SAP to open-stack ERP migration
Manufacturer, 500+ employees. Phased migration off SAP ERP onto an open stack with custom modules — run in stages while the plant kept working.
- 0 downtime
- 8 mo migration
- −40% on licences
LLM integration into a product that already has users
The hard part is not calling the API. It is adding a probabilistic component to a deterministic product without breaking the contract you already have with your customers.
What we can show you
Overview
A language model inside a live product
Adding a probabilistic component to a deterministic product is the hard part — not the API call. That is what this service covers.
What it takes to put a language model inside a live product
Adding an LLM to a product that already has users is not an API call. It is adding a probabilistic component to a deterministic system without breaking the promises the product already makes — about permissions, about latency, about what happens when something fails.
That means grounding data with permissions preserved, versioned prompts, an evaluation set in CI, budgets enforced per request, a fallback path and a feature flag that switches the whole thing off without a deploy.
When integration is the right project
When you have a product, users and data, and one clear place where a model would help. If the data is not accessible yet, the first project is a data project — and pretending otherwise wastes a quarter.
Signals your product is ready for an LLM feature
- there is one obvious task users repeat that a model could shorten;
- the data the model needs already exists and has an owner;
- permissions are enforced server-side today, not in the interface;
- somebody can define what a good answer looks like;
- you can live with a feature flag that turns it off if quality drops.
Price
From $10,900
Timeline
3–5 weeks
Payment
50/50 by milestone
Comparison
Integrating an LLM: build it in or bolt it on
The difference between a feature that survives its second month and a demo that quietly gets switched off.
Criterion
Quick bolt-onIntegrated properly
Integrated properlyBolted on quickly
Grounding
Prompt with pasted context
Indexed data with permissions preserved
Quality control
Someone eyeballs it
Labelled set scored in CI with regression gates
Failure behaviour
Feature errors, page breaks
Timeout, fallback, feature flag off without deploy
Cost
Whatever the month brings
Budget per request, caching, model routing
Switching provider
Rewrite the feature
Config change plus an eval rerun
When it wins
A one-week experiment
Anything users depend on
Anatomy
What integration actually involves
Six layers. Skipping any of them is how a working demo becomes an incident three weeks after launch.
The six layers we build
Grounding layer. Your data, chunked and indexed with permissions preserved, so the model cannot answer from documents this user may not see.
Prompt and schema management. Versioned, tested, deployable independently of the app release.
Evaluation. A labelled set from your real cases, scored in CI, with regression gates before merge.
Cost and latency control. Budgets per request, caching, model routing — small model first, escalate only when needed.
Failure design. Timeouts, fallbacks, and a feature flag that turns the whole thing off without a deploy.
Observability. Traces per request, quality and cost dashboards, alerting on drift.
Feature in your product
From $10,900
Timeline
3–5 weeks
Ongoing tuning
Core subscription $2,990/month
Providers
Model choice, honestly
Vendor lock-in in this space is a self-inflicted wound: the integration should make switching a configuration change.
How we choose
Frontier hosted models where quality decides. Smaller or open models where cost, latency or data residency decides. Local models when the data cannot leave your perimeter — with the quality trade-off stated plainly rather than hidden.
Whatever we pick, orchestration and model stay swappable, and switching means re-running the evaluation set rather than rewriting the feature.
Why us
Why teams choose us for this
Four things we can prove, not four adjectives.
Grounding with permissions intact
Your data indexed so the model cannot answer from a document this user may not open — the failure mode that ends pilots.
Quality measured, not asserted
A labelled set from your cases, scored in CI, with regression gates before merge. The number is visible before launch and tracked after.
Provider stays swappable
Orchestration and model are abstracted, so switching is a config change plus an eval rerun rather than a rewrite.
Failure designed in
Timeouts, fallbacks and a feature flag that disables the feature without a deploy. The product keeps working when the model does not.
Industries
What we buildfor your industry
LLM features land first where text is the bottleneck: support, sales operations, legal review and internal knowledge.
- IT companies and digital agencies
- Insurers and brokers
- Lenders and microfinance
- E-commerce
- Wholesale and distribution
- Manufacturing
- Freight and forwarding
- Clinics and medical centres
- Dental
- Car dealers and service centres
- EdTech
- Legal and consulting firms
- Equipment rental
- Real estate agencies
- Property developers
- Travel and events
- Beauty and wellness
- Construction and renovation
FAQ
Integration questions
What engineering leads ask before letting a model near their product.
Can you work with our existing codebase?
Yes — that is the entire point of this service. We start with a short audit of the data model and the permission system, because those decide what grounding is even possible.
How do you keep costs predictable?
Budgets in code, caching, model routing and a monthly cap. You get a cost-per-request dashboard rather than a surprise invoice.
What if the model provider has an outage?
Fallback provider or a degraded mode behind a feature flag. The product keeps working without the feature — that is designed in, not bolted on.
How do you measure quality?
A labelled set built from your cases, scored in CI, with the number visible before launch and tracked after it.
Will our data train someone else's model?
No: zero-retention configurations plus contractual terms saying the same thing.
Who owns the prompts and the evaluation set?
You do. They ship with the code, in your repository, and they are as much of the deliverable as the feature itself.
Do you offer RAG development services on their own?
Yes — RAG development services are the most common single request here: retrieval over your own documents, grounded answers with citations, and an evaluation harness that proves the answers stayed correct after the next change.
2–4×
faster on the code-writing share
in 3 weeks
from idea to MVP
230+
projects delivered
17 years
of engineering experience
Work
Work we can show you
SMARTWILL
Digital inheritance service (Web3)
AIRDEVELOPPA
Indoor air quality finder driven by sensor data
CAR5
Car-sharing app
APP.DENTAL
Cloud CRM for dental clinics
NEXORA
Warehouse and logistics management platform
As-built docs
Digital as-built documentation
EX211
Telegram Mini App for currency exchange
SMARTWILL
Digital inheritance service (Web3)
Bring one feature, not a strategy
Pick the single place in your product where a language model would obviously help. We will tell you what grounding it needs, what it costs to run, and what could go wrong.
If your data is not accessible yet, the first project is a data project — and we will say so.