Skip to content

Work

Work, labelled for what it is.

What Northkeep is delivering now, what the company delivered before it carried the Northkeep name, and the engineering track record of the founder who leads it.

How to read this page

Every case says who delivered it.

Northkeep engagements
Contracted with Northkeep and delivered by Northkeep.
AdvancIT OOD, before Northkeep
Delivered by AdvancIT OOD, the company that now trades as Northkeep, before the Northkeep name existed.
Founder's experience
Work the founder delivered in previous roles and through other firms. It shows the engineering foundations Northkeep is built on; it is not presented as Northkeep's own delivery.

Clients are described by sector, never by name. We describe the engineering in detail and publish specific figures only where they are cleared for publication.

Northkeep engagements

Delivered by the company.

Dairy and food production, Bulgaria

Established business In delivery

An AWS data platform with AI-assisted access to business data

Implemented by a Northkeep engineer under the founder's architectural supervision.

Scope

  • AWS platform and landing-zone engineering
  • Data ingestion and transformation pipelines
  • A governed analytical data platform
  • AI-assisted access to business information

Active engagement. Results will be published after acceptance, and only with the client's agreement.

  • AWS
  • Data lake
  • Analytics
  • AI-assisted access

AdvancIT OOD, before Northkeep

The same legal entity, earlier work. Presented as the company's history, not as Northkeep-branded delivery.

Non-bank consumer credit, central-bank regulated

Lending platforms and ERPs for regulated consumer lenders

Delivered by AdvancIT OOD, the company that now trades as Northkeep, before the Northkeep name existed.

The problem

A regulated lender cannot start manually and formalize later. Supervisory reporting, the data-handling model, and the audit trail have to exist before the first loan is written, and the systems have to keep up as the business grows rather than be rewritten along the way.

What was done

Loan-origination platforms built from the ground up and grown into full lending ERPs: origination, servicing, collections, and the reporting interfaces the central bank requires, with the regulated-data security model designed in rather than retrofitted. The platforms were built, extended, and supported through the lenders' growth, then handed over. On-premises, by era - this predates the cloud work, and the regulated-delivery discipline is the same one DORA and NIS2 now ask for in writing.

Scope
Origination to full ERP

Origination, servicing, collections, and supervisory reporting in one platform.

Regulatory regime
Central-bank supervised

Built to operate under Bulgarian National Bank supervision from the first loan.

  • Bulgarian National Bank
  • Regulated lending
  • ERP

Founder's experience

The track record Northkeep is built on.

More than twenty years of software, cloud, and data engineering, including three CTO roles. These cases were delivered in previous roles and through other firms - not contracted through Northkeep.

Regulated lending, UK mortgage and consumer credit

A regulated lending back office, rebuilt on AWS

Founder's experience as CTO.

The problem

The business ran on a legacy CRM that could not carry a regulated lending process end to end. The stages that mattered - qualifying a lead, assembling an application, evidencing an affordability decision, reporting to the regulator - depended on people moving data between systems by hand, which is slow, and in a regulated business is exactly where the evidence trail breaks.

What was done

A greenfield rebuild on AWS over several years, delivered by an engineering team under the founder's management. The platform carries the regulated lifecycle in one flow: lead to opportunity, application, affordability and approval, drawdown, repayment tracking, and renewals. Open Banking for affordability evidence, credit-bureau data, identity verification and KYC, and regulatory reporting all sit behind a single process designed for straight-through processing.

Built
Greenfield, on AWS

A rebuild of the business system, not a migration of the old one.

Role
CTO, leading the team

Architecture, delivery, and the engineering team, end to end.

  • AWS
  • FCA
  • Open Banking

Retail software, product company

Generative AI inside a live software product

Founder's experience, delivered through another firm.

The problem

A generative-AI capability had to run inside a live product rather than beside it as a demonstration - which meant it had to be governed, observable, evaluated against a measurable baseline, and operable by the client's own engineers. The gap was not model capability; it was everything production requires around the model.

What was done

A production architecture on Amazon Bedrock with agent orchestration, integrated into the product platform to classify and enrich product data. The design covered the parts that decide whether an AI feature survives contact with production: guardrails, identity and access for agent actions, evaluation against a graded baseline, and observability that makes failure legible to the team operating it.

Where it runs
Inside the product

Built as a product capability, not a side demo.

Operated by
The client's own team

Designed so the people running it can see when it fails, and why.

  • Amazon Bedrock
  • Agent orchestration
  • Evaluation

SaaS, file and media library

AI search, measured against acceptance criteria agreed up front

Founder's experience, delivered through another firm.

The problem

The product's keyword search could not answer the way its users actually asked. Real searches, against files that existed in the library, returned nothing - and in a self-serve product, an empty result is where users give up.

What was done

A search layer rebuilt on two paths and measured separately rather than blended. Underneath, OpenSearch reconfigured for fuzzy matching, field weighting, and dense-vector retrieval, with descriptions and multimodal embeddings generated on Amazon Bedrock. In front of it, an agent that classifies intent, builds hybrid queries, enforces per-file permissions, refines iteratively, and acts through the product's own API. Acceptance was agreed with the client in advance as a measurable bar rather than a demo, and a repeatable benchmark harness was built to score against it.

Acceptance
Agreed up front

A measurable bar set with the client before the build, not a demo afterwards.

Measurement
Repeatable, unblended

Retrieval and agent paths scored separately, with a harness that can be re-run.

  • Amazon Bedrock
  • OpenSearch
  • Multimodal embeddings

Enterprise software vendor, regulated customer base

A regulated multi-site estate, migrated and modernized on AWS

Founder's experience, delivered through another firm.

The problem

The vendor's core platform ran as an on-premises install inside each customer's own network. That pushed infrastructure, security, and scaling onto the customers, capped what the product could become, and left analytics stranded: every site held its own data, and there was no queryable foundation to build anything new on.

What was done

Assessment through build as one continuous program. A production landing zone on AWS Organizations, Control Tower, and Account Factory for Terraform, with Terraform and Terragrunt throughout, hub-and-spoke networking, and organization-wide security baselines - designed so onboarding a customer is a codified operation rather than a hand-assembled one. A repeatable migration pattern, and a database conversion onto Aurora PostgreSQL with a phased cutover runbook. At the centre, an event-driven data lake: ingestion into S3, orchestrated ETL through AWS Glue and Step Functions, and open Iceberg tables catalogued for direct SQL, with a semantic layer and BI on top and legacy reporting rebuilt serverless against the same layer. Delivered as patterns, infrastructure as code, and runbooks the client's engineers execute and extend.

Customer onboarding
Codified

Infrastructure as code and runbooks, so each additional customer is a repeat rather than a project.

Data foundation
Open table formats

An event-driven data lake the product's analytics build on.

  • Control Tower and AFT
  • Terraform
  • AWS Glue and Iceberg
  • Aurora PostgreSQL

B2B SaaS

Modeled projection

EU data residency enforced by architecture

Founder's experience, delivered through another firm.

The problem

EU enterprise prospects were raising data residency in security reviews, and the questions that follow the first one - where the control plane sits, who can reach tenant data from outside the EU, what a residency commitment is actually enforced by - needed architectural answers rather than contractual ones.

What was done

A dual-region target architecture that enforces residency structurally rather than by policy: network isolation between regions, tenant-to-region pinning at the identity and routing layer, and a cross-region control plane designed so operational metadata can move without personal data following it. A governed Amazon Bedrock AI/ML layer was designed inside the same boundary, and the whole was costed and sequenced as a phased plan.

Residency enforcement
Structural, not contractual

Designed so the residency answer survives the follow-up questions in a vendor-risk review.

Stage
Architecture and plan

A design-stage deliverable; cost figures in it were modeled, not measured.

  • AWS
  • EU data residency
  • Amazon Bedrock

Retail group, multi-brand

Recommendation systems and high-traffic APIs for a large retailer

Founder's experience, delivered through another firm.

The problem

Recommendations were a live, customer-facing surface across many brands, which makes them a latency and availability problem before they are a machine-learning problem. Traffic arrived in unpredictable spikes, and a recommendation that arrives after the page has rendered is the same as none - except that it still costs money to produce.

What was done

A data engineering and ML platform built and run over several years, orchestrating many production recommendation pipelines across the group's brands, and the serving layer in front of them. The serving tier was designed serverless so it could absorb spikes without provisioning for the peak, with cold starts engineered out rather than tolerated. Payload size, caching, and inference placement were tuned together, because at high request volume an inefficient design is a line item rather than a rounding error.

Serving
Serverless, spike-tolerant

Absorbs unpredictable traffic without provisioning for the peak.

Breadth
Multi-brand, multi-year

Production pipelines across a brand estate, built and run over several years.

  • Recommendation systems
  • Serverless APIs
  • MLOps
  • Data pipelines

More from the same track record.

Shorter entries, same rules.

Multiple client engagements

Multi-account landing zones, built to be handed over

More than ten multi-account AWS landing zones over the past three years, several built on Account Factory for Terraform, with account vending designed to extend as the estate grows.

AWS Control Tower AFT Terraform

Blockchain and GPU/ML platforms

Migrations from other platforms into AWS

Multi-region blockchain infrastructure migrated into AWS, and a GPU-heavy ML training and inference platform replatformed from another cloud provider onto Amazon EKS.

Migration Amazon EKS

Research data platforms

Cloud cost reduction from evidence, not assumption

Storage and analytics cost reduced on large research data platforms by measuring access patterns first and redesigning tiering, lifecycle, and retention on what the evidence showed - with the basis for every saving recorded so it could be checked afterwards.

FinOps Storage lifecycle

Large multi-region production estates

Well-Architected remediation at production scale

Repeated Well-Architected remediation across large multi-region production estates: threat detection, audit-trail hardening, encryption by default, and IAM rebuilt in Terraform.

Well-Architected Security

Credentials

Credentials you can check.

Held by the founder, and supporting evidence rather than the argument.

20+
Years of software engineering and technical leadership.
3
CTO roles, one ending in a profitable trade sale.
All
Active AWS certifications held - the Golden Jacket, AWS's recognition for holding every AWS certification at once, all current.

Every active AWS certification Verify on Credly

Microsoft Azure

  • Microsoft Certified: Azure Solutions Architect Expert
  • Microsoft Certified: Azure Administrator Associate

Our production delivery experience runs deepest on AWS; Azure engagements are scoped to the depth we can stand behind.

Regulatory fluency

  • DORA
  • NIS2
  • EU AI Act
  • EU Data Act
  • AWS European Sovereign Cloud

Partner

AWS Partner Network Member

Start a conversation

Tell us what you need built.

Across Cloud & Platform Engineering, Software Engineering, and Data & AI Engineering - as a project, with engineers inside your team, or behind your own client relationship.

Discuss a project Talk to us about delivery