SLG

Savinu Linath Gunarathna

AI Research Engineer

https://www.opentalent.in/savinu-linath-gunarathna

AI Research Engineer with 2+ years in NLP, LLM & RAG Systems

Key Strengths

Demonstrated ability to build and deploy end-to-end NLP/ML systems, including custom corpus creation and PyPI package publication.
Strong practical experience with LLMs, RAG, fine-tuning (LoRA/PEFT), and various transformer architectures (T5, mT5, Gemma 3).
Proficient in lightweight deployment strategies, including INT8 quantization and CPU-only hosting with CTranslate2.
Experience with hybrid retrieval pipelines (dense vector search + BM25) and relevant backend technologies (FastAPI, Qdrant).
Solid understanding of ML evaluation metrics, class imbalance handling (SMOTE, oversampling), and model interpretability.

Cultural & Operational Fit

Cultural Fit Analysis

The candidate's projects demonstrate a strong alignment with the target role of AI Research Engineer, focusing heavily on NLP, LLMs, and RAG systems. The diversity of projects, from transliteration to legal RAG and churn prediction, showcases a broad interest and ability to apply AI/ML across different domains. The independent nature of these projects, coupled with open-source contributions, indicates a self-starter mentality and a passion for learning and building, which are positive indicators for cultural fit in an innovative environment.

Soft Skills & Operational Fit

The candidate's project descriptions indicate a proactive and independent approach to problem-solving, evidenced by building systems from scratch and addressing real-world challenges like code-mixed language. The focus on interpretability (Decision Tree rules, source citation in RAG) suggests an understanding of stakeholder needs and responsible AI practices. The publication of a PyPI package and open-source contributions imply a collaborative and sharing mindset, which aligns well with operational fit in a technical team.

AI is analyzing your overall score…

Identifying your key strengths…

Evaluating your skill match against the job requirements…

Assessing your cultural and operational fit

About

Second-year AI & Data Science undergraduate who has independently built and shipped three full NLP/ML systems alongside coursework, including a self-built 40,000-pair Sinhala transliteration corpus, a published open-source PyPI package, and a hybrid retrieval legal assistant covering 100+ Sri Lankan legal documents. Comfortable across the full pipeline: data collection, model fine-tuning (T5, mT5, Gemma 3 with LoRA/PEFT), evaluation, and lightweight deployment. Seeking an internship to apply this foundation in a production ML/AI team.

Top Skills

Scikit LearnPrompt EngineeringRagLlms

Projects

Customer Churn Prediction: Neural Network vs. Decision Tree Study

January 1, 2025 – June 1, 2026

Built and benchmarked a custom ANN against a Decision Tree on structured customer data, with a full evaluation suite covering confusion matrix, precision, recall, F1-score, and ROC-AUC; addressed class imbalance using random oversampling and SMOTE so both models learned from minority churn cases. Extracted human-readable decision rules from the tree model to surface the strongest churn predictors, giving a non-technical stakeholder a clear basis for prioritising retention efforts alongside the neural network's performance metrics. Compared model performance across the full evaluation suite to recommend which model fits which use case, balancing the decision tree’s interpretability against the ANN’s stronger raw predictive accuracy.

View Project

Dual-Architecture Singlish-to-Sinhala Transliteration System & sin-transliterator PyPI Package

January 1, 2025 – June 1, 2026

Built a 40,000-pair Sinhala transliteration corpus from scratch (no pre-existing labelled data) and fine-tuned multiple transformer architectures (T5, mT5, Gemma 3); Gemma 3 with LoRA/PEFT achieved CER 0.16 and WER 0.31 on a held-out 2,549-row test set, outperforming widely-used existing tools on ad-hoc and code-mixed Singlish that they typically fail to handle. Designed a two-stage inference pipeline (lightweight seq2seq model for real-time use, LLM refinement layer for ambiguous/code-mixed input), quantised to INT8 via CTranslate2 for CPU-only hosting, and published it as the sin-transliterator PyPI package with automatic CPU/GPU detection and versioned Hugging Face weights. Built a custom stochastic data augmenter to generate realistic ad-hoc Singlish variations and scraped code-mixed training examples from YouTube Live Chat, expanding the corpus beyond formal text and directly improving the model’s robustness on real-world informal input.

View Project

MyLawLLM: Sri Lankan Legal RAG Assistant

January 1, 2025 – June 1, 2026

Built a hybrid retrieval pipeline combining dense vector search with BM25 keyword matching over 100+ Sri Lankan legal documents, pairing a plain-English explanation with the underlying legal basis so non-experts can get oriented on routine legal questions without consulting a lawyer first. Implemented end-to-end with a FastAPI backend, Qdrant Cloud for vector storage, and a lightweight web interface; BM25 re-ranking on top of pre-indexed vectors keeps response latency low. Designed the retrieval and prompting layer to cite the specific source document for every answer, so users can trace any explanation back to the original legal text rather than relying on an unverifiable summary.

View Project

Certifications

Supervised Machine Learning: Regression & Classification

ULSA

January 1, 2025 – Present

Fine-Tuning Large Language Models

Hugging Face

January 1, 2025 – Present

Professional Certificate in Machine Learning

IIT PDU

January 1, 2025 – Present

Key Strengths

Demonstrated ability to build and deploy end-to-end NLP/ML systems, including custom corpus creation and PyPI package publication.
Strong practical experience with LLMs, RAG, fine-tuning (LoRA/PEFT), and various transformer architectures (T5, mT5, Gemma 3).
Proficient in lightweight deployment strategies, including INT8 quantization and CPU-only hosting with CTranslate2.
Experience with hybrid retrieval pipelines (dense vector search + BM25) and relevant backend technologies (FastAPI, Qdrant).
Solid understanding of ML evaluation metrics, class imbalance handling (SMOTE, oversampling), and model interpretability.

Cultural & Operational Fit

Cultural Fit Analysis

Soft Skills & Operational Fit

Savinu Linath Gunarathna

Key Strengths

Cultural & Operational Fit

About

Top Skills

Skills

Education

Projects

Certifications

Key Strengths

Cultural & Operational Fit