M.Tech Data Science & AI @ PDEU (GATE 2025) · Intern @ BISAG-N · Q1 Elsevier co-author
I build ML systems end-to-end — pipelines to deployment — and lately I've been deep into multi-agent LLM systems. At BISAG-N I'm building flood prediction infrastructure for a government agency: ERA5 ingestion pipelines, HEC-HMS/HEC-RAS hydraulic simulations validated against a real 2017 flood event. Separately, I built an AI Council where 5 LLMs debate a problem across rounds and vote on the answer. Also Mimir, an LLM interface backed by ChromaDB so it actually remembers between sessions.
I want to work somewhere agents and production ML infrastructure are both taken seriously.
- Kelani Basin Digital Twin — AHP-enabled Agentic Geo-AI flood prediction engine for Sri Lanka (BISAG-N)
- AI Council — hierarchical multi-agent system with LLM debate + majority-vote consensus
- Mimir CLI — memory-augmented LLM interface using ChromaDB
ML & AI
MLOps & Infrastructure
Data Engineering
Geospatial
| Project | What it does | Stack |
|---|---|---|
| Bank Fraud Detection | Real-time fraud scoring on 1M transactions. AUC 0.89, KS 0.63 | Spark, Kafka, GBT, HDFS |
| AI Council | 5 LLMs debate problems and vote on answers. Hierarchical agent architecture | Python, HuggingFace, LLMs |
| Mimir CLI | LLM interface with persistent memory across sessions | Python, ChromaDB, Embeddings |
| Image Super-Resolution | Denoising + 2x upscaling. PSNR 38.48 dB, SSIM 0.99. Full MLOps stack | PyTorch, Docker, Prometheus |
| Fake News Detection | Multilingual classifier. Fine-tuned XLM-RoBERTa + LightGBM ensemble | HuggingFace, SBERT, FastAPI |
| VITON-HD | Virtual try-on with SegFormer parsing and MediaPipe pose estimation | PyTorch, SegFormer |
Deep Learning for Canine Cardiac Radiography: A Comprehensive Review of Automated Vertebral Heart Score Estimation The Veterinary Journal, Elsevier (Q1) · Vol. 319 · 2026 Het Buch, Prit Mayani DOI: 10.1016/j.tvjl.2026.106797