How to Connect Business Data to LLMs: Lessons from Building the InfuseNet AI Data Platform
Most of the work in an AI data platform isn’t the model. It’s getting data in safely, preparing it properly and letting non-engineers build on it
Connecting business data to large language models means three jobs. First, pull data out of the places it already lives: documents, databases, cloud storage and APIs. Second, clean it into a consistent shape a model can use. Third, give people a safe way to build AI tools on top of it. Picking the model is the smallest part. Most of the effort goes into data access, preparation and security.
The InfuseNet platform is a good example of that balance.
What InfuseNet set out to do
InfuseNet is an AI data platform. Teams use it to import data from text, images, documents, databases, cloud storage and APIs, more than 80 sources in total, and connect it to large language models through a visual drag-and-drop interface called Flow.
With it, a business can:
- Link LLMs, templates, media models and extraction tools in one workflow
- Prepare its own data for fine-tuning models
- Build a private, ChatGPT-style assistant on its own knowledge base
- Watch performance and trends with built-in monitoring
- Run everything on its own private cloud or on the hosted InfuseNet Cloud
The goal was to let business teams build useful AI workflows without writing code each time, and without sending sensitive data where it shouldn’t go.
Challenge 1: Models from different providers work differently
The platform needed to support foundation models from OpenAI, Anthropic and Google. Each provider has its own API, request format, limits and customization options. A Flow that works with one model shouldn’t break when a user switches to another.
The fix is a model abstraction layer: one internal interface the Flow canvas talks to, with an adapter per provider behind it. Adding a new model then means writing one adapter, not rebuilding workflows. It’s also the main defence against lock-in. Model quality and pricing change every few months, and a platform tied to one provider ages fast.
Challenge 2: A drag-and-drop interface that stays useful
No-code tools have a well-known trap. Too simple, and they can’t do real work. Too powerful, and they become a programming language with extra steps.
The InfuseNet team handled this by designing and testing the Flow interface iteratively with users. That’s the right instinct. The only way to find the line between simple and capable is to watch real users build real workflows and see where they get stuck. A few principles usually hold:
- Each block on the canvas should do one clear job (import, extract, prompt, filter, output).
- Users should see sample output at every step, not only at the end.
- Advanced settings should exist but stay out of the way until someone needs them.
Challenge 3: Preparing data for fine-tuning without leaking it
This was the hardest part, and the most important. Data arriving from 80+ sources comes in every format imaginable: scanned PDFs, spreadsheets, database rows, image files, API responses. Before any of it can fine-tune a model or feed a knowledge-base assistant, it has to be cleaned, de-duplicated and converted into a consistent structure.
Get Good Opportunity’s stories in your inbox
Join Medium for free to get updates from this writer.
InfuseNet built a dedicated data preprocessing framework for this, so data is consistent and ready for fine-tuning wherever it came from.
Security sat alongside it:
- Encryption to protect data in storage and in transit
- Access controls so people only see the data they’re allowed to
- A self-hosted deployment option for companies that can’t let data leave their own environment
Access control is the one teams most often get wrong. If a sales rep asks the company assistant a question, the answer must not draw on HR files they couldn’t open themselves. That rule has to be built into retrieval from day one. It’s very hard to add later.
Fine-tuning or a knowledge-base assistant?
InfuseNet supports both. They solve different problems, and it’s worth knowing which one you need.
Fine-tuning changes how a model behaves: its tone, its format, the specialist terms it uses. Use it when you need consistent output in a particular style or domain.
A knowledge-base assistant keeps the model as it is and gives it the right documents at the moment of each question. Use it when answers have to reflect current, specific company information that changes often.
Many businesses start with the knowledge-base approach because it’s faster and easier to keep up to date. Fine-tuning makes sense once you know exactly what behaviour you want to change.
The technology behind it
The platform was built with Python and JavaScript. NLP work uses spaCy, NLTK, Rasa, Dialogflow and Botpress, and the machine learning frameworks are TensorFlow and PyTorch. It runs on AWS, Google Cloud Platform and Microsoft Azure, with integrations for external services such as MySQL and Amazon AWS.
Supporting all three major clouds fits the deployment goal. A company that wants InfuseNet self-hosted can run it on the cloud it already uses, instead of opening a new account just for AI.
Lessons for anyone building an AI data platform
Budget most of the effort for data, not models. Connectors, cleaning, de-duplication and permissions are where the real work goes.
Abstract the model layer early. You’ll switch or add models more often than you expect.
Build permissions into retrieval. An assistant that ignores access rules becomes a data leak with a friendly interface.
Offer a self-hosted option if you sell to regulated industries. For many buyers in finance, healthcare and government, keeping data in their own environment decides whether they buy at all.
Test the no-code interface with real users, repeatedly. The InfuseNet team’s iterative, user-tested approach to Flow is what keeps a visual builder usable as it grows.
If you’re planning something similar, the full InfuseNet case study covers the platform’s features and stack. The model integration and data preparation work is core AI software development. The deployment choices, self-hosted or hosted cloud, are worth reviewing early with a cloud consulting team, because they shape the security design from the start.
