Softellar

AI Integration

AI Integration Services for the Systems You Already Run

We connect Azure OpenAI, Amazon Bedrock and other models to the software you run today. One gateway, a private network path and a switch per segment, so the rest of your system does not change.

Discuss your integration

We reply within 24 hours

AI you can switch off

Assistive summary, by segment

AI

  • Customer portalUsers see: Summary from the modelOn
  • Back-office teamUsers see: Summary from the modelOn
  • Partner accountsUsers see: Static help textOff

Model version

Pinned

Changed here, after the test set passes

Release needed to switch

None

Flags are read at runtime

When the model is down

Model timed out. Static help shown, and an AI unavailable event logged for support.

Clutch rating, 12 reviews
5.0
Projects since 2021
70+

Companies we have built software for

  • Multi Radiance Medical
  • Loma Lux Laboratories
  • WellVu
  • Commercial Capital Group
  • Grand Auto
  • HostMy
  • Land Base
  • Auto Tuning

Four Ways a Model Joins Your System

Every call goes through one gateway, wherever it starts. Each row shows what we add, what stays untouched and the system where we built it.

Your system

Screen
Background workflow
Queue between services
Team tools
One gatewayThe only code that talks to the model: allow-list, limits, schema checks, fallback
Private modelAzure OpenAI, Amazon Bedrock or another model, on a private network path
  1. On a screen, on request

    What triggers it

    A user asks for a summary or an explanation.

    What we add

    A read endpoint that returns a cached, versioned answer or queues a short job.

    What stays as it is

    The submit path, which never waits for the model.

    In the lending platformLoan submissions never block on the model, and assistive text comes from a separate read endpoint
  2. Inside a background workflow

    What triggers it

    A job or an incident reaches a step that needs judgment.

    What we add

    A gateway call the workflow can skip, with a versioned contract for the answer.

    What stays as it is

    The workflow’s own steps, which run the same with or without AI.

    In the ETL monitoring platformThe workflow decides when to call the model, which contract version to use and when to skip it
  3. Between services, on a queue

    What triggers it

    A document or a message lands on the queue.

    What we add

    A recognition step in the flow that reads from the same queue.

    What stays as it is

    The services on either side, which keep talking over the bus.

    In the invoice recognition SaaSRecognition runs inside the document flow, with services talking over Azure Service Bus
  4. Into the tools your team uses

    What triggers it

    A result is ready for a person to act on.

    What we add

    Messages in Slack and pull requests in GitHub, written for review.

    What stays as it is

    Code review and the on-call routine your team already follows.

    In the ETL monitoring platformRoot-cause analysis lands in Slack, and fixes arrive as normal GitHub diffs

Models Inside Running Systems

Three products where we wired a model into the flow, then the integration API and data work that systems like these depend on.

United States

AI lending platform for a US private credit marketplace

Outcomes

  • Faster cycle times on cleared applications
  • Fewer rework loops on document packages
  • Higher borrower completion rates
  • .NET
  • Angular
  • Azure
  • Azure OpenAI

The Gateway, Layer by Layer

Your code never calls a model directly. Each layer below owns one job, so a limit, a key or a provider changes in one place.

  1. Your code

    Calls one interface, so the provider can change behind it.

    • IChatClient from Microsoft.Extensions.AI
    • Feature flags read at runtime
    Any .NET library with an LLM client can implement IChatClient
  2. .NET gateway

    Decides what may leave and what comes back.

    • Allow-listed fields and a versioned contract
    • Timeouts, retry budgets and a circuit breaker
    The same layer, in the ETL monitoring platformA .NET gateway with strict timeouts, retry budgets and circuit breakers
  3. Azure API Management

    Shares the model between applications and teams.

    • Token limits per consumer
    • Usage metrics and one place for keys
    Limits token use per key, by rate or quota
  4. Model endpoint

    Answers only on a private network path.

    • Azure OpenAI behind a private endpoint
    • Amazon Bedrock over AWS PrivateLink
    The same layer, in the ETL monitoring platformThe model in the client’s own cloud tenant, with network rules that deny unapproved hosts

Side laneMCP server

The other direction: your system’s operations offered to assistants and agents.

  • Lookups and actions you choose to expose
  • The sign-in and roles your system already enforces
An open-source standard for connecting AI applications to external systems

Where should the gateway sit in your system? Software architecture covers the structure around it.

Where AI sits in the architecture

Clients on Systems We Joined

An API for systems already in use, a migration done while everything kept running and CRM and ERP data connected, each linked to its project.

What AI Integration Covers

AI integration is adding a model, such as Azure OpenAI or a model on Amazon Bedrock, to software that already runs. We find where it plugs in, put one gateway between your code and the model, keep the traffic on a private network path and switch it on in steps. The screens, services and data around it stay as they are.

The goal

A model inside the systems you already run, switched on one segment at a time and off without a release.

Is Your System Ready for a Model?

Four questions we ask before connecting anything. A no is not a stop, it points to the work that comes first.

  1. Is there an API or a queue to attach to?

    A model joins through a clean interface. Logic buried in a desktop app or a stored procedure has to come out first.

    If not, start withLegacy Application Modernization
  2. Can the model read clean data it is allowed to see?

    Fields scattered across exports and spreadsheets give the model nothing reliable to read, and no way to limit what it reads.

    If not, start withData Architecture
  3. Is there a cloud tenant or account for a private model?

    A private endpoint needs an Azure tenant or an AWS account that your security team controls.

    If not, start withCloud Migration
  4. Has anyone decided where the model sits?

    Calls scattered through the code are hard to switch off, test or move to another provider.

    If not, start withSoftware Architecture

Not sure how your system scores? The fit check on the AI transformation page takes a few questions.

Take the AI fit check

Where Does a Model Plug In?

Tell us what the system does today and where AI should help. We will say which of the four ways fits and what has to be ready first.

We reply within 24 hours and can sign an NDA first.

Have files or a long brief? Use the full form

Azure OpenAI or Amazon Bedrock?

Both give you a private, governed model. Each row shows how the platform handles it and what we build on top either way.

Already settled on a cloud? See what else we build there.

Which Model, and Where It Runs

Azure OpenAI and Amazon Bedrock cover most systems we connect. The rest fit specific clouds, prototypes or data that may not leave your servers.

Around the model

  • .NET
  • Temporal
  • RabbitMQ
  • Azure Service Bus
  • OpenTelemetry
  • Application Insights
  • Azure Key Vault
  • Entra ID

Switched On One Segment at a Time

A model reaches users in steps, and each step can be undone without a release. This is the order we follow.

  1. Shipped behind a flag

    New AI paths go to production switched off, so the release and the switch-on are separate decisions.

    In the lending platformThe first bureau connector shipped behind feature flags
  2. On for one segment

    One group of users gets it first. The people who run that group can switch it off again without a release.

    In the lending platformLending officers turn model assistance off per segment, without redeploying
  3. Version pinned

    The model version is set in configuration and changed from an admin screen, after the test set passes.

    In the lending platformAdmin controls for model versions
  4. Failures recorded

    A slow or failed call shows the non-AI path and logs an event, so support knows before users report it.

    In the lending platformFailures degrade to static help and log a structured event for support
  5. Watched in production

    Traces and dashboards for every call, and a runbook for whoever is on call.

    In the lending platformOpenTelemetry traces, Application Insights dashboards and on-call runbooks

US Office, European Engineering

Headquartered in Warsaw, Poland, with a client office in New York. A European engineering team that works in your hours, with a US point of contact.

Headquarters

Warsaw, Poland

Engineering and delivery, and the main contact point for clients in Europe.

Client office

New York, United States

Kickoffs, contracts and account management for North American clients.

Shared hours with North America

Clients in Europe share the full working day with the team.

8 AM to 12 PM New York time, 2 PM to 6 PM in Warsaw.

Rankings for Our Engineers

Independent rankings for the team that connects models to client systems, each next to the service it relates to.

Related reading

Read Before You Connect

Guides from our team on application modernization, web application security and CI/CD.

All Insights

Frequently Asked Questions

Questions teams ask before a model touches a system that is already in production.

Have a different question? Talk to our team

Ready to Add AI?

Tell us the system, the cloud it runs on and where a model should help.

We reply within 24 hours, and we can sign an NDA before you share details.

Talk to the people who will scope your project