I'm an AI Engineer with an MSc in Data Science (University of Hertfordshire, UK), specialising in building production-grade agentic AI systems - from multi-agent LangGraph pipelines to full-stack LLM applications.

• 🏗️ Currently building: multi-agent orchestration systems, RAG pipelines, LLM-powered applications, and MLOps infrastructure with LangGraph, FastAPI, MongoDB, and Docker
• 🔍 Focus areas: Agentic AI · Multi-Agent Systems · RAG & Semantic Caching · LLM Evaluation · MLOps
• 🌱 Actively learning: distributed training, LLM evaluation frameworks, agent memory architectures, vector-DB internals
• 🟢 Open to: AI Engineer · ML Engineer · LLM/Agentic Systems Engineer roles in the UK

Languages

Agentic AI & LLMs

ML / DL / Data

MLOps & Infrastructure

Cloud & Tools

The work I'm proudest of - production-grade systems spanning agentic AI, full-stack engineering, and applied ML.

A multi-service AI system: 9 LangGraph agents, ML drift monitoring (MLflow), Airflow ETL pipelines, GDPR-compliant CV parsing, and a TypeScript/React frontend - built end-to-end as a production platform. 🌐 marketforge.digital

Procurement and supply chain operations for mid-market companies - an EU-based engagement point in Sweden paired with senior-led delivery from Sri Lanka. Three decoupled services: a live site with a real-time disruption monitor, an intelligence pipeline that turns the raw GDELT global-events feed into classified, scored disruptions with AIS ship-tracking, and a LangGraph deep-agent "war room" that investigates a single incident end-to-end into a cited impact briefing.

MultimodalAI'26 Hackathon — Clinical Strand · London, June 2026 · 🏆 Winner. Built a complete AI governance pipeline for predicting 24-hour ICU deterioration across 5,000 post-surgical patients. Three competing models: a PyKale multimodal fusion network (tabular + TF-IDF NLP + hourly vitals time series), LightGBM tabular baseline, and Logistic Regression interpretable anchor. Full evaluation suite: AUROC, ECE calibration, Decision Curve Analysis, SHAP explainability, MNAR-aware subgroup equity analysis. Two production Streamlit UIs — a 6-tab clinical Evidence Dashboard (AUROC 0.85–0.89, ECE 0.044) and a dark-theme ICU Nursing Monitor (bedside IoT screen with LangChain/Gemini 2.0 Flash clinical agent, ElevenLabs voice AI, and patient priority board for 1,000 patients). Key clinical insight: note absence is MNAR — concentrated in the highest-acuity patients — and must be modelled explicitly, not imputed away. 🔗 UKOMAIN26_Hackathon_Clinical

MongoDB Agentic Evolution Hackathon - London, May 2026. Fake signals move real markets. Phantom Trade is a dual-pipeline autonomous agent system that detects fabricated supply-chain headlines before they reach risk models - combining a 5-scorer ML forensics engine (TF-IDF, spread velocity, linguistic anomaly, source credibility, template matching), multi-source news aggregation (GDELT, NewsAPI, RSS, X API v2, Reddit, Wayback CDX), and adversarial LLM debate (MAD-Sherlock: PRO-AUTHENTIC vs PRO-FABRICATED agents). A LangGraph Oracle (PLAN→ACT→OBSERVE→REACT) then generates bi-temporal commodity risk theses for 4 materials, with a 3-layer MongoDB memory system (short-term TTL, long-term Voyage AI vectors, reasoning bank) enabling agents to self-improve across runs. 🔗 Phantom-Trade

Discovers, evaluates, and prepares personalised applications for AI Engineering, ML Engineering, and Data Science roles across the UK - autonomously, end to end. 🔗 JobFinder

Match outcomes, score forecasts, and squad optimisation using ensemble ML, LSTM, and genetic algorithms — ~72% match-outcome accuracy. 🔗 CricOracle2026

Multi-agent pipeline that gathers news and research papers, evaluates relevance, synthesises analysis, and auto-generates email digests or LinkedIn posts with AI-generated images. 🔗 AI-News-Analyzer

I care less about a +0.2% F1 score in a notebook and more about whether the system is observable, reproducible, and recoverable in production.

That means:

• Evaluation-first — agents get evals before they get features
• Memory-aware — systems that improve across runs, not just across epochs
• Production-honest — latency budgets, drift monitoring, and human-in-the-loop gates built in from the start, not bolted on after

If you're hiring for a role where the AI system has to actually run, that's where I do my best work.

I'm actively looking for AI Engineer / ML Engineer / LLM Systems Engineer opportunities in the UK, and I'm always happy to chat about agentic AI, LLM evaluation, or multi-agent architecture.

• 💼 LinkedIn: linkedin.com/in/viraj97
• 📬 Email: amanthavirajavb@gmail.com
• 🌐 Portfolio: viraj-bulugahapitiya.vercel.app

_{⭐ If any of these projects help you, a star is hugely appreciated — it helps others discover the work too.}