All four came out of shipping something real, under load, with someone waiting on it.
Ingest, transcoding, packaging and playback. The whole delivery path, including the telemetry that tells you which sessions are failing and why.
MoreProvisioning, over-the-air updates you can roll back, and telemetry that stays affordable when the fleet grows an order of magnitude.
MoreAuthentication, authorisation and audit trails built to be queried later. Plus on-premise and hybrid deployment where data cannot leave the country.
MoreSpeech and document pipelines in Arabic and English, wired into systems that already exist, with evaluation so you can tell when quality moves.
MoreWe are a small studio, so we are deliberately narrow. These are the areas where we have shipped enough to know where the problems hide.
Playback breaks in specific, boring ways: a transcode ladder that does not match the audience’s bandwidth, a CDN caching the wrong manifest, a player that gives up on the first 5xx. We work the delivery path end to end, not one box in it.
A device you cannot reach means sending someone out to it. So provisioning has to work first time, updates need a way back when they go wrong, and history has to be kept without paying to store all of it forever.
Most incidents come down to a stale credential, a permissive default, or an audit trail nobody can query. We build and harden the access layer, and make what happened afterwards reconstructable.
Most of the effort goes into the pipeline around the model, not the model itself: pulling clean input out of source systems that were never designed to hand it over, handling the cases the model gets wrong, and noticing when the output has drifted.
Nisaba was the Sumerian goddess of writing and record-keeping, the one who kept the accounts straight. That is roughly the job.
We have spent over ten years on production software in Germany and Iraq. In Europe that meant large projects for companies where a bad release is measured in revenue per minute. In Iraq it meant building the systems themselves, at national scale, under constraints most vendors never have to design around. Nobody reads that market from a slide deck; we built inside it, and very few teams know it the way we do.
So we take a small number of engagements at a time, and the people you talk to are the people who write the code. There is no bench, no account layer, and nothing is handed to someone junior once the work is won.
If a problem sits outside what we have actually shipped, we will tell you and point you elsewhere. Learning on a client’s budget is a bad trade for both sides.
No bench, no handoff after the pitch. Whoever you speak with is the person writing the code and answering for it later.
Documented, tested and deployable by your own team. You should not need us on retainer to change a configuration value.
Over ten years building production software in Germany and Iraq, across streaming delivery, connected-device fleets, security systems and applied AI. Writes most of the backend and the infrastructure it runs on.
Interface and interaction design, brand and visual identity. The parts a user actually touches. On every engagement from the first sketch, not brought in to decorate it at the end.
Nearly everything we have built lives inside someone else’s product and is covered by an agreement that says so. What follows is the shape of the work, not the logo on it. It runs in Iraq and across Europe.
Live and on-demand video going out to subscribers over networks where available bandwidth moves hour to hour. The failure mode that mattered was not total outage. It was sessions that started slowly, stalled once, and never came back.
Work covered the path from ingest through transcode ladders and packaging to the player, plus playback telemetry so a bad session could be traced to a segment and a CDN edge instead of argued about.
Hardware deployed where nobody can drive out to it, across several revisions live at the same time, on connectivity that drops without warning.
We built provisioning and device identity, an over-the-air update path with staged rollout and rollback behind health gates, and a telemetry pipeline tiered so that keeping history did not become the largest line on the infrastructure bill.
A permission model that had accumulated years of one-off exceptions, and logs that were written but could not usefully be queried after an incident.
We replaced the access layer with a role model that maps to how the organisation is actually structured, and rebuilt audit so that “who changed this, and when” is a query rather than an investigation. Deployed on infrastructure the client controls, because the data was not permitted to leave the country.
A backlog of scans, forms and recordings in Arabic and English that people were working through by hand, one item at a time.
We built the pipeline outward from the source data: normalising what came out of it, sending anything the model was unsure about to a person, and keeping an evaluation set running so a quality regression showed up before a customer found it.
A paragraph is enough to start. If it is not something we can help with, we will say so quickly instead of booking a discovery call about it.
Mail reaches the whole team. We are in Munich and Baghdad, so between us there is usually someone at a desk.
The things people ask before they get in touch. If yours is not here, ask it directly. We would rather answer than have you guess.
Ask us directly. Technical questions get technical answers, not a callback from someone who will have to check.