Benchmarking Statistical, ML, Deep Learning, and Hybrid Models for Global Renewable Energy Consumption
With Paramita Roy — walk-forward CV, structural breaks, and SDG-relevant energy forecasting (open access).
Applied AI research · production analytics engineering
Building systems that make decisions under uncertainty
Building systems that make
Hi, I'm Shaon Biswas — AI & Analytics Engineer. Most of what I do comes down to one problem: making a model's uncertainty explicit enough that someone can act on it.
That runs through two strands of work. Peer-reviewed research on uncertainty-aware machine learning — conformal prediction for trade execution, tail-risk estimation, calibration-aware deep learning. And production analytics engineering at ShelfTrak, where AI-assisted tools, semantic models and custom visuals serve global brands across 70+ airports. MSc (Distinction) in AI & Data Science.
Open to: Research collaboration · Building with businesses · Open-source
Python
SQL
Power BI
Machine Learning
Experience
From business analytics to applied AI — and the products built along the way.
Building applied AI and analytics products since 2024, on six years of business analytics across the UK, Canada, Sweden, and Bangladesh.
Feb 2025 — Present
United Kingdom
Business Intelligence Analyst
Building applied-AI and analytics products for leading global FMCG brands across travel retail — serving 70+ airports and 90+ stores. Engineer end-to-end Power Query (M) pipelines and DAX analytics, and built two AI-assisted production tools (Promo Depth & Historic Exchange Rate calculators, developed with Cursor) that feed ShelfTrak's SQL Server reporting pipeline and replace error-prone manual work.
ShelfTrak Limited · part of the team named New Business of the Year (York Press Business Awards 2025)
Mar 2024 — Jan 2025
Canada · Remote
Data Scientist
Delivered data science and ML solutions across multiple industries — reducing client reporting cycles by approximately 25% and improving operational decision-making through deployed BI dashboards and predictive models.
Digilyzent Inc.
Dec 2019 — Oct 2023
Sweden · Remote
Business Data Analyst
Drove up to 20% quarter-over-quarter growth through data-driven sales analysis, reduced supply chain delays by ~30%, and supported a business pivot that contributed to ~200% revenue increase.
Scandinavian Designs Group West AB
Dec 2021 — Apr 2023
Bangladesh
Technical Lead — Access to Market
Led a team of 10 to deliver digital transformation strategies across 98 SMEs — achieving up to 85% increase in customer reach and 20% revenue growth through e-commerce onboarding, social commerce setup, and digital marketing implementation.
Rezia Management Consulting · Swisscontact project funded by the Embassy of Switzerland
Decision tools
Market-intelligence and stock-analysis apps, automated reporting, decision-support systems.
Operational intelligence
Demand forecasting, segmentation, process automation, real-time BI pipelines.
Applied-AI research
Uncertainty quantification, forecasting benchmarks, model evaluation and deployment.
Retail & travel retail · Finance & FinTech
Research Work
Two published, one under review.
One question runs through all of it: when a model is uncertain, how do you make that uncertainty explicit enough to act on? Conformal prediction is the recurring tool — applied so far to trade execution, tail risk, medical signals and security.
Financial ML & uncertainty quantification
Peer-Reviewed Journal Article · Open Access
Uncertainty-Aware AI: Conformal Prediction vs Reinforcement Learning for Optimal Trade Execution
With Asadullah Irshad (University of Hull, DAIM) — reframes VWAP trade execution as uncertainty-aware control. A normalised split-conformal predictor is used as an explicit decision gate, reaching 90%+ empirical coverage and cutting execution-cost variability from 19.1 to 10.0 bps, while a multi-seed PPO reinforcement-learning agent proves high-variance and unreliable. Validated on 30 US large-cap equities at 5-minute resolution — the first use of conformal intervals as a decision gate in trade execution.
Journal Article · Under Review
Deep Learning for Value-at-Risk and Expected Shortfall Estimation
Tail-risk estimation with deep learning, evaluated on the criteria that matter to a risk function rather than average-case accuracy alone. Under review at the Journal of Risk.
Applied AI beyond finance
Peer-Reviewed Journal Article
Benchmarking Forecasting Models for Global Renewable Energy Consumption
With Paramita Roy — benchmarks statistical, ML, deep learning, and hybrid models (including a novel ETS–GRU hybrid) on World Bank renewable energy data under expanding walk-forward cross-validation, Diebold–Mariano tests, and Model Confidence Set analysis, with structural break considerations for policy-relevant forecasting.
MSc Dissertation
AI & Platelet Proteomics in Cardiovascular Risk of Diabetic Patients
Investigated the molecular mechanisms of cardiovascular complications in diabetic patients by applying AI and bioinformatics to platelet proteomics data — utilising PCA, clustering, GSEA, and pathway enrichment analysis to identify potential cardiovascular risk biomarkers.
Ongoing Work
Other active threads
Further applied work in progress, all sharing the same uncertainty-quantification core:
- Conformal prediction for phishing URL detection — coverage guarantees in adversarial security classification.
- CalGlaucoma — calibration-aware deep learning for glaucoma detection, where a confident wrong answer is the expensive failure.
- Platelet proteomics ML — continued analysis following the MSc dissertation, including negative results confirmed under proper normalisation.
Selected Work
Systems built end to end — from data pipeline to deployed decision tool.
Financial & quantitative systems
Paper → reproducible code
Open source · MIT
Conformal VWAP Execution
The reference implementation behind my published trade-execution paper — conformal prediction used as an execution gate, with the whole study reproducible from a clean checkout in about a minute on a laptop CPU.
- Contains
- A market simulator with stochastic volatility, gradient-boosted return forecasting, split-conformal prediction intervals, classical execution benchmarks (TWAP, VWAP, Almgren–Chriss), and a PPO reinforcement-learning comparator.
- Reproduces
- Walk-forward multi-seed backtests, cost–risk frontier plots, and empirical coverage validation — every figure and table in the paper regenerates from one command.
Open source · Deployed
Breaking News Market Sentiment
A production-grade NLP dashboard aggregating real-time financial news from Bloomberg, CNBC, and Reuters — using VADER sentiment analysis, Fear & Greed Index, VIX, and S&P 500 correlations. Processes 500+ daily news sources with ML data pipelines, containerised with Docker and deployed on Render.
Hosted on a free tier — the demo may take up to 30 seconds to wake.
Production analytics engineering
70+ airports · 90+ stores
Production · ShelfTrak Limited
AI-Assisted Production Tools for Travel-Retail Reporting
Two internal tools — a Promo Depth Calculator and a Historic Exchange Rate Calculator — built with Cursor and shipped into ShelfTrak's live reporting pipeline.
- Problem
- Promotional depth and multi-currency normalisation were calculated by hand for global FMCG brands across travel retail — slow, and error-prone at scale.
- Built
- Two AI-assisted calculators feeding end-to-end Power Query (M) pipelines and DAX analytics into ShelfTrak's SQL Server reporting layer.
- Effect
- Replaced recurring manual work in the production reporting path serving 70+ airports and 90+ stores.
Built on demand
Production · ShelfTrak Limited
Custom Power BI Visuals
A routine part of how I build: when the standard visual library can't express what the analysis needs, I build the visual. Ring charts, fill-tube gauges, score panels, an in-report HTML document viewer — packaged as .pbiviz and shipped into live client reporting rather than left as prototypes.
The point isn't the individual visual — it's not treating the chart library as a constraint on the analysis. Each one carries its own data-view mapping and formatting options, and gets reused across client reports once it exists.
Decks generated from data
Production · ShelfTrak Limited
Programmatic Client Reporting
Client decks generated directly from source data rather than assembled by hand — a 21-slide spirits review, an 18-slide category deck, and category reviews across confectionery, generated from code against a shared brand system.
- Problem
- Recurring client reviews were rebuilt manually each wave — slow, and every rebuild is a chance to introduce an error.
- Built
- A code-generated deck pipeline reading straight from the reporting layer, with brand styling, chart formatting and category logic expressed once and reused.
- Effect
- Reviews regenerate on new data instead of being rebuilt, and the output is consistent across brands and waves.
12+ global brands
Production · ShelfTrak Limited
Enterprise Semantic Models & Data Pipelines
Power BI semantic models with advanced DAX for global spirits, confectionery and tobacco brands across international airport retail — distribution and Perfect Store logic, must-stock-list compliance, promotional depth, market share versus share of space, and multi-currency price tracking down to terminal level.
Alongside the models: the pipelines that feed them, including brand normalisation against master mappings across tens of thousands of rows, plus a reusable theme toolkit standardising 13 client report designs.
Recognition & Media
Recognised in leading global travel-retail trade publications for advancing data-driven analytics in international airport retail — and part of the ShelfTrak team named New Business of the Year at the York Press Business Awards 2025.
The Moodie Davitt Report
The world's leading travel retail business publication.
GTR
Magazine
Global Travel Retail Magazine — industry authority covering 50+ international markets.
DFNI Online
Duty Free News International — global duty free and travel retail industry publication.
Open Source & Writing
Work I do outside the job description.
Open source
Public repositories covering applied NLP, market-sentiment pipelines, and analytics tooling — built to be read and reused, not just shipped.
GitHub →Writing
Technical and industry writing on applied AI, travel-retail analytics, and decision intelligence — on Medium and here.
Insights & Articles
Thoughts on AI, retail, finance, and decision intelligence.
Building a Real-Time Financial Sentiment Intelligence Dashboard
End-to-end sentiment analysis dashboard for financial news using NLP, Python Flask backend, and React frontend.
The Passenger-to-Sale Conversion Crisis in Travel Retail
Record passengers, declining spend. How AI-powered demand forecasting, personalisation, and shelf intelligence can close the conversion gap.
Academic Background
My academic foundation underpins my technical expertise. I hold an MSc in Artificial Intelligence & Data Science (Distinction) from the University of Hull, where I focused on utilizing AI in cardiovascular research.
Additionally, I hold an MSc in International Business Management (Merit) from Sheffield Hallam University and a BBA (Magna Cum Laude), giving me a unique edge in translating complex data into strategic business value.
2024
Applied AI since
12+
Global Brands Servedvia ShelfTrak reporting
2
Published Papersplus further work under review
Technical Skills
Data Science & AI
BI & Data Engineering
Tools & Deployment
Research & Modelling
Certifications & Professional Development
Let's Connect
Research collaboration
Co-authorship and technical discussion on uncertainty-aware ML, forecasting, and decision support. Currently working across quantitative finance, energy, and healthcare AI.
Start a conversationBuilding with businesses
Happy to talk with teams about analytics and applied-AI systems — what's worth building, what it takes to get it into production, and where the numbers can and can't be trusted.
Let's talk