Custom AI development
Krikir handles custom AI development for startups and growing firms, tested on real data before launch.
Fixed quote after one callReply within one working day
Example: one summary feature added to a support tool
What custom AI development covers
Custom AI development means building AI software around one task in your business, on a provider's model, with your data, rules and screens. It suits a task close to how you make money, and it is the wrong choice when a product already does the job or nobody will own it after launch.
One task, one measure
We name the task, the person who does it now and what a good result looks like to them.
AI proof of concept
Working code on your real data, with cost per task measured and a pass mark agreed first. Stop is a valid result.
Generative AI or integration
A new generative AI app, a copilot inside your product, or LLM integration into software you already run.
A person stays in charge
Someone checks the output before it counts, and a non-AI path keeps the product working when the model cannot.
Running it live
Monitoring, cost limits and model changes, the work also known as MLOps or LLMOps.
Buy, add, build or train?
Pick the cheapest option that passes your own test, and move down a row only when the one above fails.
The task is common, a product already does it, and it gives you no edge.
You carry the risk that its terms or features change.
Your product has users and data, and it needs one new ability such as search or summaries, with no rebuild.
Poor data underneath shows up in the answers.
The task is close to how you make money, several systems are involved and no product fits.
It needs a measure of success first, and a named owner for as long as it runs.
Prompts and retrieval have fallen short on a measured test, and you hold a large stock of good labelled examples.
New facts belong in retrieval. Tuned models end with their base model.
A sample test report
This sample was written for this page and is not from a client. It is the sheet you read before deciding whether a build goes live.
Talk it through on a call
We look at what you have now and what you want it to do.
You hear how we would approach it and who would work on it.
You get a fixed quote in writing after the call, with nothing to sign until you are ready.
How a custom AI project runs
Founder and CEO Krishnkant Patidar leads the project with AI and ML developer Deepak Patidar, who brings ten years in this field. Earlier client systems stay under NDA, so this page shows the method. Test results reach you in writing, and call times show in your own time zone.
Choose one task
One call names the task and the measure. If nobody can mark a result right or wrong, the project is not ready.
Build the test set
Real past cases go in, including ones where the right answer is "I don't know". This is the most valuable file in the project.
Prove it on real data
The proof of concept answers four questions: does it pass, what does one task cost, what happens on the bad cases, and who runs it.
Build with limits
Permissions apply before the model sees anything, output is checked in code, and a non-AI path keeps working when the provider is down.
Launch in stages
It runs unseen on real input first, then with a small share of users, then everyone, with a way back at each stage.
Watch and maintain
We track quality on live samples, cost per task and response time. Every failure becomes a new test.
What keeps it safe and affordable
Most stalled pilots stall on the work around the model.
Cost per task is measured
We count finished tasks, not API calls: every model call, retry and review. Limits apply per user and per task, and the provider bills your account directly.
One layer for model calls
All model calls sit in one place and use a dated model version, so changing provider is one planned change and not a rebuild.
Limits sit outside the prompt
The UK's NCSC said in December 2025 that prompt injection, hidden text that tricks a model, "may never be totally mitigated". So keys hold only the rights the task needs.
Drift gets a named cause
Your data changes, your users change, or the model changes with no edit on your side. Tests on samples of live work show which one it is.
What you keep
Where data goes and when models change
Vendor terms and dates were read on 7 November 2026 and can change.
What is sent
Only the fields the task needs, plus whatever retrieval returns. Data handling is agreed in writing before any real data is used.
Who keeps it
OpenAI keeps API abuse logs up to 30 days. Anthropic deletes API data within 30 days, and flagged content can stay up to 2 years. Google's paid Gemini API keeps it 55 days.
When the model changes
OpenAI states at least six months for generally available models and less for some others, Anthropic at least 60 days, and Google no fixed period. From 1 December 2026, Google routes requests for two older Gemini models to other models automatically, so behaviour can change with no edit on your side.
Who owns the output
Provider terms give output to the customer, but owning output differs from holding copyright in it. The US Copyright Office said in January 2025 that prompts alone do not make a user the author. Your lawyer answers how this applies to you.
Other pages for this work
Questions before you build custom AI
Ownership, running cost, stalled pilots and data.
Book a 30-minute call
In custom AI development with Krikir, the code, prompts, test set and scoring scripts are transferred to you on final payment, under the contract. Model, cloud and monitoring accounts sit in your name. The provider's model itself stays the provider's, and third-party licences stay with their owners.
The running cost of custom AI depends on the model calls per task, their length, the model and your volume. We measure cost per task in the proof of concept, and the provider bills your account directly.
Often a product will do, and we say so. Custom AI earns its place when the task depends on your own rules, data or screens and no product passes your test.
AI proofs of concept usually stall on the work around the model: clean demo data, no owner, and no measure of cost or bad cases. RAND's 2024 interviews with practitioners name the wrong problem, thin data and weak infrastructure among the causes.
Custom AI sends only the fields a task needs to your chosen model vendor, and OpenAI, Anthropic and Google's paid Gemini API do not train on it by default. Data handling is agreed in writing before any real data is used.
A custom AI system is retested on its test set before any model switch. All model calls sit in one layer, so the change stays contained, and who runs the retest is agreed in writing before launch.
Let's talk about your next project
Book a 30-minute call or send a brief. A founder replies within one working day.