Janith Chathuranga Mahanama
Computer Science & Engineering, University of Moratuwa
About
I am a third-year Computer Science and Engineering honors student at the University of Moratuwa, specializing in Data Science and Engineering. Driven by a profound interest in academic research and intelligent system design, I have maintained a rigorous academic record and am a consistent high-achiever on the Dean's List. My practical experience bridges the gap between theoretical research and scalable enterprise solutions by integrating machine learning architectures and explainable artificial intelligence with full-stack development and DevOps practices. Whether I am building robust digital platforms, advancing transparent AI methodologies, or navigating high-stakes global programming competitions, I am passionate about engineering end-to-end solutions for complex real-world challenges.
View CVEducation
University of Moratuwa
2024 – PresentB.Sc Engineering (Hons), Computer Science and Engineering
Moratuwa, Sri Lanka- CGPA: 3.95 / 4.00
- Stream: Data Science and Engineering
- Dean's List (SGPA)
- Semester 1 - 4.00
- Semester 2 - 4.00
- Semester 3 - 3.92
- Semester 4 - 3.91
Trinity College Kandy
2009 – 2023Primary and Secondary Education
Kandy, Sri Lanka- GCE Advanced Level (Physical Science) 2022/23, Z-Score 2.6570 · District Rank 10 · Island Rank 106
- GCE Ordinary Level 2019, 8 A's, 1 B
Research Experience
Predicting Post-Crisis Daily Tourist Arrivals
Accepted, MERCon 2026- Sri Lanka's tourism sector has had to rebuild forecasting trust after repeated shocks (the Easter Sunday attacks, COVID-19); this work asks whether real-time search behavior can shorten that recovery gap. Built a multi-model framework (SARIMA, SARIMAX, XGBoost, Random Forest, LSTM, SVR, and a sequential hybrid) fusing official daily arrivals with Yandex search queries, Google Trends interest, exchange rates, and weather data.
- Benchmarked all seven architectures with dedicated feature engineering and explainability analysis to see which signals actually drove predictions, not just which model scored best. The RBF-SVM configuration led, reaching a rolling MAPE of 8.94%.
Feature Selection Research (NIPS 2003)
Not Published- Ran a controlled comparison of filter (ANOVA), wrapper (recursive feature elimination), embedded (Random Forest / XGBoost importance), and hybrid selection pipelines on Madelon, a dataset purpose-built for the NIPS 2003 challenge to punish methods that can't separate 5 informative features from 15 redundant and 480 pure-noise ones.
- Held the downstream evaluator (RBF-SVM) constant across all four families for a fair comparison. Selection quality followed a clear hierarchy, hybrid > embedded > wrapper > filter > no selection, with the best hybrid pipeline lifting balanced accuracy from 0.58 (all 500 features) to 0.89 and reaching 0.97 AUC-ROC.
Projects
AudioLens, Explainable AI Research Workbench for Speech Models
Ongoing- Tech: Next.js, FastAPI, Celery, Redis, Docker, PyTorch, Hugging Face Transformers.
- An interpretability workbench for speech models (Whisper, Wav2Vec2) that fills a gap general tools like Google's Learning Interpretability Tool leave for audio: saliency and attention-based attribution, layer-wise representation probing, Jacobian-based sensitivity analysis, perturbation-driven robustness testing, Whisper hallucination detection, and accent/language fairness analysis.
- Researchers upload their own models and datasets and trace a bad prediction back to the exact audio evidence or acoustic condition behind it. Heavy analysis jobs run asynchronously through a Celery/Redis queue behind a FastAPI control plane, decoupled from the Next.js frontend so long-running inference never blocks the UI, with multi-GPU support (CUDA, ROCm, Metal) across whatever hardware a researcher has.
Physics-Informed GNNs for Multi-Horizon Disease Forecasting
Ongoing- Tech: PyTorch Geometric, Spatio-Temporal GNNs (GCN/GAT), GANs.
- Building a spatio-temporal Graph Neural Network baseline to forecast weekly, district-level dengue incidence across Sri Lanka.
- Enhancing the baseline model by integrating an epidemic-mechanistic (SEIR-SEI) physics-informed loss function, conditional GAN-based data augmentation to counter scarcity, and a learned adaptive graph to capture unobserved inter-district transmission dynamics.
Smart Waste Management System
2026- Tech: GitOps via ArgoCD, Kafka/EMQX event streaming, Kong API gateway, Keycloak/Vault identity & secrets, Prometheus/Grafana monitoring, Hyperledger Fabric ledger for tamper-proof records.
- An IoT-to-insight pipeline for municipal waste collection: bin-level sensors feed an edge layer for local processing, sensor data streams through Kafka/EMQX into a GitOps-deployed platform, and a Hyperledger Fabric ledger keeps collection records tamper-proof for audit, across four repos spanning edge, IoT firmware, mobile, and a routing simulator.
- Worked in the DevOps and Platform Infrastructure team, standing up the deployment and monitoring backbone (ArgoCD, Kong, Keycloak/Vault, Prometheus/Grafana) the rest of the system runs on.
Pharma Availability Control Tower
2026- Tech: FastAPI, Next.js, SQLAlchemy, MySQL, XGBoost, OR-Tools, LangGraph, LangSmith, Chroma, Gemini, Sentence-Transformers.
- Planners previously tracked pharmaceutical distribution across spreadsheets, juggling DC replenishment speed, simultaneous vessel manifests, cold-chain constraints, lorry capacity, and day-level availability by hand. This platform folds shortage forecasting, shipment prioritization, dispatch planning, and execution tracking into one console.
- The core design choice: planning state never directly mutates physical stock. Approvals create scoped reservations and in-transit transfers, and real inventory only changes through explicit business events, keeping the system auditable end-to-end.
- Implemented the Machine Learning module (priority-scoring manifest clearance, XGBoost-based 48-hour shortage forecasting, OR-Tools dispatch optimization) and a multi-agent RAG chatbot (LangGraph supervisor routing to knowledge, operations, analytics, and human-in-the-loop recommendation agents over a Chroma vector store, fully traced with LangSmith), with a CI-gated LLM-as-judge evaluation harness and a self-correction node verifying answer groundedness before response.
RPAL Interpreter
2026- Tech: Java.
- A from-scratch interpreter for RPAL, a lazy functional language, built without lex/yacc or any parser generator: a hand-written lexer, a recursive-descent parser producing an AST, a standardizer that rewrites language constructs down to lambda-calculus primitives, and a Control/Stack/Environment machine that evaluates the result.
- Recursion is call-by-value through a Y*-style fixed-point combinator that wraps functions in eta closures so they can reference themselves. The runtime handles closures, tuples, and 14 built-in operators end-to-end, verified against 17 test programs from Fibonacci to simultaneous bindings.
Banking Management System
2025- Tech: Next.js, Express.js, Neon Postgres, Docker.
- A digital core-banking platform modeled on a Sri Lankan microfinance bank serving rural communities: customer onboarding and agent assignment across branches, tiered-interest savings accounts, fixed deposits across 6-month, 1-year, and 3-year terms, joint accounts, and monthly interest auto-crediting with full transaction audit trails.
- Enforces real banking constraints rather than happy-path CRUD: minimum balances, no overdrafts, one fixed deposit per savings account, and role-based access separating customers, agents, and branch managers. Contributed to the full-stack build and a CI/CD pipeline automating build verification and unit testing.
12-bit Nanoprocessor
2025- Tech: VHDL, Xilinx Vivado, FPGA.
- Designed a processor from the gate level up in VHDL, an ALU, register bank, and instruction decoder, iterated across three passes: a basic 4-bit build, an optimized revision, then an extended 12-bit version, rather than a single one-shot design.
- Verified functionality through simulation and synthesis in Xilinx Vivado, targeting integration into larger FPGA projects.
Competitions
AITHON 2026 by Hemas Holdings PLC
2026- Top 10 Finalist.
IEEEXtreme 19.0 Programming Competition
2025- Global Rank 51.
Coderally 6.0, Hackathon by IEEE Computer Society of IIT
2025- 4th Place (Advanced Tier).
CodeArena, Coding Competition by IET on Campus KDU
2025- 4th Place (Finalist).
Devthon 2.0, Web App Development Competition
2025- Semi-Finalist with "Safe Steps: Women Safety Website."
Certifications
- AWS Academy Graduate, Cloud Foundations
- AWS Academy Graduate, Microservices and CI/CD Pipeline Builder
- Machine Learning Specialization, Coursera (Ongoing)
- Python Programming Online Course, University of Moratuwa
- Web Development Online Course, University of Moratuwa
- Level 3 Certificate in Information Technology, School of Computing, Esoft Metro Campus
Experience & Extracurriculars
University of Moratuwa
2024 – Present- Member of IEEE Student Branch and IESL.
- Member, Strategic Events Committee, Computer Science & Engineering Student Society.
- Member, Publicity Committee for SLIoT 2026.
Trinity College Kandy
2009 – 2023- Teaching Assistant, Advanced Level Section (2023).
- Committee Member, Social Service Union (2019), Science Society (2015, 2017).
- Member, Astronomical Society, Automotive Club, Design Club.
- School Hockey Team (2013–2014, 2016–2017).