural language models that introduced human‑interpretable controls for text generation, enabling users to steer style and structure without retraining and resulting in a peer‑reviewed publication at an international ML conference. It showcases Nick’s ability to turn complex autoregressive models into practical, controllable creative tools through advanced conditioning techniques for generative systems.

• Developed new methods for controlling sequence generation in neural language models, leading to a peer‑reviewed publication at an international machine learning conference.​
• Conducted exploratory research using Gaussian processes to model and quantify predictive uncertainty in language model outputs.​
• Trained neural language models on large‑scale text corpora using distributed GPU clusters, including data preprocessing, experiment configuration, and results analysis.​​
• Regularly reviewed and presented state‑of‑the‑art machine learning and natural language generation papers to the lab, distilling key ideas and proposing follow‑up experiments.