EU AI Act high risk obligations are now enforceable. Check your exposure
Insights About us Careers
Contact us
Generative AI & LLM Engineering

Custom LLM Application Development

A production application built on large language models, designed around your data, your permissions model and your cost ceiling rather than around a demo script.

8 to 14 weeks
Typical build
Fixed scope
Commercial model
Yours in full
Code and IP

Almost every organisation can build an impressive large language model demo in a fortnight. Very few can turn that demo into software that a thousand people use on a Tuesday morning without a support queue forming behind it. The gap is not the model. It is everything the demo quietly skipped.

In one paragraph

Custom LLM application development is the design and engineering of production software built around a large language model: the retrieval layer that grounds it in your data, the orchestration that decides what happens on each request, the evaluation harness that proves quality before release, and the guardrails, observability and cost controls that keep it dependable once real users arrive.

What custom LLM application development actually involves

The model is a component, not the product. In a typical build it accounts for perhaps a tenth of the engineering effort. The rest is the work that decides whether the thing survives contact with real users:

  • Grounding. Which of your systems the application can read, how that content is chunked, indexed and retrieved, and what happens when the answer is not in the corpus at all.
  • Orchestration. Whether a request is a single model call or a sequence of steps with tool use, validation and fallbacks, and where the boundaries between them sit.
  • Permissions. Retrieval that respects the access rights of the person asking, so the application cannot summarise a document its user was never allowed to open.
  • Evaluation. A scored regression suite that runs before every release, so a prompt change cannot silently degrade quality in a corner of the product nobody tested by hand.
  • Guardrails. Input and output validation, prompt injection resistance, refusal behaviour, and a defined path for the cases the system should not attempt.
  • Economics. Token accounting per request and per user, caching, model routing and a projection of what the application costs at ten times current volume.

Skip any one of these and the application still demos beautifully. It just fails later, more expensively, and usually in front of somebody senior.

How we build it

Narrow the scope until it is defensible

We start by cutting the brief down to the smallest thing that is genuinely useful. A product that answers four question types correctly is worth more than one that attempts forty and is right most of the time. Scope discipline early is what makes the evaluation set possible, and the evaluation set is what makes everything after it measurable.

Design the retrieval before the prompt

Output quality is capped by retrieval quality, so the first architectural decisions are about content: what is indexed, how it is segmented, how freshness is handled and how conflicting sources are resolved. Where retrieval is the dominant risk we run it as its own workstream, described in more detail under RAG system development.

Build the evaluation harness alongside the feature

A golden set of real questions with agreed correct answers is assembled in the first fortnight, not the last. Every subsequent change is scored against it. This is the single practice that most separates applications that improve over time from applications that drift, and it is covered in depth in LLM evaluation and benchmarking.

Make failure visible and recoverable

Every request is traced. Latency, token spend, retrieval hits, tool calls and user feedback are recorded against a request identifier so that a complaint on Thursday can be reconstructed exactly. Rate limits, timeouts and model outages have defined fallbacks rather than a spinner.

Worth knowing

Model choice is a runtime decision, not a religion

We build against an abstraction so the underlying model can change without a rewrite, then benchmark candidates on your data. Prices fall and capabilities move every few months; an application welded to one provider is a migration project waiting to happen. Where the choice is contested before the build starts, AI vendor and model selection settles it with evidence.

Where LLM applications usually break

FailureWhat it looks like in productionWhat prevents it
Silent quality driftA prompt tweak improves one use case and quietly breaks three othersGolden set plus regression run in CI
Retrieval starvationConfident answers assembled from the wrong three paragraphsChunking strategy, hybrid search, reranking, retrieval metrics
Cost surpriseUnit economics that only work at pilot volumeToken accounting per request, caching, model routing
Permission leakageA user is shown content from a document they cannot openAccess-aware retrieval and per-user index filters
No rollbackA bad release cannot be reversed without a redeployVersioned prompts and configuration, feature flags

None of these are exotic. They are the predictable consequences of treating an LLM application as a prompt with a user interface attached, and each one is cheaper to design against than to diagnose later.

Process

How the engagement runs

Every stage produces something you can inspect, so you always know what is being built and what it currently scores.

Weeks 1 to 2

Scope and evaluation set

We agree the question types in scope, assemble a golden set of real examples with correct answers, and set the quality threshold the application has to clear before release.

Weeks 3 to 5

Retrieval and architecture

Content is indexed, the retrieval strategy is benchmarked against the golden set, and the orchestration, permissions and fallback design is fixed and documented.

Weeks 6 to 10

Application build

The product is built end to end: interface, orchestration, guardrails, tracing and cost accounting, with the regression suite running on every merge.

Weeks 11 to 13

Hardening and pilot

Prompt injection testing, load and latency work, cost projection at target volume, then a controlled pilot with a defined user group and a feedback loop.

Week 14

Handover

Code, prompts, evaluation sets and infrastructure definitions transfer to you, with a working session for the team that will own it.

Deliverables

What you receive

Everything needed to run, change and defend the application after we leave. Delivered in editable form, with full IP transferring on final payment.

01

Production application

The deployed application with source code, infrastructure as code and a documented build and release path.

02

Retrieval and index design

The indexing pipeline, chunking strategy and retrieval configuration, with the benchmark results that justified them.

03

Evaluation harness

The golden set, the scoring code and the regression suite wired into your pipeline so quality is checked on every change.

04

Guardrail and safety report

Injection testing results, refusal behaviour, validation rules and the residual risks we could not design out.

05

Cost model

Token accounting per request type with a projection at your target volume and the levers that move it.

06

Operations runbook

Tracing, alerting, failure modes, rollback procedure and the first ninety days of recommended monitoring.

Fit check

Is this the right engagement?

Worth being direct. Custom LLM Application Development is the wrong spend in some situations, and those are listed rather than buried.

Good fit if

  • A demo already exists internally and the question is how to make it dependable enough to release.
  • The application has to read your own content and respect who is allowed to see what.
  • Quality has to be provable to someone outside the team before launch.
  • You need a defensible unit cost per request before committing to a rollout.
  • The team will maintain the application afterwards and needs the evaluation practice as much as the code.

Choose something else if

  • You have not yet decided which use case to build. Start with AI use case discovery.
  • The real question is whether an off-the-shelf product would do. That is build vs buy advisory.
  • You need to test feasibility cheaply first. A four to six week proof of concept is the smaller bet.
  • The system needs to take actions across your systems rather than answer questions. That is agent work, not application work.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

How long does custom LLM application development take?

Eight to fourteen weeks for a first production release, depending on how many source systems are involved and how strict the security review is. The evaluation set is built in the first fortnight, so you have a quality signal long before launch rather than an opinion at the end.

Do we need our own machine learning team to maintain it?

No. Almost all of the ongoing work is ordinary software engineering: updating prompts and configuration, adding cases to the evaluation set, watching cost and latency. We hand over the evaluation practice with the code, because that is the part teams most often lack rather than the deployment skills.

Which model will the application use?

Whichever one performs best on your evaluation set at an acceptable cost and latency, and we build so it can be changed later. In practice most applications end up routing simple requests to a cheaper model and reserving the strongest model for the cases that need it.

Can the application run inside our own infrastructure?

Yes. Where data residency, procurement or sector rules require it, the same architecture is deployed against a self-hosted model. That path is described under private and self-hosted LLM deployment.

What happens if quality is not good enough?

You find out during the build rather than after it, because the threshold is agreed in week one and scored continuously. If the target is not reachable with the data available, we say so and set out what would have to change, which is usually a content problem rather than a model problem.

Is this the right engagement?

Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.