LLM integration for products and business processes
NDDev integrates LLMs into companies’ products and workflows: generated descriptions, document summaries, request classification, field extraction and assistants inside the interface. We connect models from several providers with a fallback route and track the quality and price of every request. For Citilink we built a prototype in which AI searched for products and wrote marketplace descriptions.
Who it is for
Product teams adding language model features to their service, and companies whose staff already paste text into chatbots by hand and want this built into their working systems.
- Several providers and a fallback
- Answers in a strict format
- Control over request costs
- Tracing in Langfuse
What your product gets
We connect the model through a LiteLLM gateway: your code calls one API, and behind it sit models from several providers. If the main provider returns an error or responds too slowly, the request goes to a fallback model. Simple tasks go to fast, low-cost models and complex ones to stronger models.
- selection of features where the model saves users time
- a LiteLLM gateway with fallback providers
- versioned prompts and test example sets
- a strict answer format with schema validation
- masking of personal data before sending
- limits, caching and spend tracking
- request tracing in Langfuse and OpenTelemetry
The model’s answer arrives in a strict format: JSON against a schema your code validates before writing to the database. Prompts are versioned, and every version runs on a set of examples before release. Requests, answers, latency and cost are visible in Langfuse, so any odd output can be found and examined.
Before connecting an external provider we agree which data may be sent to it, what we mask and how long it keeps the data. When data has to stay with you, the same integration moves to a model deployed on your own server.
Shipping an LLM feature
Every feature is measured on your examples before users see it.
Choosing scenarios
We pick the features where the model saves users time and describe a good result for each.
Prototype and benchmark
We build a prototype and compare models on your examples for quality, speed and price.
Integration into the product
We build the calls into your code and set up the gateway, fallback route, limits and tracing.
Release and monitoring
We switch the feature on for users, watch quality, spend and complaints and improve the prompts from the data.
Work in this area
Citilink
A 2024 prototype for a major electronics retailer: AI finds products from a shopper's query, helps new staff get up to speed and drafts product descriptions for marketplaces.
Questions about LLM integration
Which provider should we choose?
We choose by benchmark on your tasks: quality, speed, price per request, processing region and data retention terms. The gateway separates your product code from the provider, so switching models comes down to a setting in one place.
How do we keep model costs under control?
We set limits per user and per feature, cache repeated requests and send simple tasks to small models. The cost of each request is visible in Langfuse, so you see the spend for every feature.
What if a provider changes its model or price?
The test example set shows how a new model handles your tasks before you switch. While we choose a replacement, the fallback route keeps the feature running.
May we send customer data to the model?
We settle this before work starts. An external provider does receive the data, so we agree the permitted data classes, masking, region and retention period. If sending is impossible, we use a model on your servers.
Where have you integrated LLMs before?
In a prototype for Citilink (2024) AI searched for products, helped new employees settle in and wrote marketplace descriptions. Another example is Declaro, a product we build as a technology partner, which assembles a draft customs declaration from documents with links to the sources.
How much does LLM integration cost?
The starting price sits in the price block on this page. The total comes from the number of features, the complexity of the integration and the size of the test sets. Model provider fees are billed at cost; we fix the price and timeline in the specification.
from $2,300
- Free warranty
- 20 days
Prices are starting prices for information only and are not a public offer. The final price is fixed in the specification.
Language models and company knowledge
Related
Tell us which feature you want to build on an LLM
Describe your product, its users and the task for the model. We will suggest a scenario and models to compare, and give you a price.
We reply within 24 hours