EP6: From Rust to Meth Labs: Putting Science at the Heart with Jim Carrol, Digica

From detecting rust on painted surfaces to identifying meth labs for law enforcement, Digica is tackling some of the most unusual and challenging problems in applied AI — often where usable data simply doesn't exist. In a recent episode of the UK AI podcast, Tim spoke with Jim Carroll about Digica's data-science-led approach to building AI solutions, and how synthetic data is helping solve some of the hardest data scarcity problems in computer vision.

An AI Software Company Built Around Data Science

Unlike many AI software companies, Digica's technical team is made up primarily of data scientists rather than software developers. This allows the company to start from a dataset, an academic paper, or even just a concept, and build AI solutions from the ground up — with software engineering discipline applied later to deploy those solutions into real-world applications.

Jim notes that most of Digica's clients are relatively well-educated about AI and data science — including finance sector clients seeking to replace traditional financial models with AI-driven alternatives — though some still need guidance on the basics, particularly around the importance of a strong starting dataset.

Tackling Legacy Data and Fragmented Systems

A common challenge across industries is consolidating data scattered across legacy systems, different databases, and platforms — a problem that becomes even more pressing with AI. Digica still relies on custom pre-processing pipelines, but increasingly uses AI-powered tools, including language models, to automate parts of that standardisation process. Even so, Jim is clear that off-the-shelf AI tools don't solve every problem — bespoke development, guided by experienced data scientists, remains essential.

Why Data Science Is Still "The Wild West"

Jim draws a striking comparison between the maturity of software engineering — a discipline with decades of established best practice supporting everything from banking systems to spacecraft — and data science, which he describes as still being in its "foothills" by comparison. While the underlying mathematics and statistical models are well established, the methodologies for applying data science reliably within business processes are still catching up. Digica positions itself at the forefront of what Jim calls "practical AI" — solutions built for real-world deployment, not just the lab.

Solving Data Scarcity with Synthetic Data

Synthetic data is one of Digica's core specialisms, particularly within computer vision, where off-the-shelf models often fall short of solving highly specific, real-world problems. Jim shares two striking examples:

• Detecting rust for an industrial paint manufacturer — Since rust develops slowly over days or weeks, gathering enough real-world images is impractical. Digica generated synthetic images of rust formation, combined with limited real photographs, to train an effective detection model — accounting for the fact that rust, while it looks two-dimensional, is actually a three-dimensional phenomenon.
• Detecting meth labs for law enforcement — With no publicly available photographs of real meth labs to train from, Digica built a synthetic data pipeline using generic lab equipment imagery and reference material (including media depictions) to generate realistic synthetic training data, combined with limited real photographs, to build an effective detection model.

Digica has also applied similar synthetic data techniques to numerical datasets — for example, predicting failures in wireless routers — demonstrating that the underlying approach extends well beyond computer vision.

How Synthetic Data Is Actually Built

Jim explains that the process typically starts with real photographs or 3D CAD models, from which images are rendered with randomised object placement, lighting, and texture to reflect real-world variability. Generative AI can then add an additional layer of realism on top. In cases where CAD models aren't available — such as medical instruments — Diga has used 3D scanning technology, supplemented by design teams manually filling in gaps the scanner missed.

From Experimentation to Business-as-Usual AI

Digica's clients often begin their AI journey in an experimental mindset, and Jim notes that managing expectations around AI's non-deterministic nature — the fact that 100% accuracy is rarely achievable — is a key part of client education, especially for organisations newer to AI. For clients already experienced with AI deployment, this is a familiar and accepted part of the process.

Crucially, Digica doesn't treat AI projects as one-off engagements. The company continues working with clients well beyond initial deployment — refining accuracy, maintaining systems, and supporting new users — treating AI adoption as an ongoing journey rather than a single delivery milestone.

Measuring ROI: Moving Beyond Technical Metrics

One of the more practical challenges Jim discusses is how to demonstrate the business value of AI to non-technical stakeholders. While technical metrics like precision, recall, and F1 scores are meaningful to data scientists, they mean little to business decision-makers focused on return on investment. Digica now focuses on translating AI performance into tangible business outcomes — for example, comparing the cost of an AI-powered inventory-counting system against the labour cost of manual stock checks, to clearly demonstrate value.