Turning complex data into business value
DevStat is a data science and AI consultancy specialising in enterprise machine learning, applied AI solutions, and statistical modelling.
Get in touch Our servicesServices
What we do
End-to-end data science solutions, from strategy through to production deployment.
Applied AI & Automation
End-to-end AI systems that automate complex business processes — from unstructured data extraction to classification and validation. Production-proven across 20+ use cases.
Predictive Analytics & ML
Forecasting, classification, anomaly detection, and optimisation models deployed as production APIs. Deep experience in time series, Bayesian networks, and ensemble methods.
Cloud & MLOps
AWS-native architectures with automated training, testing, and multi-environment deployment pipelines. Infrastructure-as-code with Terraform and CI/CD via GitHub Actions.
Generative AI Integration
LLM-powered solutions using Claude, GPT, and Amazon Bedrock. Few-shot prompting, retrieval-augmented generation, and multimodal AI pipelines for enterprise use cases.
Statistical Consulting
Rigorous statistical analysis for research, regulatory, and business applications. Specialising in credit risk modelling, clinical trials, spatial statistics, and causal inference.
Training & Mentorship
Upskilling teams in data science, machine learning, and cloud infrastructure. Curriculum development with over a decade of postgraduate teaching experience.
About
Built on deep expertise
DevStat was founded by Dr Theodor Loots, a data scientist and mathematical statistician with over 15 years of experience across academia, financial services, and the automotive industry.
As technical lead of a data science competency at a global automotive OEM, Theodor has designed and deployed production AI systems at scale, built MLOps platforms on AWS, and led multi-year optimisation programmes integrating real-time enterprise data.
Prior to industry, he served as a senior lecturer at the University of Pretoria, supervising 30+ postgraduate students and publishing 30+ peer-reviewed papers. He holds a PhD in Mathematical Statistics and has been recognised by TWAS, the Heidelberg Laureate Forum, and the Japanese STS Forum.
Domain Experience
Industries we've worked in
Cross-sector experience means we understand your data challenges, not just the algorithms.
Software we have shipped
Methods are worth what their implementations are worth. These are on public registries, under test, and in use — the same engines behind the consulting work.
Induction Period Reader live
Reads the oxidative induction period off a Rancimat conductivity trace in the browser. Fits a four-parameter kappa response, returns both readings admitted by the standard with bootstrap intervals, and says whether the two can be reconciled for that oil. Nothing is uploaded; the fit runs on your machine.
arcstat R & Python
The shared C engine behind our distributional work: four-parameter kappa fitting, arc-length methods, L-moment machinery, and exact inference. One back end, two fronts, tested for parity between them.
Distribution and filtering packages R
quatstat for quaternion-valued data, lulustat for LULU smoothers and nonlinear filtering, regstat for regime and change detection, vslstat for variable-scale likelihood work.
Where the methods come from
Consulting that reaches past the standard toolbox needs someone who builds the tools. These are the areas we work in, and the kind of problem each one answers.
Accelerated stability and shelf-life claims
How long does a product last, and what can you defend saying about it? We model the whole measured curve rather than a single threshold crossing, which gives an induction period with an honest interval attached, and shows when an extrapolation from accelerated conditions to ambient has stopped meaning anything.
Distribution theory for awkward data
Real measurements are bounded, skewed, heavy-tailed or censored, and the textbook families fit none of them well. We work in flexible families that contain the standard ones as special cases, so a model can be chosen on evidence rather than convenience.
Exact inference when the sample is small
Asymptotic intervals quietly fail at the sample sizes most projects actually have. We build exact and permutation-based procedures whose stated coverage is the coverage you get, which matters when a number has to survive a regulator or a referee.
Signal, image and spatial structure
Nonlinear filtering, edge-preserving smoothers and level-set methods for data where the interesting part is a boundary or a shape rather than an average.
Risk, credit and finance
Credit-risk modelling and validation, low-default portfolios, and the measurement problems that appear when the events you care about are rare by design.
Machine learning with statistics underneath
Neural and shrinkage methods treated as estimators: what they converge to, how uncertain they are, and when a simpler model would have done the same job more defensibly.
Accessibility apps
Built alongside the consulting work, for iPhone, and designed first for people who use a screen reader.
ScoreNavigator in development
Sheet music made navigable without sight: move through a score by bar, voice and section, and read it in braille music notation.
CurbNavigator in development
Pedestrian navigation that pays attention to the things that actually decide a route on foot — crossings, kerbs and obstacles.
Peer-reviewed publications
Twenty-four papers, 2013 to 2025, in statistics, materials science, image analysis and the life sciences, each entry linked to its DOI. See the full list.
Contact
Let's talk about your data
Whether you need a production AI system or a one-off statistical analysis, we'd love to hear from you.