Applied AI R&D

Research with a route to practical use

Razor created a dedicated AI team as rapid advances in machine learning and large language models began opening up new kinds of customer problem. I was the Lead Developer in that team: exploring unfamiliar techniques hands-on, turning ambiguous ideas into testable systems, and helping decide which approaches had enough practical value to take further.

This was not research isolated from software delivery. Each investigation had to connect a model to real data, an application, and an operational question. Some ideas became foundations for later products; others established what was possible, what data would be required, and where the limits were.

Early RAG and language-model systems

I carried out the initial R&D behind Razor's later LLM-based chat and data products while LangChain, RAG patterns, and Semantic Kernel were still young and changing quickly.

That work included document-powered retrieval in Python, natural-language access to relational databases, and evaluation techniques for answers that could be semantically correct without matching an expected sentence word for word. Those prototypes fed directly into the direction that became DataQI.

Recognising materials from images

One computer-vision prototype analysed photographs of mixed waste at a recycling or disposal site and reported which material types were present.

The interesting problem extended beyond choosing a classification model. Real piles are visually messy: materials overlap, lighting and scale vary, and the useful output is a description of a mixture rather than one clean object centred in an image. The prototype explored how a custom model and an application layer could turn that imperfect visual input into structured information.

Measuring throughput from live video

Another investigation analysed a live factory-floor video stream to compare the throughput of components across different tooling arrangements. I extracted image frames from the feed, ran them through TensorFlow models, added informational overlays to the video, and recorded statistics for later comparison.

This required treating video as a changing process rather than a collection of unrelated images: maintaining enough state to measure activity over time and producing an operational result that could inform how work was organised. It demonstrated the wider engineering around a vision model—stream handling, inference, visual feedback, aggregation, and a usable reporting path.

From IoT collection to energy forecasting

For an internal energy project, I built a system that collected electricity usage from smart plugs across Razor's office into a central dataset. The collection platform was valuable in its own right, but it also created the history needed for predictive modelling.

I then used that data to model future energy consumption and cost. It was a compact example of an end-to-end AI system: physical devices, continuous ingestion, data preparation, prediction, and an output connected to a real planning question.

Further applied experiments

The broader programme also covered OCR, further detection and classification work, and statistical models.

Not every R&D result needs to become a standalone product. A useful prototype can establish feasibility, reveal the data and operational constraints, or show that a different approach would be more sensible before a company commits to a full delivery programme.

Why I include this work

This work shows breadth beyond integrating a hosted language model:

• Leading applied AI exploration while remaining the engineer building the prototypes.
• Moving between LLM systems, computer vision, streaming analysis, IoT data, and predictive modelling.
• Connecting models to applications, live data, evaluation, and operational decisions.
• Learning emerging technologies early and carrying useful patterns into production-oriented products.
• Treating R&D as disciplined risk reduction rather than technology theatre.