Building Value with AI and
Large Language Models

Who Am I?

I’m a GenAI engineer and consultant with a PhD in cognitive neuroscience, specialising in building intelligent agent systems that adapt, iterate, and scale, while helping others to do the same.

I’ve delivered LLM-powered applications across biotech and education, from a patent-focused retrieval system at Insmed to live LangGraph agents used by more than 3,000 learners on DataCamp. From working with orchestration tools like LangChain and LangGraph, LLM providers such as OpenAI or Hugging Face, or designing semantic search with various vector databases, I turn complex, domain-specific data into fast, functional workflows while helping teams skill-up along the way.

My machine learning journey began in 2013, decoding high dimensional fMRI data to understand how the brain learns from visual input. Today, I build and explain GenAI systems that help people learn, reason, and create across engineering, product, and strategy.

If you are looking for someone who can prototype quickly, teach clearly, and connect technical and non-technical teams, let’s talk!

Dilini K. Sumanapala, PhD
Founder, Generative AI Engineer
Genverv Ltd.

Book a Call

AI Agents course for DataCamp

I developed and deployed advanced AI agents to 3000+ DataCamp learners, showcasing how to rapidly prototype and orchestrate chatbot and language-based AI applications using LangChain and LangGraph. I guided users through robust development pipelines in Python, emphasizing practical implementation of both foundational and advanced tools for building multi-step agents capable of handling user queries and executing autonomous tasks.

Large Language Models are a rapidly evolving domain, and I led the development and deployment of a chat application designed to interact with hundreds of patent data documents. To mitigate the risk of “hallucinations,” I implemented Retrieval-Augmented Generation (RAG), ensuring the LLM’s responses were constrained to specific data stored in vector databases on cloud platforms such as Qdrant. This approach enables organisations to guarantee that all outputs from the LLM are reliably grounded in their proprietary datasets.

Chatting with Patent Data using LLMs

Using current NLP tools, it is possible to quantify the similarity between different forms of text, which can easily be translated into recommendation engines. While I developed this pipeline to assess the similarity between different books, it is possible to calculate similarity on any type of content such as scripts, qualitative research, organisational documents, etc.

Book Recommendations using NLP Text Similarity

HuggingFace is a powerful library that includes tools for complex text processing. Here, I use HuggingFace transformers to create summaries of well-known public domain sci-fi novels downloaded from Project Gutenberg. Similar techniques can be deployed for summarising many forms of text documents, increasing the speed of research within organisations for fast paced environments.

Text Summarization using HuggingFace NLP Transformers

Natural Language Processing and Twitter: A Case for Fashion Retail

Using real Twitter data, this project showcases how natural language processing can be used to track emotional sentiment towards a retail brand in a social media landscape.

Incorporating RoBERTa, Facebook’s sophisticated sentiment analysis tool, I quantified positive and negative attitudes towards a clothing brand across 60,000 tweets collected over the course of a year.

Named Entity Recognition Database for European Parliament Proceedings

Using Named Entity Recognition tools featured in the Spacy NLP library, I created a database of important entities (dates, events, persons, organisations, etc.) from 9000 publicly accessible European Parliamentary documents. Similar databases can be created whenever documents with similar semantic content need to be rapidly accessed together to facilitate research within an organisation.

Machine Learning Projects

  • Deep Learning and Predicting Classroom Absenteeism

    Investigating how self-reported time-management ratings can be used to predict how often a student attends college classes.

  • Support Vector Machines and Predicting Gender Using Academic Performance

    Using reading, writing, and maths test scores to determine student gender, which could be used to address emergent gender disparities in learning.

  • K-Nearest Neighbour and Predicting Sports Purchases based on Cardio

    Exploring ways to classify treadmill model purchases using customer demographics.