AI practice
AI that survives contact with production
We build the AI systems companies are being asked about, and we tell leadership teams the truth about what those systems can do. Engineering first, with thirty years of delivery behind it.
pannonrex / delivery-agent
LiveWe run a 15-year-old Java banking core. We cannot take it offline. Where do we start?
Start where a mistake is cheap. Here is the sequence we would run:
- Mapped 1,240 stored procedures and 38 external integrations
- Identified 4 bounded contexts safe to extract first
- Estimated migration: 3 slices, no downtime, 9 months
- Drafting the characterisation test harness
Slice 01
Reporting
Read-only. Lowest risk, proves the pattern.
Slice 02
Payments
Shadow traffic first, then cut over by tenant.
Slice 03
Core ledger
Last. Only once the harness is trusted.
Ask about modernising a system you cannot switch off
Ask usAnswered live by our assistant, from what is on this site. Not a quote.
Try it yourself
Two of our own tools, open to anyone
Both are the real thing rather than a demo reel: the same first pass we would do, run by the same kind of system we build for clients.
Scope my project
Describe what you are building or modernising and get the brief back: the approach, three slices in delivery order, the risks and the questions we would still ask.
Write my briefAI readiness check
Six questions about what your system does and how it is governed, and a short memo back: the likely EU AI Act tier, the gaps, and what to do first.
Take the checkTwo ways in
Start with the conversation, or start with the code
Most clients do both, in that order. Neither requires you to have an AI team already.
Advisory & enablement
An AI expert in the room with your leadership team: what the technology actually does, where it pays off in your specific operation, what it costs, and what to do in the next two quarters. Followed by hands-on enablement so your people can run the next project.
- Executive briefings and board sessions
- AI readiness assessment with a ranked opportunity map
- Architecture reviews and vendor second opinions
- Developer and analyst workshops
Development & delivery
The build itself: retrieval that returns the right document, agents with scoped permissions, evaluation suites that catch regressions, and integration into the ERP, CRM and legacy services you already run. Deployed in your cloud, owned by you.
- Retrieval-augmented generation on your own corpus
- Agentic workflows with audit trails
- Evaluation, observability and cost control
- Integration with existing enterprise systems
Under the hood
The parts that decide whether it works
Retrieval
Getting the right document is most of the problem
A model can only be as good as what you put in front of it. We tune chunking, hybrid search and re-ranking against your actual corpus, and enforce access control at retrieval time so nobody can prompt their way into a document they are not cleared for.
- Hybrid keyword and vector search with re-ranking
- Permissions applied before the model ever sees a document
- Citations users can click through to the source
retrieval/answer.ts
1// Retrieval tuned against your corpus, not a benchmark23const hits = await retrieve({4 query,5 mode: 'hybrid', // BM25 + vector6 rerank: true,7 filter: { tenant, clearance }, // access control at retrieval8})910return answer(query, hits).withCitations()Indexing 412,880 documents — 94% complete
Evaluation
A number that either went up or went down
Without evaluation, every prompt change is a guess and every model upgrade is a risk nobody can size. We build graded test suites from your real cases and wire them into CI, so quality becomes a release gate rather than an opinion.
- Golden datasets built from your own traffic
- Automated graders plus targeted human review
- Quality, latency and cost tracked together per release
Release 2.14 — evaluation
PassingCompared against the previous production build
- Answer accuracy
- 94.2%+3.1
- Citation coverage
- 99.1%+0.4
- Hallucination rate
- 0.6%−1.2
- p95 latency
- 1.4s−0.6s
- Cost per 1k queries
- $2.10−41%
Integration
The model is the easy part to replace
What takes the time is connecting AI to systems that were designed before any of this existed. We keep the integration layer provider-agnostic, so the expensive work survives your next change of model, vendor or pricing plan.
- Deployed into your tenancy, inside your data residency rules
- Hosted or open-weight models, chosen per workload
- No rewrite of the systems you already depend on
- SAP / ERP
- Salesforce / CRM
- Data warehouse
- Legacy services
- Document store
- Identity & SSO
Where it pays off
Use cases that tend to survive the business case
Drawn from what has actually worked for clients, rather than what demos well.
Knowledge retrieval
Contracts, specifications, tickets and manuals made answerable in natural language, with every claim traceable to the paragraph it came from.
Process automation
Agents that read a case, gather what they need from your systems, prepare the decision and escalate the exceptions to a person.
Engineering acceleration
AI-assisted development inside your own codebase, with the review discipline and test coverage that keeps the output trustworthy.
Analysis and reporting
Structured extraction from unstructured sources, feeding the dashboards and models your analysts already trust.
Customer-facing assistants
Support and sales assistants grounded in your product knowledge, with hard limits on what they are allowed to promise.
Risk and compliance
Monitoring, classification and audit trails built to satisfy the people who will eventually ask how a decision was made.
Governance
The questions your risk team will ask
We would rather answer them at the start of the project than at the end of it.
Your data stays yours
Deployed into your cloud account or on-premises. Enterprise model endpoints with no-training guarantees, or open-weight models where that is not enough.
Every answer is traceable
Citations, execution traces and immutable logs, so an auditor can reconstruct why the system said what it said.
Humans keep the consequential decisions
Agents propose; people approve, wherever the outcome is expensive, irreversible or regulated.
Regulation is designed in
EU AI Act classification, data lineage and human-oversight mechanisms handled during the build rather than retrofitted before an audit.
Questions
What companies ask us first
We keep hearing we need an AI strategy. Do we?
You need a short list of problems where AI is genuinely the cheapest solution, and the discipline to ignore the rest. That is what an assessment produces. A strategy document with no use case attached to it is not worth paying for.
What does a first AI project typically cost?
A readiness assessment is a fixed-price engagement of two to three weeks. A first production pilot usually runs four to six weeks of a small team. We give you both numbers before you commit to either.
Can you work with our existing AI vendor or platform?
Yes. We are frequently brought in as the engineering side of a platform somebody else sold, or to give an independent second opinion on whether a proposal is realistic.
What happens to the system when the engagement ends?
You own the code, the infrastructure definitions, the evaluation suite and the documentation. Enablement is built into the engagement specifically so your team can carry it forward.
Do you build with a specific model provider?
No. We choose per workload and keep the integration layer provider-agnostic, so switching model or provider later is a configuration change rather than a rebuild.
Bring an AI expert into your next meeting
A free first consultation with a senior engineer who builds these systems. Come with your actual problem — we will tell you whether AI is the right answer to it.
Typical reply within one business day · Initial consultation is free