Nadia Kowalski
Co-founder and technical lead. Former principal data scientist at a FTSE 250 insurer. Specialises in time-series forecasting and anomaly detection. Holds a PhD in applied statistics from Imperial College London.
How we started
Pillars of AI began as a two-person experiment in a shared office in Keelingbury. Our founders, Nadia Kowalski and James Hartley, had spent the previous decade working inside large consultancies. They kept seeing the same pattern: companies would pay six-figure sums for an AI proof of concept, watch a dazzling slide deck, then never see the model reach production. Data sat in silos, engineering handoffs failed, and the proof of concept quietly died.
They decided to build a firm that only takes on projects it can ship. That principle still governs every engagement we accept. If we believe the data quality is too poor, the budget too thin, or the timeline unrealistic, we say so during the scoping call rather than three months in.
By mid-2021 we had delivered nine production systems and hired our first two engineers. Today the team stands at eleven people, all based in England, all working on no more than two projects at a time so nobody is context-switching across five clients.
Milestones
2019 — Founded
Nadia and James registered Pillars of AI Ltd and signed the first client, a regional logistics company that needed route-optimisation predictions. The model cut average delivery times by 11 per cent in its first quarter.
2020 — First NLP deployment
We built a ticket-classification engine for a financial-services firm processing over 4,000 support emails a week. The system categorised incoming requests into 23 queues with 94 per cent accuracy, freeing up the equivalent of three full-time agents.
2021 — Team growth
Hired four ML engineers and a dedicated data-ops specialist. Moved into a permanent office at 541 Uriah Paddock. Started offering ongoing model-monitoring retainers alongside project-based work.
2023 — Computer vision practice
Launched a dedicated CV practice after completing three consecutive manufacturing quality-inspection projects. Invested in an on-premise GPU cluster for clients who cannot send image data to public cloud providers.
2025 — Present
Eleven people, 47 completed projects, 12 sectors. We are currently expanding into healthcare data pipelines and exploring federated-learning architectures for clients with strict data-residency requirements.
What we care about
Four commitments that shape how we run projects and treat clients.
Honesty at the scoping stage
If your problem is better solved with a spreadsheet formula or a simple rule engine, we will tell you. We have turned down roughly one in five inbound leads because the project did not genuinely need machine learning. That saves both sides time and money.
Production first
Every deliverable includes deployment scripts, monitoring hooks and documentation. A model that only runs in a Jupyter notebook is a research artefact, not a business tool. We write code that your internal team can maintain after the engagement ends.
Realistic timelines
We quote four-week prototype cycles because that is what the work actually takes when you account for data cleaning, stakeholder reviews and iteration. Promising a working model in five days is a recipe for corners cut and trust lost.
Knowledge transfer
We run a two-day hands-on workshop at the end of every project so your developers understand the model architecture, the retraining pipeline and the alert thresholds. The goal is to make ourselves unnecessary as quickly as possible.
Key people
A few of the engineers and researchers who lead our client work.
James Hartley
Co-founder and client lead. Background in software engineering at two London fintech startups. Manages project scoping, contracts and the operational side of every engagement. Keeps engineers focused and clients informed.
Ravi Anand
Senior ML engineer. Joined in 2021 after five years at a computer-vision startup in Cambridge. Leads our manufacturing inspection projects and maintains the on-premise GPU cluster. Contributor to two open-source object-detection libraries.