Sunanda Gamage Email & Phone Number

London, On, Ca

Singapore

I am leading a team of AI engineers to build a new analytics engine for high-volume call centers, working on speech-to-text and LLM model development.- Co-authored the AI capabilities roadmap for the company, including analytics products, data generation pipelines, LLM architectures and fine-tunes, evaluation benchmarks, and GPU budgets.- Developed a range of AI capabilities for the analytics engine: speech-to-text, summarization, classification. real-time and batch processing of large audio datasets.- Training robust open source speech-to-text AI models (Whisper) for state-of-the-art performance in Asian Pacific locales and implementing efficient inference pipelines (4% - 5% word error rates).- Fine-tuning open source LLM models (Llama, Mistral) to generate high quality outputs (summarization and classification) for in-domain data using parameter-efficient fine-tuning techniques (LoRA, QLoRA).- LLM distillation: Developing LLM-based data generation pipelines to generate high quality training data for conversational settings, and using them to fine-tune smaller LLMs.- Implemented and successfully demoed AI Voice Assistant for customer care call centers (stack: speech-to-text, LLM workflows with RAG, text-to-speech).- Designed and implemented comprehensive benchmarks to evaluate speech-to-text and NLP tasks (accuracy, robustness to noise and input variations, efficiency).

London, Canada Area

Research topic: Attribution Robustness of Neural Networks- I developed a novel framework of loss functions for attribution robust training of neural networks. A new robust objective derived from this framework achieves 5% higher attribution robustness with 2x - 10x speed up over current robust neural net training methods.- I further improved the efficiency of the robust training by developing new methods including hybrid standard and robust training, using a fast one-step attack, and optimized computation of integrated gradients.- I also designed and implemented efficient and accurate classification systems (neural net-based) for several domains, including; - Network intrusion detection (network packet data) - Structural damage detection (vibration signals) - Industrial time-series classification (multi-variate sensor data)Skills: deep learning experimentation at scale on distributed GPU clusters, doing research and science communication, designing and implementing data and model pipelines for applied machine learning problems.Publications: 3 journal papers and 1 conference paper, 450+ total citations.Key coursework: Machine Learning, Applied Probability Theory, Random Signals & Adaptive Filters.

Classes: Cloud Computing, Programming Fundamentals, Web Technologies, Software Requirements AnalysisHelped students work through programming problems, debugging, and designing & developing software systems.

Ladenburg, Germany

I developed a self-supervised neural net pre-training flow for label efficient industrial process signal classification (a domain where labeled data is hard to obtain), achieving over 95% accuracy with 10's of labeled data examples.A year long collaboration with the ABB research team led to a conference paper and poster publication.

Colombo

I designed and implemented a high performant market data system (client: Italian Stock Exchange) - Designed a micro-services architecture with highly concurrent processing modules to best utilize hardware.- Wrote computation modules to approximate relevant market data statistics (eg: information on order book depth, trade execution, volume and liquidity metrics etc.)- Wrote high-throughput incoming message gateways and information broadcast components in C++.- Designed and implemented fault tolerance mechanisms for all components of the system (process replication, cold and warm startup procedures) and wrote protocols to test fault tolerance.- Test-driven C++ development (unit, component and performance regression test suits on Atlassian Bamboo) in an Agile project management setting.- Developed project onboarding material and conducted training sessions for new team members.

Malambe, Sri Lanka

As a member of an R&D team, I developed modules of a new application framework and prototyped stock exchange systems on it.- Writing components in C++ in a large, high-throughput distributed system (gateways, message validators, order book matching engines, asynchronous data storage).- Profiling and optimizing C++ programs for latency, throughput and memory correctness (Valgrind tools).- Writing and profiling fast inter-thread data transfer methods (classic locked queues and lock-free queues).

As an intern, I worked on the following key projects of the company involving large clients, and learned a lot about how networking solutions are deployed in the real world.- Implemented wireless networks and wi-fi offloading solutions (Cisco switch, router, access point installation, configuration & monitoring) for several large clients (banks, internet service providers, hotels).- Deployed and maintained Network Management Systems (NMS: Cisco and SolarWinds) and Network Access Control (NAC) systems on large networks (1000+ globally distributed servers).- Developed a software solution to monitor network usage of clients connected to a wireless LAN andto display real-time visualizations on a dashboard (connectivity, throughput, latency).- On-site troubleshooting of issues in client networks (access issues, speed issues & network optimization, switch and router installation & configuration).

Doctor Of Philosophy - Phd, Artificial Intelligence, 94% Average For Coursework

Activities and Societies: Founded the Machine Learning Journal Club at Western ECE Research topic: Attribution robustness of neural.

Bachelor Of Science (Bsc), Electronics And Telecommunication Engineering, 3.76/ 4.2 (First Class Honours, Dean'S List X 6)

Activities and Societies: Rowing Crew of University of Moratuwa Awards: Fully funded admission and completion of the 4-year B.Sc. Hons..

Education record

Activities and Societies: The United Nations Club (committee member). International Junior Science Olympiad, Indonesia (member of the Sri.