Native Android and iOS applications built and shipped through Google Play and the App Store — Kotlin and Swift, real release engineering, signed builds, and staged rollouts rather than prototypes that never leave a laptop.
Speech recognition, computer vision, and language models trained on rented GPU clusters, then quantized and exported to run entirely on a phone — no round trip to a server, no per-request inference bill.
Python API services backed by PostgreSQL, deployed behind nginx with TLS on cloud infrastructure we provision, monitor, and keep running. Including search platforms handling millions of documents.
Fine-tuned a transformer speech-recognition model for a historic liturgical language with almost no existing training data. That meant building the corpus itself — sourcing and rights-clearing audio, aligning it to text, and running a disciplined ledger of every training experiment against word-error-rate, including the runs that did not work.
The trained model was then exported to a mobile inference runtime and shipped inside native Android and iOS apps, so recognition happens on the device itself.
Built a production search engine over a large body of ancient-language texts in Greek, Latin, and Syriac alongside English translations. Combines traditional full-text search with vector similarity, so a query finds passages that mean the same thing rather than only those sharing a word.
Required an ingestion pipeline for messy scanned and OCR'd sources, specialist natural-language tooling for ancient languages, and a deployment that has been live and serving real users on custom domains.
A consumer mobile app that runs instance segmentation on camera input and generates narration from a quantized small language model — both entirely on-device. When the standard export toolchain produced silently incorrect output, we diagnosed it, exported only the clean part of the network, and re-implemented the detection decoding and suppression stages in the app runtime ourselves.
Shipped to both Google Play and the App Store, including the model training, the mobile port, and the store release process.
We start by finding out what actually has to be true for the project to succeed — the constraint, the deadline, the thing that will break. Scoping honestly at the start is cheaper than discovering it in month three.
Every project has one piece that might not work. We build that piece first and measure it honestly — including when the result is negative. A model that does not hit its accuracy target is worth knowing about early, not after the UI is finished.
Tested code, reproducible builds, documented deployment runbooks, and infrastructure defined well enough that someone else could take it over. We write things down as we go, because the handover is part of the job.
App store submissions, production deployments, certificates, monitoring, and the unglamorous maintenance afterwards. We have taken projects all the way through, so we know what the last ten percent actually costs.
Native apps for both platforms, taken through the full release process.
Training, evaluation, and getting models small enough to run on a phone.
API services, relational and vector data, and the pipelines that fill them.
Provisioning, deployment, and keeping production alive afterwards.
Not every project needs all of it. Most need a few pieces done well and connected properly — which is usually where projects go wrong.
We take on a small number of projects at a time, across mobile, machine learning, and backend work. If you have something in mind — an app, a model, a platform, or an existing system that needs finishing properly — send us a short description of the problem and the constraints you are working under, and we will tell you honestly whether we are the right fit.
Email Us