AI automation for US companies, built by an engineer in Lithuania

Lukas Ceponis writes the workflows, wires the integrations, runs the deploys and answers the email. There is no account manager sitting between you and the person doing the work, which is why replies come back with specifics rather than a promise to check with the team.

Everything is a fixed scope at a published price, and the work happens on accounts registered in your name, so a US client ends up owning every credential, workflow and repository at the end of it. Previous builds and rescues are written up on the work page.

A US company paying a foreign entity is expected to hold a valid Form W-8BEN-E before the first payment goes out. Without one, the payer is generally required to withhold 30 percent under Chapter 3 and remit it to the IRS. We complete the form and send it with the first invoice, or earlier if your finance team wants it sitting in the vendor record before anything is approved.

Two things usually settle the withholding question. The income tax treaty between Lithuania and the United States means business profits are taxable in Lithuania unless the supplier has a permanent establishment in the US, and we have none. Separately, services performed entirely outside the United States by a foreign entity are generally treated as foreign-source income, which falls outside Chapter 3 withholding to begin with. Every hour of our work happens in Lithuania.

Your accountant has seen this before. Confirm the treatment with them before anything reaches a filing, because we build software for a living and give no tax advice.

Form 1099-NEC reports payments made to US persons and pairs with a Form W-9. A foreign entity hands you a W-8BEN-E instead, and payments to it do not belong on a 1099-NEC at all. Where a payment to a foreign vendor is US-source and reportable, the form involved is 1042-S, which is one more reason the source-of-income point above is worth getting into your records properly. If your bookkeeping software insists on a tax identifier, use our EU VAT number; we have no EIN to give you.

Application-to-person messaging that touches a US mobile network runs through the 10DLC framework, and there are two registrations to complete. Brand registration identifies the legal entity behind the messages using your company details and EIN. Campaign registration describes the specific use case, the sample messages, the opt-in language and how opt-out is handled, and it is reviewed before traffic is allowed to flow.

Unregistered traffic does not fail loudly. It gets filtered, which means messages leave your provider, come back accepted, never arrive on the handset, and show a delivery status close enough to success that teams often lose weeks before anyone works out the campaign was never approved in the first place.

Treat registration as a lead time item and start it before the launch date rather than after it. We handle the submission and write the opt-in copy as part of any build that sends SMS, and the registration sits under your brand, because the brand is yours.

US carriers authenticate caller ID with STIR/SHAKEN. The provider originating a call signs it with an attestation level that records how much they know about the caller and whether that caller is entitled to use the number. Full attestation requires the originating provider to have a direct relationship with you and to have verified the number you present.

A call arriving with weak or missing attestation feeds the analytics engines that decide what shows on the screen, and the practical result is a call labelled as spam, or blocked before it rings. For an agent calling your own customers back, that label is the difference between a booked job and a missed one. We provision numbers so the originating carrier can attest to them properly, and we check what the label looks like on a real handset before anything goes live.

We do not build unattended outbound cold calling, and we do not build SMS blasting.

That answer is fixed.

The Telephone Consumer Protection Act is the reason. Marketing calls and texts placed with an automatic dialing system or an artificial or prerecorded voice generally require prior express written consent from the person on the other end, and national and company-specific do-not-call rules sit on top of that. Damages are statutory and counted per call or per message, so the arithmetic on a list of fifty thousand numbers gets ugly fast.

Inbound work is most of what we do: answering the calls your customers place to you, booking them into a calendar, and writing the outcome back into your systems. Outbound that continues an existing conversation with a customer who asked for a callback is workable too, with the consent trail recorded where your lawyer can find it. If the plan is a dialer pointed at a purchased list, say so in the first email and we will save you the call.

By default, workflow state and application data sit in EU regions, and model calls go to the endpoints the model vendor operates. Call audio and transcripts from a voice agent are stored wherever the voice platform is configured to store them, which we set deliberately per project instead of leaving on a default nobody read.

When a US client requires US processing, we configure it: US regions for the database and the workflow runtime, US endpoints where the vendor offers them, and a written note in the handover document naming which vendor holds what and for how long. Retention is a setting. Name the number of days you need and we set it, and since the accounts are yours, you can read every one of those settings back yourself whenever you like.

Field service and SMB software in the US looks different from the European equivalent, and integration work is where a build either succeeds or quietly rots. We have connected HubSpot, Salesforce and QuickBooks, and the same method applies to ServiceTitan, Housecall Pro and Jobber: read the API surface, map the objects that matter, respect the rate limits, and write back in a form an office manager can audit later.

We have not yet shipped a ServiceTitan integration for a paying client, so there is no case study here pretending otherwise. What you get instead is a scope that names which parts of a vendor API are documented and which are guesswork, with discovery priced separately when the risk is real. The fix library shows the failure modes we deal with most, and the production readiness checklist is the bar anything has to clear before we call it done.

Lithuania runs seven hours ahead of New York and ten ahead of Los Angeles, and we shift the day so the second half of ours overlaps the first half of yours.

An email from the East Coast sent before lunch gets an answer the same afternoon. From the West Coast, a message sent in your morning comes back by your early afternoon, and anything sent in your evening is waiting when you open the laptop. Calls get booked inside your working window. During a build sprint there is a scheduled call each week and a shared channel between them, and care clients get fixes inside 24 to 48 hours depending on tier.

Every number on this site is a US dollar number, fixed before any invoice and unchanged by which state you are in. Here is the ladder.

Invoices are issued in USD and paid by bank transfer. Builds run 50 percent on signature and 50 percent on delivery. Retainers bill monthly and roll, with 30 days notice to stop. Model usage, phone numbers and hosting are paid by you directly to those vendors at cost with nothing added on our side, and the ones people forget to budget for are itemised in the post on running costs nobody mentions. What each tier includes is on the pricing page.

British buyers arrive with a different list of questions, mostly about VAT, data adequacy and IR35, and those are answered on the UK page. Everything else on this site reads the same from either country: the rescue service for a build that is failing right now, and the questions worth asking any AI supplier before you sign with us or with anyone else.