LLM development
LLM development services for applications built on language models
Our LLM development services build custom applications on large language models: we choose between hosted APIs and open-weight models, decide whether prompting, retrieval or fine-tuning fits, make outputs reliably structured, and set latency, cost and privacy limits. We do not train foundation models from scratch. A first LLM feature is indicatively ₹60,000–2,00,000 ($2,200–7,500) over 2–4 weeks.
- First LLM feature
- ₹60,000–2,00,000 indicative
- Models
- Hosted APIs or open-weight
- Not offered
- Training a foundation model
- Larger products
- Custom quote after discovery
Updated · By the Codebrix team
What our LLM development services cover, and what they do not
'LLM development' gets used for two very different jobs. One is training a new foundation model, which takes research teams, large datasets and compute budgets that only a handful of companies spend. The other is building a dependable application on top of an existing model. We do the second.
| We do | We do not do |
|---|---|
| Pick and compare models on your own task set | Pre-train a model from scratch |
| Prompting, retrieval and tool use around a model | Claim a 'proprietary LLM' that is a prompt on a public API |
| Fine-tune hosted or open-weight models when tests justify it | Promise that fine-tuning will add facts the model can recall reliably |
| Deploy open-weight models on your cloud when data rules require it | Run GPU clusters on our own infrastructure for you |
Typical builds: a contract clause extractor that outputs structured fields, a drafting tool trained on your house style, a classification service routing thousands of tickets a day, or a domain assistant your customers pay for. Where the work centres on documents, see RAG development; where it takes actions, see AI agent development.
Hosted API or open-weight model
This is the first real decision, and it is about data rules and volume as much as quality.
| Factor | Hosted API (OpenAI, Anthropic, Google) | Open-weight model you host |
|---|---|---|
| Quality on hard reasoning | Usually the strongest models available | Varies; strong on narrower tasks, especially after tuning |
| Where data goes | To the provider, under their terms and your plan's settings | Stays in your cloud account and region |
| Cost shape | Per token, near zero at low volume | Fixed GPU cost whether busy or idle |
| Operations | Provider handles scaling and uptime | You own scaling, monitoring and updates |
| Licence | Provider's commercial terms | Each model's licence, which can restrict use; check before building |
For self-hosting we typically use vLLM, an open-source library for LLM inference and serving that provides an OpenAI-compatible API server (vLLM docs). That compatibility means the application code can move between a hosted API and your own server with little change.
Fine-tuning: when it helps, and when prompting or RAG is enough
Fine-tuning trains a model further on your examples. Google's Vertex AI documentation advises starting with prompting, suggests a dataset of around 100 examples or more, and recommends supervised fine-tuning for tasks such as classification, sentiment analysis and entity extraction where the output is easy to define (Google Cloud). OpenAI's model optimisation guide also starts with evals and prompting before tuning.
- Fine-tune when: you have hundreds of checked examples, the output format or tone must be very consistent, or a smaller tuned model could replace a larger one at high volume.
- Do not fine-tune when: the model needs current facts (use retrieval), you have only a few dozen examples (use them in the prompt), or the task changes every month.
- Always: compare the tuned model against the best prompt on the same evaluation set before switching.
Structured outputs your code can trust
An LLM inside a product usually has to return data, not prose: fields from an invoice, a category label, a JSON payload for another service. OpenAI's Structured Outputs constrains completed, non-refusal output to the JSON Schema you supply, and Anthropic's structured outputs constrain Claude's responses to a schema, including a strict mode for tool calls (OpenAI, Anthropic).
A valid shape is not a correct answer. OpenAI's guide itself notes that applications still need to handle refusals, incomplete responses and errors, and validate business rules the schema does not express. So we add:
- Business-rule checks after parsing: totals that add up, dates in range, IDs that exist in your database.
- A confidence or 'unsure' field, with low-confidence items routed to a person.
- Field-level accuracy tracked on the evaluation set, not only pass or fail per document.
Latency and cost trade-offs
Bigger models are usually slower and more expensive per token. Most products do not need the biggest model on every request.
| Lever | Effect | Cost |
|---|---|---|
| Route easy requests to a small model | Lower cost and latency on most traffic | A routing step that must itself be tested |
| Stream the response | Users see text start quickly | Harder to validate structured output mid-stream |
| Cache repeated prompt sections | Cheaper, faster repeated calls | Prompts must be ordered so the fixed part comes first |
| Batch non-urgent jobs overnight | Lower price for work nobody is waiting on | Results arrive later |
| Shorter outputs and tighter prompts | Fewer tokens per call | Needs re-testing for quality |
We report p95 latency and cost per request next to quality scores in every weekly demo, so a model change is judged on all three. Current per-token prices are on each provider's pricing page; we estimate your monthly spend from expected volume during scoping.
Data privacy and self-hosting considerations
- Redact before sending: a pre-processing step replaces names, phone numbers and ID numbers with placeholders and restores them in the output where needed.
- Read the provider terms together: whether your plan allows inputs to be used for training, how long requests are kept, and which settings change that.
- Choose the cloud region for hosting, logs and any self-hosted model.
- Keep prompts, outputs and training data in your own accounts, with a retention period you set.
- For self-hosted models, plan for GPU capacity at peak load, security patches, model upgrades and someone on call.
Self-hosting makes sense when data must not leave your infrastructure, or when volume is steady and high enough that a fixed GPU cost beats per-token billing. Below that, a hosted API with careful data handling is usually cheaper and simpler. We build with the data protection law that applies to you in mind; your legal adviser signs off.
Indicatively, a first LLM feature is ₹60,000–2,00,000 ($2,200–7,500) over 2–4 weeks; products with fine-tuning, self-hosting or their own interface are custom-quoted. See AI development for the full product process, or hire AI developers to work inside your team.
Frequently asked questions
Can you build us our own LLM like ChatGPT?
We can build an application that looks and works like a private assistant for your business, running on a hosted or open-weight model, with your data, permissions and branding. We do not train a new foundation model from scratch; that takes research teams and compute budgets few companies need. What usually matters is the application around the model.
How much do LLM development services cost?
With Codebrix, a first LLM feature is indicatively ₹60,000–2,00,000 ($2,200–7,500) and 2–4 weeks. Products that need fine-tuning, a self-hosted model or their own interface are custom-quoted after discovery. Model usage or GPU hosting is billed separately to your accounts, and we estimate it from your expected volume.
Should we fine-tune a model or use RAG?
Use RAG when answers depend on facts in your documents that change or need citations. Consider fine-tuning when you need a consistent format, tone or classification and have hundreds of checked examples. Start with prompting for both; many projects never need tuning once the prompt and retrieval are right.
Can we use an open-source LLM so our data stays in India?
Yes. An open-weight model can run on GPU servers in an Indian cloud region in your own account, so prompts and outputs stay there. Check the model's licence, and expect a fixed monthly GPU cost plus operational work. We compare its quality against hosted models on your task set before you commit.
How do you choose which LLM to use?
We run your evaluation set, built from real examples, on a shortlist of hosted and open-weight models, then compare accuracy, p95 latency and cost per request. The model is named in one place in the code, so switching when prices or quality change is a configuration change and a test run.
What does an LLM development company need from us?
Real examples of the task with correct outputs, a person who can judge quality, access to the data involved, a clear note of anything that must not leave your systems, and model provider or cloud accounts in your company's name. Everything we build, including prompts and training data, is committed to your repository.
Sources
Related
AI development
An AI development company for products that have to work after the demo
AI development company that scopes, builds, tests and runs AI products end to end: data, model choice, RAG or agents, evals, deployment and cost limits.
Read moreRAG development
RAG development services for answers your team can check
RAG development services: ingestion, chunking, hybrid search, reranking and permission-aware retrieval, with retrieval and answers measured separately.
Read moreAI agents
AI agent development company for agents that act inside your systems
AI agent development company building agents that use your CRM, email and APIs, with scoped permissions, human approval on risky steps and full traces.
Read moreAI integration
AI integration services for the software you already run
AI integration services that add LLM assistants, document search, summaries and classification to your product, tested on your data with cost limits set.
Read moreHire developers
Hire AI developers who build, test and ship LLM features
Hire AI developers who build LLM features with retrieval, agents and evals in Python or TypeScript. Paid first milestone, your keys, rate after a call.
Read moreAI development, US
Outsource AI development without losing control of your data, costs or IP
Outsource AI development to an India-based team and keep control: model accounts in your name, evals as acceptance tests, cost limits, IP assigned to you.
Read more