Preplo: read the case study
Turns cooking videos into recipes you can cook
We build AI products, and add AI to the ones you have
Nano Soft builds new AI products and adds AI to existing ones: agents, assistants, search over your own data, extraction and automation. We plan for quality, cost and privacy from the first week, because we run AI in our own products, including Preplo, Just Add and Kraterion.
At a glance
- AI in our apps
- Preplo, Just Add, Kraterion, Perqly
- Recognition
- Winner, RevenueCat Shipyard: Creator Contest 2026
- MCP servers
- Kraterion, Just Add, Inkray
- Experience
- 10+ years in software
How we help
Two ways we help with AI
Start from a new idea, or from the product you already run.
Build a new AI product
- For
- Founders and teams with an idea that has AI at its core
- You get
- A prototype tested on real examples
- A written plan with model costs worked out
- A launched product with accounts and payments
- Quality checks that run before every release
- First step
- Send us the idea and a few examples of real inputs
Add AI to your product
- For
- Product and engineering teams adding AI to a product people already use
- You get
- One task picked where AI clearly helps
- A test set built from your real cases
- The feature built into your product
- Cost, speed and quality tracked after launch
- First step
- Tell us the task you want AI to take on
What we build
What we build with AI
These are the AI features we design, build and run. We pick the model for each task by testing it, not from a shopping list.
AI agents that take actions
Agents that call tools, read your data and finish multi-step tasks, with every step logged.
Assistants inside your product
Chat and assistants that answer from your product's own data and hand off to people when needed.
Search over your own data
Retrieval that finds the right passage and cites it, so every answer can be checked.
Extraction from video and documents
Structured data from messy inputs, the way Preplo turns cooking videos into recipes.
Workflow automation
Repetitive steps handed to AI with clear rules, review points and a record of what ran.
MCP servers
Let assistants such as Claude and ChatGPT use your product safely, with permissions you control.
Evaluation and monitoring
Test sets, quality checks and cost tracking that run before every release and after launch.
Approach
How we approach AI work
Start with one job
We pick one task worth doing and make it reliable before adding a second.
Measure with real cases
We build a test set from your real inputs and check quality before and after every change.
Budget cost and speed early
We price each model call in the plan, then design caching and limits to stay inside it.
Keep checks where mistakes cost
Where an error is expensive, a person reviews it or plain code checks it, as in Just Add.
Proof
Where we have done this
AI runs in most of what we ship. These are the clearest examples.
Kraterion: read the case study
A runtime for AI agents you can audit
Inkray
A publishing platform where writers own and earn from their work. Articles are stored on Walrus and registered on Sui. There is a web app, an iOS reading app and an MCP server for AI agents.
- Funded by the Walrus Foundation
- Invite-only beta
Perqly
A paid directory of startup credits and discounts from cloud, AI and developer tool vendors, with eligibility details, step-by-step claim guides and an AI assistant that matches perks to your stack and stage.
Which AI models do you use?
Nano Soft picks the model for each task by testing it on your real cases, weighing quality, speed and cost.
We keep the model behind a thin layer, so switching later is a small change rather than a rebuild. Where a task needs no model, we use plain code, which is cheaper and easier to test.
How do you keep our data private?
Nano Soft sends a model only the data a task needs, and we agree in writing, before the build, how your data is handled, stored and deleted.
Our own products show the habits we bring: Just Add reads meal photos and discards them, and Kraterion encrypts every file before it is stored, so the platform cannot read it once access is revoked.
How do you know the AI is good enough to ship?
Nano Soft decides what good looks like before building an AI feature, then tests against it.
That means a set of real examples with expected results, run before every release, plus checks in code for anything that must never happen. Just Add, for example, runs a code gate on every answer, and a failing test for any of its rules stops a release.
What does an AI feature cost to run?
The running cost of an AI feature depends on the task and the volume, so Nano Soft estimates it in the written plan before any build starts.
Then we design to keep it low: caching repeat requests, using smaller models where they are good enough, and skipping the model where plain code works. Preplo caches every extraction, so a popular video is processed once, however many people open it.
Can you add AI to a product we already run?
Yes, Nano Soft adds agents, assistants, search and automation to existing products, working inside your codebase and alongside your team.
We start with one task, build a test set from your real cases, and ship the feature so you can roll it out to a few users first.
Do you build MCP servers?
Yes, Nano Soft builds MCP servers, which let AI assistants such as Claude, ChatGPT and Cursor use your product with permissions you control.
We built MCP servers for three of our own products: Kraterion, Just Add and Inkray. Just Add's server can update a shopping list but can never generate a meal, by design.
Have an AI feature in mind?
A few lines are enough: what the product is, where it stands and when you need it. We reply to every message within 2 working days.
- hello@nano-soft.tech
- Reply
- Within 2 working days
- Based in
- Bucharest, Romania