AI that runs on the device.
And actually works.
We build intelligent features that never send your users' data anywhere — and we test the AI systems that traditional QA breaks on. Seventeen years across quality engineering, full-stack, cloud and mobile, in one studio small enough that you talk to the person writing the code.
Three things, done properly.
We're deliberately narrow. These are the problems we've solved in production — not a capability list assembled to look broad.
On-device & privacy-first AI
You want intelligent features. Your legal team won't let user data go to a third-party API. We build models that run on the phone, so the question never arises.
- Image & text models running locally
- Classification and scoring engines
- Cross-script multilingual matching
- Zero-telemetry architectures
Quality engineering for AI
AI systems don't give the same answer twice, so pass/fail testing falls apart. Seventeen years of QA, pointed at the systems most teams are shipping untested.
- Test strategy for non-deterministic systems
- Evaluation sets & accuracy measurement
- Regression suites for models and prompts
- Automation across mobile, web and API
End-to-end product delivery
When you need the whole thing built rather than a piece of it — from architecture through store review, by someone who has done every layer.
- Native Android (Kotlin, Compose)
- APIs, backends & data pipelines
- Cloud architecture & scheduled jobs
- Web applications
The compliance layer nobody wants to own
If your product touches children, health data or the EU, the regulation is the hard part. We've built to COPPA, GDPR-K and India's DPDP Act — encrypted storage, consent flows, and data-safety declarations written to survive review rather than to sound reassuring.
Kidzy — a content safety engine
we built end to end.
Our own product, and the clearest evidence of what we do. Every claim on this page is something Kidzy required us to actually solve.
The problem
Let a young child watch video without handing them the open internet. Existing filters read English — but most children's content in India isn't in English, and a four-year-old can leave a safe app in a single tap.
What we built
A five-layer scoring engine that assesses every video before it's shown — keywords, language, behavioural risk, publisher reputation and on-device thumbnail analysis — with a threshold the parent controls.
Multilingual matching
15 languages in native script and romanised spelling, plus obfuscation
recovery — g.h.o.s.t resolves before matching. The part most systems skip.
On-device inference
Text recognition and image analysis run locally. No thumbnail, image or user signal is ever uploaded for processing.
Tested, then tuned
Automated suite across the scoring engine, plus live-corpus audits. Filtering throughput improved roughly 4× after profiling — 58 ms to 14.7 ms per page.
Zero-telemetry architecture
No analytics SDK, no ad SDK, no account system. Encrypted local storage with the key sealed in hardware.
Cost-aware backend
A scheduled catalogue build keeps third-party API cost flat regardless of install count, instead of scaling per user.
Compliance built in
COPPA, GDPR-K and DPDP addressed in architecture — consent flow, PIN-gated controls, and a published data inventory.
Small studio. No account managers.
You talk to the person writing the code. We take on a few projects at a time — and if yours isn't a fit, we'll say so rather than take the work.
Discovery sprint
One week, fixed price. We scope the problem properly — architecture, risks, and an estimate you can budget against. You keep the output whether or not we build it.
$1,500 · fixed, most projects start here
Build
Fixed-scope project work with a defined deliverable. Weekly demos — working software rather than status reports.
Quoted from the discovery output
Ongoing
Monthly retainer for teams needing continuous capacity — feature work, quality engineering, or both.
Monthly, cancel with 30 days' notice
Time zones and working language
We're based in Chennai (IST) and work with teams across Europe, the UK, the Middle East and North America — overlapping mornings with the US and full days with Europe. All work is delivered in English.
Four rules we hold client work to,
same as our own.
Safe by default
Nothing we ship exposes users to harmful or adult content, and nothing is designed to distress, addict or manipulate. Safety arrives switched on.
Simple by design
If it needs a tutorial, it was built wrong. That holds for a five-year-old and for someone using their first smartphone at sixty.
Private by construction
Collecting data and promising to handle it well is the wrong shape of promise. We prefer architectures with no collection mechanism to misuse.
Honest about limits
Every system has failure modes. We tell you ours during the project, not after. Someone who over-trusts a tool is worse off than someone who never had it.
Seventeen years, seven disciplines.
Ryddhu AI is run by Karthick Navin from Chennai, India. Seventeen years spanning quality engineering, full-stack development, AI systems, APIs, mobile, web and cloud — the unusual part being depth in both building and proving it works, which are normally different people.
That combination is the reason the AI testing work exists. Most teams shipping AI features today have no meaningful way to know whether they still work after the last change. Seventeen years of quality engineering turns out to be exactly the background that problem needs.
We're a small studio and would rather say so than perform being a larger one. What that buys you: no handoffs, no juniors on your codebase, and an honest answer when something can't be done in the time you have.
Tell us about your project.
What you're building, what's blocking it, and when you need it. You'll get a real reply from the person who'd do the work — not a sales sequence.