cv
Basics
| Name | Sachith Gunasekara |
| Label | Senior AI Engineer |
| sachithgunasekara@outlook.com | |
| Url | https://sachith-gunasekara.github.io |
| Summary | AI engineer and researcher working on LLM reasoning, instruction tuning, LLM evaluation, multi-agent systems and retrieval-augmented generation. |
Education
Work
-
2026.02 - Present Colombo, Sri Lanka · Hybrid
Senior AI Engineer
Gapstars, placed at Envision AI Labs (Leiden, Netherlands)
- Chose the retrieval embedding model on evidence: built a synthetic question–answer benchmark linked to source records and compared four leading models on recall, nDCG and MRR, with significance tests
- Building a ground-truth benchmark from client-provided documents to measure retrieval and answer quality on real data
- Designed the evaluation framework for RAG and agentic systems using LLM-as-a-judge metrics, claim-level faithfulness checks and expert-validated gold-standard datasets
- Built and benchmarked LLM-based insight-generation features on client-approved pharmaceutical and clinical data
- Contributed to the multi-agent architecture behind narrative trend generation: hybrid retrieval, guardrails, citation grounding and context management
-
2025.09 - Present Online
Instructor, AI Engineering Bootcamp
StemLink
- Teach an applied LLM curriculum: transformers, prompting, agents (LangChain, LangGraph, AutoGen), RAG (LlamaIndex, Pinecone) and LoRA/QLoRA fine-tuning
- Supervise capstone builds through to working demos
-
2025.06 - 2026.02 Toronto, Canada · Remote
Senior AI Engineer
Arcadea Group
- Automated structured data extraction from a US state police agency's records with an offline multi-agent system of 10+ planning, generation, execution and evaluation agents that write named-group regex patterns
- Made accuracy verifiable with a generate–test–revise loop scored against ground truth, delivering a validated, reusable regex instead of an ongoing model dependency
- Built a customer-care assistant grounded in company knowledge and historical phone calls, comparing traditional RAG with graph-based RAG and adding topic modelling and source traceability
-
2025.04 - 2026.07 Colombo, Sri Lanka
Machine Learning Engineer
Surge Global
Full-time, then part-time contract
- Led AI delivery of an orchestrated multi-agent system for an insurance warranty-claims intermediary, giving service technicians a ready recommendation for each claim
- Researched connecting a real-time GPT-4o voice agent to multi-agent workflows for a founder relationship assistant
-
2023.10 - 2025.04 Colombo, Sri Lanka
Associate Machine Learning Engineer
Surge Global
- First author of iSelf-Discover, showing unstructured reasoning plans beat structured JSON plans across BBH, T4D and MATH (up to 18.9% relative on MATH)
- Co-developed OpenBezoar, a low-cost recipe for small open instruction-following models (synthetic data, QLoRA SFT, DPO)
- Wrote nanoJAXGPT, a JAX/Equinox rewrite of nanoGPT, as a teaching resource
-
2023.03 - 2023.10 Colombo, Sri Lanka
-
2018.11 - 2019.02 Battaramulla, Sri Lanka
Volunteer
-
2024.10 - Present Colombo, Sri Lanka
Judge
IEEE Young Professionals - Sri Lanka
Judge at Sri Lanka Artificial Intelligence Competition (SLAIC) 2024
-
2023.12 - Present Colombo, Sri Lanka
-
2017.11 - 2018.08 Colombo 7, Sri Lanka
Secretary and IT Assistant
International Vipassana Meditation Center
Secretary and IT Assistant to Chief Incumbent Thero, Ven. Udadumbara Kashyapa Thero
-
Awards
- 2025.08
Youth Delegate, Global Leadership Program
Headway Institute of Strategic Alliance, Brasenose College, University of Oxford
Partially funded fellowship on global policy dialogue and strategic innovation
- 2024.03
Gold Medal for Best Performance in Computer Science
University of Kelaniya, 142nd General Convocation
- 2023.10
Best Presenter
4th International Conference on Applied and Pure Sciences (ICAPS 2023), University of Kelaniya
- 2022.11
-
Skills
| Research areas | |
| LLM reasoning | |
| Instruction tuning | |
| LLM evaluation | |
| Multi-agent systems | |
| Retrieval-augmented generation |
| Methods | |
| Supervised fine-tuning | |
| LoRA/QLoRA | |
| DPO | |
| Synthetic data generation | |
| LLM-as-a-judge | |
| Calibration |
| Tools | |
| Python | |
| JAX/Equinox | |
| Hugging Face | |
| LangChain | |
| LangGraph | |
| AutoGen | |
| LlamaIndex | |
| Pinecone |
Talks
- 2025.04
A Step Beyond the Hype: A Journey Through Modern AI
IEEE Student Branch, Wayamba University of Sri Lanka
Invited talk (online)
- 2024.08