Generative AI development
Generative AI Development Services
We help teams find answers in company policies, tickets, contracts, and manuals. We build systems that retrieve relevant information, cite source documents, and pass accuracy tests before each release. Scoping takes two weeks and ends with a fixed quote and an assigned engineer.
Question
Can a contractor install the VPN client on a personal laptop?
Answer, from your documents
No. Contractors connect through the managed browser; the VPN client is for company devices only.
- Remote access standard, section 3.2
- Device policy, section 2
- Average eval score
- 97%
- To a fixed quote
- 2 weeks
- AI engineers
- 30+
- Generative AI services
- 6
Models and sources
Models and sources behind our generative AI development services
Generation
OpenAIGeneration
GeminiGeneration
Mistral AISelf-hosted model
LlamaIndexRetrieval
ElasticsearchHybrid search- Generation
- OpenAIGeneration
- GeminiGeneration
- Mistral AISelf-hosted model
- LlamaIndexRetrieval
- ElasticsearchHybrid search
PostgreSQLVector store
AWS BedrockPrivate endpoint
Vertex AIPrivate endpoint
SharePointPolicies
ZendeskTicket history
SalesforceCase notes- PostgreSQLVector store
- AWS BedrockPrivate endpoint
- Vertex AIPrivate endpoint
- SharePointPolicies
- ZendeskTicket history
- SalesforceCase notes
What we do
What our generative AI development services cover
A system built by our generative AI development services is a retrieval layer, a model behind it and a gate in front of it. It takes a question in your team's own words, finds the passages in your corpus that bear on it, answers from those passages, and shows where each claim came from. The output is an answer people can use and check, not simply a paragraph that sounds right.
These projects often become difficult because the corpus needs care, an answer cannot be traced to a source, or a release was approved from a promising demo rather than a representative evaluation. A dependable system needs to show when an answer is supported and when a person should review or take over.
So our generative AI development services build the checkable part first. One corpus, one owner on your side, a set of real questions drawn from what your people already ask, and a measured score on those questions before anything reaches a user. Breadth comes later, from a second corpus, not from asking one system to be good at everything.
- Scope
- One corpus, one owner
- Every engagement names one owner on your side who can decide what the system is allowed to say.
- Ceiling
- Answers from your sources
- The system answers from retrieved passages, and says the corpus does not cover it rather than filling the gap.
- Proof
- Eval gate, every release
- A question set drawn from your own history, a threshold agreed up front, and only the run that clears it ships.
- Ownership
- Yours at handover
- Code, prompts, the index, the evaluation sets and the dashboards, handed over with the training to run them.
01 · Sub-services
Custom generative AI development services, by what you need built
Six ways into our generative AI development services. Each one opens its own page, with the integrations it needs, the way it is evaluated and its own cost band. The line that matters on each card is the last one: the way this particular technique fails, and what we do about it.
01AI chatbot development
A chat interface over your own content that answers in your language and refuses the questions it should not attempt.
Where it goes wrong
A chatbot needs particular care just outside its scope, where an unsupported answer may be less helpful than a clear handoff. We set a refusal threshold and connect the customer with a person, with the transcript attached, whenever confidence or policy calls for review.
Typical timeline
6–10 weeks
Includes
- Retrieval over your corpus
- Refusal and handover rules
- Conversation memory
- Channel integration
- Answer-level analytics
02RAG development
The retrieval layer itself: ingest, chunking, embeddings, ranking and the citation path, built to be measured rather than demonstrated.
Where it goes wrong
Retrieval quality, not the model, decides whether this works, and it degrades quietly as the corpus grows. We score retrieval separately from generation, so a drop shows up as a number on a dashboard rather than as a complaint from a user.
Typical timeline
6–10 weeks
Includes
- Ingest and chunking strategy
- Embedding and index choice
- Reranking and filtering
- Citation and traceability
- Retrieval scored on its own
03AI knowledge assistant
An assistant across the places your knowledge actually lives, that respects who is allowed to see what and knows which version is current.
Where it goes wrong
The corpus contains contradictions, superseded versions and documents most people are not allowed to read. We record precedence and permissions at ingest and check access when each question is asked, so the assistant inherits your access rules instead of flattening them.
Typical timeline
6–10 weeks
Includes
- Multi-source ingest
- Permission inheritance
- Version and precedence rules
- Freshness and reload schedule
- Usage and gap reporting
04LLM development
An application with a model inside it: the prompts, the routing, the caching, the fallbacks and the cost path, built as software rather than as a prototype.
Where it goes wrong
The prototype works and the cost per call makes it unshippable at volume. We cost the call path during scope, with caching and routing in place, rather than discovering it in the first month's bill.
Typical timeline
6–10 weeks
Includes
- Prompt and context design
- Model routing and fallback
- Caching and cost control
- Structured output and validation
- Latency budget per call
05LLM fine-tuning
Training a model on your examples to fix a format, a tone or a classification that prompting keeps getting wrong.
Where it goes wrong
Fine-tuning is mainly a format, style and task-specialisation tool, not a reliable way to add changing knowledge. We first check whether retrieval or prompting meets the need more simply, then recommend fine-tuning only when the evidence supports it.
Typical timeline
Set after dataset review
Includes
- Dataset design and labelling
- Training and evaluation runs
- Base-model migration plan
- Held-out test set
- Serving and version control
06Conversational AI development
Spoken and multi-turn systems where the answer has to arrive inside a human pause, not inside a page load.
Where it goes wrong
Latency is the product. An answer that is correct at four seconds has already lost the caller. We budget the round trip per turn during scope, and the retrieval design has to fit inside it.
Typical timeline
8–14 weeks
Includes
- Turn and latency budget
- Interruption handling
- Telephony or channel integration
- Escalation to a person
- Transcript and quality review
02 · Grounding
The same question, grounded and ungrounded
In our generative AI development services, grounding means the answer is assembled from passages retrieved from your corpus, and every claim points back to its source. The fictional policy example below shows the difference between an answer with retrieval and one without it; a live build uses your own documents and reviewed questions.
Question
What approval is required for business travel costing more than $5,000?
Retrieved passages
- HighTravel policy · Section 4.2Used
- HighProcurement guide · Section 2.1Used
- LowExpense FAQ · General guidanceBelow threshold, dropped
Passages are ranked by similarity to the question and anything below the threshold is dropped before the model sees it.
Answer
The fictional travel policy requires finance-director approval for travel above $5,000, and the procurement guide requires a purchase request before booking.
- Travel policy · Section 4.2
- Procurement guide · Section 2.1
Verdict
Every claim traces to a retrieved passage
Faithfulness
Pass
Higher-cost travel may require manager or finance approval. Please check the current policy before booking.
Verdict
Plausible, with nothing to check it against
Faithfulness
Not grounded
Illustrative example using fictional policy text. No client information or production result is represented here.
03 · Choosing
RAG, fine-tuning or prompt engineering: which one your case needs
These techniques solve different problems and carry different costs. Choosing the simplest one that meets the need protects your budget and makes the system easier for your team to operate. Here is how we assess the options.
| Technique | When it is the right answer | What you need first | What it costs | Verdict |
|---|---|---|---|---|
| Prompt engineering | The knowledge is already in the model and you need the output in a particular shape. | A dozen worked examples of the output you want, and someone who can say which is right. | Cheapest to build and cheapest to change. Highest per call once volume arrives. | Start here. A useful share of cases stop here. |
| Retrieval (RAG) | The answer lives in your documents and those documents change without warning. | A corpus you can get at, with its permissions and versions intact. | $12k–$59k for a production system, then index and query costs each month. | The right answer for most of what gets called generative AI. |
| Long context | The whole relevant set fits in one window and rarely changes. | Documents small enough to send with every single question. | No build cost, and a token bill that scales with every question asked. | Fine for a pilot. Expensive as a product. |
| Fine-tuning | You need a house format, a register or a classification that prompting keeps getting wrong. | Thousands of labelled examples, and someone whose job it is to maintain them. | Quoted after dataset review, plus evaluation and retraining when the base model changes. | Rarely the answer to a knowledge problem. Buy it for format, not for facts. |
04 · Evaluation
How an answer is proved before it is released
Nothing our generative AI development services build ships on a demo. Every release runs against a question set built from your own history, and it goes out only after scoring at least 90% on two consecutive weekly runs, whatever the date in the plan says. The average across the systems we have shipped is 97 percent. The threshold is the floor, not the average.
Evaluation is split in two, because one number hides the failure. Retrieval is scored on whether the right passage came back at all. Generation is scored on whether the answer is supported by the passages that did come back. A system can look healthy overall while its retrieval quietly rots, and that is the failure that turns up six months in.
The set is yours. It lives in your repository next to the code and it runs on every prompt change, every model change and every corpus reload, so the gate is still there after we have gone.
E1
A question set from your own history
The questions come from what your team has actually asked, with the expected answers reviewed by a subject-matter expert.
E2
Groundedness, per answer
Every claim in an answer is checked back against the retrieved passages, and an unsupported claim scores as a failure even when it happens to be true.
E3
Retrieval scored on its own
Whether the right passage was returned at all is measured separately, so a drop in retrieval quality is visible before it reaches an answer.
E4
A gate on every change
Prompt, model, chunking and corpus changes all rerun the suite, and the run that clears the threshold is the run that ships.
E5
An adversarial pass
Questions the system should refuse, questions with no answer anywhere in the corpus, and questions built to pull an answer out of training data instead of a document.
05 · Stack
The generative AI stack we build on
Models
- Claude
- OpenAI
- Gemini
- Llama
- Mistral
Chosen per task on accuracy against your evaluation set first, then latency, then cost. The choice goes into the scope, and because the model sits behind our own interface, switching later is a configuration change and a rerun of the suite.
Orchestration
- LangGraph
- LlamaIndex
- Pydantic AI
- Plain state machines
We reach for a framework when it removes work and write the state machine ourselves when it does not. The framework never decides the architecture.
Retrieval and vector stores
- pgvector
- Pinecone
- Qdrant
- Weaviate
- Elasticsearch
If your data already lives in Postgres, pgvector usually wins on operational cost. A dedicated store earns its keep at scale, or where filtering does most of the work.
Data and ingest
- Airflow
- Dagster
- dbt
- Unstructured
- Docling
Chosen to fit the pipelines your data team already runs, because an ingest they cannot operate stops working the month after handover.
Serving and infrastructure
- AWS
- Azure
- Google Cloud
- Vercel
- Kubernetes
Your cloud account, your region, your network boundary. Your corpus does not move to us.
Evaluation and observability
- Ragas
- LangSmith
- Langfuse
- OpenTelemetry
- Grafana
Whatever the rest of the stack is, the evaluation suite and the traces run in your account and belong to you, so the gate survives the handover.
06 · Process
How our generative AI development services run, step by step
01
Scope
2 weeks
We read the corpus, sit with the people who answer these questions today, write down what the system may and may not say, and hand back a fixed quote with the name of the engineer who will build it.
02
Corpus and evaluation set
1–2 weeks
Ingest, permissions, versioning and precedence, and a question set built from your own history with reviewed answers, before a single response is generated.
03
Pilot
4–6 weeks
A working system on your own documents, answering real questions with citations, measured against the set and put in front of a small group of your people.
04
Build and integrate
6–10 weeks
Production hardening, the connectors into the places your people already work, access controls, monitoring, and the evaluation gate wired into your release process.
05
Run
Ongoing
Monitoring, evaluation reruns on every model and corpus change, drift response and tuning, either as a managed service or handed to your team with the runbooks.
The pod
Who works on a generative AI development project
Five roles on a typical engagement of our generative AI development services. The same people stay on it from scope through to handover, and you have their names before the contract is signed.
| Role | Owns | Allocation |
|---|---|---|
| AI engineer | Retrieval, prompts, the answer path and the cost per call | Set in scope |
| Data engineer | Ingest, chunking, permissions, versioning and the reload schedule | Set in scope |
| Evaluation lead | The question set, the scoring, the threshold and the regression suite | Set in scope |
| Delivery lead | Scope, the fixed quote, your owner's time and what stays out of scope | Set in scope |
| Platform engineer | Your cloud account, the serving path, monitoring and the release process | Set in scope |
07 · Industries
Generative AI development services by sector
- InsurancePolicy wordings, endorsements and claims notes, where the current version and the one in force at the time of loss are different documents.
- BankingProduct terms and procedure manuals that change by jurisdiction, and answers an adviser has to be able to defend later.
- HealthcareProtocols and payer rules where an unsourced answer is not an inconvenience, and access has to follow the same rules people already work under.
- ManufacturingWork instructions, maintenance histories and supplier specifications, most of it in PDFs and scans rather than in a system.
- Supply chainContracts, tariffs and supplier terms that decide whether an exception is chargeable, spread across systems that were never joined up.
- LogisticsFreight paperwork and customs rules, where the answer has to arrive while the shipment is still moving.
- RetailProduct data, policies and supplier terms, feeding both staff answers and customer-facing copy at a volume no one can review by hand.
- ConstructionDrawings, specifications and variations, where the question is usually which revision governs and who approved it.
08 · Work
Generative AI systems we have shipped
- Optician training · Role-play with an AI patientAverage role-play session: about 25 minutes down to under 15Opticians practise consultations on an AI patient. 85 to 92% of consultation stages are identified correctly in common scenarios, and feedback runs on a separate model to keep the running cost down.Read the case study
- Small business lending · Assistant for three roles20 to 30% fewer support tickets asking what happens next, in trial usageOne assistant answers merchants, funders and sales reps about their applications, each from their own role's view, and hands a frustrated conversation to a rep.Read the case study
- Service businesses · Conversational voice agentFirst response to a new lead: 2 to 6 hours down to under 1 minuteCalls, WhatsApp, SMS and email handled as one conversation per lead, with replies to messages in 10 to 20 seconds and a summary written for the rep who takes over.Read the case study
Read before you buy generative AI development services
All guides- GuideAI agent costWhat assistants, agents and automation cost to build and to run, by tier, and the factors that move the number.
- GuideEnterprise AI agentsEnterprise platforms compared by requirements, human approval, billing units and publicly available pricing.
- GuideAgentic AI companiesHow providers compare on published client evidence, human review and the questions worth asking before a shortlist becomes a contract.
FAQs
Questions buyers ask about generative AI development services
Straight answers on what generative AI development services cost, how long they take, how accurate they are, where your data goes and what happens after launch.
How much do generative AI development services cost?
A scoped pilot on one corpus is typically $8k–$35k. A production retrieval or knowledge-assistant system is usually $12k–$59k to build. Monthly model, index and support costs are modelled separately at your expected usage, and the build quote is fixed at the end of the two-week scope.
How long do generative AI development services take?
Allow two weeks for a scoped plan and fixed quote, one to two weeks to prepare the corpus and evaluation set, four to six weeks for a pilot, and about six to ten weeks for a production system. Clean, versioned documents with clear permissions move faster than content that still needs review and ownership decisions.
How do you reduce unsupported answers?
Three mechanisms. Answers are assembled from retrieved passages and cite them, so an unsupported claim is visible rather than hidden. Every claim is scored against those passages, and an unsupported one counts as a failure even when it is true. And a release scoring below 90% on the question set does not ship. The grounding section on this page shows the same question answered with and without retrieval.
Do we need RAG or fine-tuning?
Almost certainly retrieval. Fine-tuning changes how a model writes, not what it knows, so it is the right tool for a house format or a stubborn classification and the wrong one for a knowledge problem. The decision table on this page sets out what each needs and what each costs. If prompting alone solves your case, we will say so.
What happens to our documents, and where do they go?
They stay in your cloud account and your region. We build in your infrastructure, the index lives there, and the model is called over an API with no training on your data. Retention, residency and audit requirements are set during scope, before anything is ingested, because they change the architecture rather than the configuration.
How should we evaluate a generative AI development company?
Ask five things early: an indicative price band, an example trace showing retrieval and citations, the release threshold, who owns the code and evaluation sets, and who would do the work. Specific answers help your technical, security and procurement teams compare providers consistently.
Can generative AI development services work with the systems and content platforms we already run?
Yes, and that is a substantial part of the build. We connect to SharePoint, Confluence, Google Drive, your document management system, service desk and data platform, preserve the permissions each source already enforces, and surface answers inside tools your team already uses.
Which models do you use, and can we change later?
Claude, OpenAI, Gemini, Llama or Mistral, chosen per task on accuracy against your evaluation set first, then latency and cost. The choice is written into the scope. Because the model sits behind our own interface, changing it later is a configuration change plus a rerun of the evaluation suite, not a rebuild.
Is this generative AI consulting, or do you build it?
We build it, and the scope phase is the consulting. If you want the decision made and the roadmap written without a build attached, that engagement lives on the AI consulting page with its own fee and its own deliverables. Most people arriving here want the system, not the report.
What is the difference between a generative AI system and an AI agent?
A generative AI system answers, and its output is a response with sources attached. An agent acts: it plans over a goal, calls tools against your systems, writes results back and stops at a gate. That needs a tool list, permissions and escalation rules that an assistant does not, so it is built and priced differently. Agents are on the AI agent development page.
Bring one corpus and one question. We will scope the system that answers it.
Tell us where the answers live today, who has to be able to see them and what your people keep asking. Within two business days, the engineer who would build it replies with a scoped plan and the cost band that applies.
Reply within two business days · Every answer cites its passage · You speak with the engineer, not a sales team