> qai run --mode reactor --stream✓ gate passed, providers 10+QAI Lab
LLM orchestration platform with isolated memory per domain, graph-backed retrieval and a gate that scores work before it reaches the user.
Systems · Infrastructure · Cloud · AI
I build the layer underneath the product.
The services, the network, the infrastructure they run on, and the retrieval that makes model output dependable enough to rely on. Before this, twenty-two years running international operations: shift patterns, bad data, audits, and people who need an answer now. Those are the conditions I design for, because they are the ones software actually meets.
whoamiAI systems & infrastructure engineer, Glasgowsystems --built5 end to end: QAI Lab, DocuMiner, Newis, Cairn, Kelvinstack --coreAWS, Terraform, Python, FastAPI, PostgreSQL, Linuxcredentials --verifyCisco CCNA, 2 papers with permanent DOIsexperience --prior22 years running international operationsfocusAI infrastructure, retrieval, cloud, networkingstatusAvailable now, hybrid, remote or project
AvailabilityImmediate · UK right to workLocationGlasgow · hybrid or remoteFocusAI infrastructure · cloud · backend
WHAT ONE ENGINEER
SHIPPED, END TO END
5 SHIPPED · 0 HANDOFFS
✓Takes a product spec → system design
✓Builds infra with Terraform + AWS
✓Orchestrates LLMs, GraphRAG, retrieval
✓Ships backend + frontend, no handoff
5 SHIPPED · 0 TEAM
QAI Lab
CV ON REQUEST
Selected systems
Five products, built end to end> documiner ingest ./bundle --local✓ entities linked, provenance keptDocuMiner
Document intelligence that answers across a bundle rather than inside one file, with every claim keeping the page it came from.
> terraform apply -auto-approve✓ multi-AZ, private subnets, ECSNewis
Fleet intelligence and compliance on a Terraform-provisioned AWS stack: load balancing, containerised app tier, isolated data tier.
> cairn solve --graph glasgow-metro✓ detour 4.8ms, SLA heldCairn
Route optimisation and in-cab co-pilot sharing the same trip log as the fleet platform, with a hazard model that keeps its evidence.
> kelvin night-cover --verify strict✓ frozen rules, maker / checkerKelvin
Autonomous CRM operations under maker and checker governance, with frozen rules that sit above the model and cannot be reasoned around. Tested across 33 sets, 330 runs.
> ls ./other✓ storefront, logistics, captureFurther work
A reseller storefront with an automated product pipeline, an operational logistics platform, and a mobile capture app that streams documents over the local network.
Methods
How the systems above are builtFour techniques, and the question that decides between them. Does the answer sit inside one document? Then retrieval is enough. Do the relationships carry the meaning? Then a graph. Does the output vary between runs? Then a loop with a gate. Can the system change something in the world? Then a harness. Choosing the heavier method when the lighter one suffices is the most common way these systems get expensive without getting better.
Pattern
The field, and where I sit in it
flowretry or stopfeedbackPattern level, not implementation
What this is
The shape of AI engineering as the field describes it: building and deploying applications, software engineering fundamentals underneath, coding agents alongside, and shaping what gets built. The four methods below are how the middle two are actually done.
Why it is here
A stack list says what tools someone has touched. A map says whether they understand the field they are working in, and where their own work sits inside it.
Where I am partial
Operating at scale. The systems are built, deployed and tested, but none has carried production traffic from the public, and I would rather mark that than imply it.
Everything else
Covered by the five systems above, end to end: applications, data, architecture, security and reliability.
QAI Lab, the certification gate
Console- Problem
- One assistant serving research, engineering and operations carries context across domains it should never cross. Answers drift, and nothing notices.
- The platform
- A REST API over the whole platform, running in two modes: domain reasoning, and an agentic development environment. Ten or more providers behind one interface.
- Gate and learning path
- Work is scored before it reaches the user, weak output regenerates, and the review seat fails closed. What each run establishes carries forward.
Product interface, scripted sequenceNot connected to production data
DocuMiner, retrieval that crosses documents
Console- Problem
- Legal and accounting firms hold the answer across a bundle, not in one file. Similarity search cannot answer what is connected to what.
- Extraction
- Each page is read by whichever method suits it, chosen automatically. A scanned 29-page bundle is transcribed in under a minute, in reading order, and the offline path works with no network.
- Failure
- Confidence is computed from what extraction produced, so a twenty-page PDF that yields zero words is stored as failed, never silently as confident.
Product interface, scripted sequenceIllustrative comparison
Newis, fleet command centre
Console- Problem
- Any organisation running vehicles has to prove compliance across drivers, documents and depots, and keep the driver safe while doing it.
- Infrastructure
- Terraform-provisioned AWS: load balancer across two public subnets, containerised app tier on ECS Fargate in private subnets, PostgreSQL in an isolated data tier with no default route, Cognito groups, Secrets Manager.
- Decisions
- Defined in code, so it rebuilds the same way every time and costs nothing while down. Capacity and network layout were planned before the first deployment.
Product interface, scripted sequenceSeeded dataset
Cairn, route optimiser
Console- The co-pilot
- Cairn is the driver-facing layer, sharing the trip log, so every active shift is a live route session with an in-cab voice co-pilot across seven languages.
- Hazard model
- Each alert keeps its confidence score and the reports behind it, a per-stop composite risk breakdown, and a lone-worker check-in.
Product interface, scripted sequenceNot connected to production data
Kelvin, night cover
Console- Governance
- Maker and checker on every action: the platform proposes, a verifier clears it, and consent and policy are checked before anything leaves.
- Frozen rules
- Above the model and not negotiable: no financial figures without approval, no credit issued autonomously, sensitive cases routed to a named human.
- Measured
- The signal tracked is how often a human had to correct it, not how much it produced. Tested across 33 sets, 330 runs, before it was trusted with real work.
Product interface, scripted sequenceShown with synthetic data
Working with me
What I am strongest at
- Backend and APIs. REST services other systems call, with the reliability work that decides whether an integration survives its first bad week.
- Cloud infrastructure. AWS defined in Terraform, private by default, rebuilt from code and switched off when nobody is using it.
- Networking. Addressing, routing, segmentation, TLS. Cisco CCNA certified, and the network was designed before the servers.
- Retrieval and model orchestration. Dependable answers out of models, and recognising when an answer is not dependable enough to send.
How I work
- Every number here came from a run. Not an estimate, not a round figure chosen because it looked convincing.
- I hand over systems, not secrets. Documented, reproducible, runnable by someone who has never met me.
- I fail closed. When a check cannot return a verdict the work stops.
- Working and finished are different. Finished is when someone else can run it on a Sunday without calling me.
In a team
Two decades of leading international operations: teams across multiple plants and countries, supplier and logistics networks, compliance, and delivery dates set by customers rather than by me. Handover, audit and working to someone else's standard were the job.
That is the half engineers usually learn on the way up. I bring it already, alongside five platforms built end to end, so I can hold a technical conversation and a commercial one in the same meeting.
Experience
Two careers that feed each other2023 – PRESENT
CaledonianAI
Glasgow, Scotland
AI systems & infrastructure engineer
- Designed, built and deployed five products from a blank page as the only engineer, across backend, data layer, infrastructure and frontend.
- Built the orchestration platform behind them: isolated domain memory, graph-backed retrieval, a scoring gate in front of every response, and a review loop that fails closed.
- Provisioned the AWS environments in Terraform: load balancing, containerised app tiers, private subnets, managed PostgreSQL, identity groups and secrets handling.
- Published two technical papers with permanent DOIs, grounded in what the platform does.
2007 – 2023
International operations
Supply chain and logistics
Operations & supply chain leadership
- Ran international operations end to end, from supply chain through to delivery at the customer.
- Just-in-time automotive distribution, where a late delivery stops a production line and the tolerance for error is zero.
- Built and ran an industrial facility, then managed operations across multiple plants, including the people, the compliance and the systems.
- This is the reason the software is built the way it is: audits, shift handovers, incomplete data and deadlines are the normal case.
Research
Permanent identifiersDomain-isolated architecture for reliable LLM inference
The isolation and routing approach behind the orchestration platform: separate execution memory per domain, explicit routing between them, and the effect on answer reliability when domains stop sharing context.
DOI 10.5281/zenodo.19264497The biological prerequisite for artificial general intelligence
Argues that probabilistic computation alone cannot produce cognition, developed against operational testing across more than fifty model configurations.
DOI 10.5281/zenodo.19264199Credentials
Labelled as it standsBSc Computing ScienceUniversity of the West of Scotland, expected 2028
In studyBA Business AdministrationAkdeniz University
EarnedCisco CCNACisco, enterprise networking
CertifiedAWS Solutions Architect AssociateSAA-C03
In progressAWS Generative AI ProfessionalAmazon Web Services
In progressClaude Partner NetworkAnthropic, Scotland
MemberContact
Open to agentic AI systems, AI infrastructure, AI platform, LLM systems, backend and cloud engineering roles, and to project work for companies that need this built properly the first time. Scotland based, hybrid or remote.
Social responsibility
Work with no commercial returnThe systems on this page were built to be sold or to be hired for. This one is neither. Autistic children are among the heaviest users of assistive technology and the least served by it, because the products that reach them are built for a median user who does not exist. I would rather spend the same engineering discipline on a problem where the return is somebody else’s day getting easier.
A companion that listens all day, and keeps what it hearsProposal · nothing built, no data held
A child’s patterns show up in small things spread across a whole day, and by the time a parent or a therapist sits down to write notes, most of it is gone. What survives is what an adult remembered to record, which is not the same thing as what happened.
Something that is simply present and can be talked to for as long as the child wants, keeping behaviour and speech as a continuous record rather than a survey. Turn taking has to feel immediate, so speech in and speech out both run at conversational latency instead of the pause a round trip to a hosted model costs.
One child, one domain, no shared execution memory and no route between domains. This is the architecture in my first paper, applied where the reason for it is not reliability but the fact that one child’s day must never be able to reach another’s.
Where something outside the sandbox would genuinely help the child, a narrow MCP server exposes it deliberately, one capability at a time, each reviewed before it is added. The default is that nothing leaves.
This is health data about a minor, so it is special category data under UK GDPR and it sits inside the PVG scheme in Scotland. Guardian consent before anything runs, withdrawn at any time with the record destroyed. Everything runs locally on open source models, on the family’s own device, so there is no processor and no transfer. Nothing the child says is used to train anything. A named clinician reads the record rather than a model acting alone, and the system offers observations, never conclusions.
It does not get built without a clinical partner and an ethics review. I am an engineer, not a clinician, and a system like this designed only by an engineer is exactly the kind that causes harm while meaning well.
Architecture view · scripted sequenceSynthetic records · nothing built