Token bills, compliance packets, and a human on the money

We work with product finance, platform, and compliance teams who already pay model vendors and already answer questionnaires. The person we sit with can open last month’s OpenAI or Anthropic invoice, or they can open the folder where the last risk assessment lived. They know which feature blew the bill and they cannot prove it in a meeting. They know which document counsel asked for and they rebuilt it from a wiki. Cyriac Zeh founded Nidrosoft in San Diego to put those two objects (the bill and the packet) into systems the company owns. If you need a charter, a BIN, or a money transmitter license, you need a different kind of firm.

A normal month inside an AI-heavy product looks like this. Engineering ships a new agent. Support starts using it on tickets. Marketing pastes a campaign into another vendor. Three invoices arrive with different date ranges and no feature names. Finance asks which team to charge. Nobody can say. A spike on a Tuesday dies in a Slack screenshot. TokenTra exists because that month is already a known shape: providers, keys, features, and a person who has to explain the number. We start there when the pain is money you already spent, not money you wish you could move.

A normal quarter on the compliance side looks like a questionnaire with a deadline. Someone is asked for a risk classification, a model card, a technical file, and a list of where the agent writes. The answers live in five repos and a Notion page that is two months stale. Counsel wants a PDF. A buyer in Europe wants a badge and a date. Protectron exists because that packet is already a known shape under the EU AI Act. We start there when the pain is a document you must produce, not a theory about regulation.

We are the wrong seat when you want Nidrosoft to hold customer balances, issue cards, underwrite loans, or operate as your bank. We do not. We are the wrong seat when you want an agent that wires funds, books a payment, or changes a ledger without a named human. We will refuse that scope in the first meeting. If you want ChatGPT seats for the finance intern, buy seats. If you want the bill by feature and a packet you can hand to counsel, keep reading and then use /services/audits.

Token invoices are a translation job. The vendor speaks in models, tokens, and projects you named once in a dashboard. The company speaks in features, teams, and customers. A single key used by two services will lie in both directions. A proxy that nobody documented will hide the expensive call inside a cheap-looking feature. Finance cannot charge back what it cannot name. Engineering cannot cache what it cannot see. The first operational fact we collect is the map from key to service to owner. Without that map, every later conversation about “saving money” is a guess.

Budgets fail at the threshold, not at the slogan. A team says they have a cap. The cap lives in a spreadsheet. The key keeps working on Saturday. Monday’s invoice is the news. Alerts only help if someone owns the channel and has the right to disable a key. Plenty of shops have the alert and no owner. The on-call is a designer. The person with billing admin is on leave. We write the owner into the scope. A meter without a hand on the switch is a report you will ignore twice.

Questionnaires fail at the exhibit. Counsel asks for the intended purpose, the data categories, the human oversight, and the logging. The engineer who could answer shipped last quarter. The doc template is from a vendor blog. Someone pastes last year’s answers and changes the date. A buyer’s security person catches the contradiction on a call. The deal slips. We treat the packet as a file with required sections and an owner, the same way we treat an eviction notice as a file with a clock. The law is different. The habit of a complete packet is not.

Approval paths fail in the ticket comment. A product manager writes “finance signed off” and attaches nothing. A model starts drafting customer emails that mention balances. A batch job retries a payout. After the fact, nobody can show who said yes. We do not fix that with a longer prompt. We fix it with a record: who approved, what they saw, and what the system was allowed to write. If your company cannot name that person, you are not ready for an agent on a money-adjacent path. That is an operational finding, not a moral speech.

Protectron is a live Nidrosoft system for EU AI Act work. It runs a risk classification questionnaire, drafts the documents that questionnaire implies (technical notes, risk assessments, model cards), and keeps a checklist that changes with the risk level. Teams use it to get a packet into the shape counsel and buyers already ask for. A person still reviews every page. A lawyer still decides what you file or send. We built the product because the packet was eating weeks of copy-paste, not because Nidrosoft is a law firm. We are not.

Classification is a set of answers, not a vibe. Intended purpose, users, whether the system is a component, whether it touches a listed high-risk area: those questions have to be answered with the product you actually shipped, not the pitch deck. We have watched teams classify a support helper as “minimal” while it already drafts messages that mention balances. Protectron will not save you from that contradiction if you lie to the form. It will make the contradiction visible because the answers and the generated sections sit together. That visibility is the product.

Agent trails matter when the system writes. If you run agents, someone will ask what the agent did, with which tool, at what time. Protectron’s job on that side is a trail you can show, not a story you tell on a call. The trail is only as good as the instrumentation you allow. A shop that will not log tool calls will not get a trail by buying a badge. We say that in the first week. A badge on a marketing site with an empty log is a buyer problem waiting for a screenshot.

The output is a packet with owners and dates. Who answered the questionnaire. Who edited the model card. Who said the residual risk was acceptable. That is the same discipline we use on money-adjacent writes: a name, a timestamp, a thing they saw. If your counsel wants a different template, they win. Protectron is a draft and a checklist, not a substitute opinion. If you need a legal memo on whether you are in scope, hire counsel. If you need the draft to stop living in a stale wiki, this is the system we already run.

Money-adjacent work shows up even when the company is not a bank. A billing agent that issues credits. A support agent that promises a refund. A collections note that changes a due date. A payout batch that retries. Those writes hit a customer and a ledger your controller already owns. Nidrosoft will draft, route, and file. A named human still approves the write that moves money or creates a promise that must be honored. We put that sentence in the scope. If you delete it, we delete the engagement.

The approval has to see the same record the model saw. A thumbs-up on a Slack blur is how refunds go wrong. We show the customer, the amount, the reason, and the policy line. We show what the model proposed and what it was not sure about. The approver can edit or reject. The trail stores the decision. That is slower than an unattended agent. It is also how you still have a company after a bad Tuesday. We will not compete with shops that sell unattended payouts as a feature.

Ledgers stay where they already live. We do not become your core, your processor, or your recon tool. If you run Stripe, Adyen, a bank partner, or an internal ledger, we read and we propose. The posting stays on their side of the line. Nerlude, our infrastructure control center, can hold credentials and renewal dates for the services around a product. It does not hold customer money. We mention it here so a founder who is drowning in keys and invoices has a place to put those objects that is not a Slack DM.

Customer-facing chat that can see a balance is a special case. People will ask the bot to “just refund it.” The safe product is a draft ticket for a human, not a silent credit. We have no interest in a demo that completes a refund in the meeting. If your risk committee needs that demo to feel modern, we are the wrong vendor. If they need a path that is boring and logged, we can build it. The difference is the whole job.

Nidrosoft is the San Diego studio of Cyriac Zeh. He ships end-to-end AI product work under the name The AI Architect. Before the studio he worked at Anthropic, Microsoft on Bing and Edge, Intuit, and Gap Inc. Intuit is a finance-adjacent past, not a claim that we ship TurboTax. We keep 16 public systems in the archive and we have shipped 125 products. TokenTra and Protectron are the two you should open if this page is why you came. The others (Omnira Dental, Eviction Wizard, Nexuvo, Octana) show that we ship into regulated-feeling desks without pretending we hold a license we do not hold.

We will not list invented bank logos. We will not say we processed a volume. We will not say we are SOC certified unless you can find that claim on a product page we control, and this long page will not invent it. If a number is not on facts we already publish (10 to 30 percent wasted spend cut on TokenTra, the studio counts above), we will not dress it up for a FinTech reader. That restraint is the point. A shop that invents volume will also invent an approval.

Helia, Guidera, Rizzer, Night Owl, and Nourishr are consumer systems in the same archive. They are not payment products. We mention them once so you can see the studio finishes software, then we return to invoices and packets. If you want those apps, they are on /work. If you want the bill and the EU packet, stay here. If you want a core banking build, you already know we are the wrong tab.

Embed means we sit with the person who already explains the invoice or the questionnaire. We watch them export a CSV, join it to a team list, and fail. We watch them hunt a model card in a repo. We take the admin access they already have to the provider dashboards, under their policies. We do not start with a workshop about agents paying invoices. The first week is access, shadowing, and a list of keys. The longer path is on /how-we-work.

Diagnose means we cost the repeating work and the repeating failure. Hours spent reconciling the bill. Hours spent rebuilding the packet. Incidents where a key had no owner. Deals that slipped because a document contradicted a demo. We score whether we can reach the providers and the places documents already live. A team that will not show a redacted invoice cannot buy a cost project in good faith. We will say that in week one rather than invent a number.

Scope means you buy a result. For spend, the result is named features, an owner on each key, alerts a human reads, and a monthly file finance can use. For the Act, the result is a classified system, a drafted packet, and a named reviewer. For a custom money-adjacent agent, the result is a draft plus an approval record, never an unattended post. We refuse “automate finance.” If you cannot name the write, you are not ready.

Build means we ship into your accounts. Your provider logins, your doc store, your ticket tool. Your people sit with us while the first dashboard or the first packet goes live. Someone in-house has to be able to add a feature label, rotate a key, and explain a classification answer on a Friday. That person is part of the build. Twelve-plus years of product work has made us impatient with a vendor-only admin.

Own means the credentials, the classifications, and the approval policy stay yours. We maintain what you ask us to maintain. We stay out of the rest. If a provider adds a model or the Act guidance shifts, the owner of the packet is you, and we update the draft with you. Continuous work is a cleaner bill and a packet that still matches the product you shipped, not a retainer that quietly becomes staff aug.

Do not hire Nidrosoft to be your bank, your EMI, your broker-dealer, or your processor. We will not hold customer funds, issue cards, or settle to a merchant. If your RFP is a core replacement, a card program, or a lending book, you already have a shortlist that is not us. We can still talk if a slice of that company is a model bill or an EU packet. We will not let that slice pretend to be a license.

Do not hire us for a legal opinion. Protectron drafts. Counsel opines. The difference is the whole product. A founder who wants a PDF that says they are “compliant” with no lawyer in the loop is asking us to sign our name under theirs. We will not. The same no applies to tax, to money-transmitter questions, and to “can we launch in this country.” Those are firms with bar cards. We build the file those firms read.

Do not hire us to hide a payout in an agent. If the demo requires an unattended wire, refund, or ledger edit, we will stop the demo. There are shops that will do it. They can have that work. Our public systems (TokenTra, Protectron, Eviction Wizard, Nexuvo) all leave a human on the commit that hurts. That is a studio rule, not a FinTech slogan we invented for this page.

We are a poor fit for a team that only wants a keynote, and for a two-person startup that has no invoice yet and wants a “FinOps platform” for theater. Nidrosoft’s main seat is an operator with a repeating bill or packet and a decision maker in the room, often 30 to 250 people, sometimes smaller when one meter pays for itself. If you have no spend and no questionnaire, wait. Write us when the PDF from the vendor makes someone angry.

An AI audit is two weeks in the work. We read a redacted invoice, a key list, and a sample questionnaire. We sit with finance and with the engineer who actually created the keys. We write down every money-adjacent write an agent already does or that someone has proposed. The deliverable is a ranked map: meter first, packet first, or “stop the unattended refund idea” first. Details live on /services/audits.

Access is the gate. Provider admin, a doc store, and a ticket project if approvals already hide there. We use your accounts and your policies. We do not ask you to paste customer ledger rows into a consumer chat so we can “try a prompt.” If counsel needs a DPA, we do that before we open the box. If you cannot show a bill, we cannot rank waste, and we will not invent a 10 to 30 percent for your slide.

We score failure the way a close or a buyer scores it. An unlabeled key is a fight in the close. A stale model card is a slipped deal. An unattended credit is an incident. Those outrank a slightly nicer dashboard chrome. Sometimes the first recommendation is “name the keys and pick an owner,” with TokenTra as the second step. That is a successful audit. You can take the map to another shop.

At the end you get a build order in plain language. If the first line is “stand up TokenTra and label features,” we write that. If the first line is “hire counsel and stop sending the wiki to buyers,” we write that even if it does not need us. The audit is the product. The build is a separate buy. Start at /contact or send cyriac@nidrosoft.com a redacted invoice and the last questionnaire you answered.

Software teams are the closest sibling because the bill is often theirs. Agents in the repo, retries in a queue, a feature flag that quietly doubled calls: that page is for the people who write the code. This page is for the people who have to explain the code’s invoice and the packet. You may need both. We will not sell you two copies of the same audit. We will say which owner should sit in the first week.

Professional services transfer the counsel desk. Protectron packets often land on a firm that sells hours. Intake, a deadline, a human who will not let software file the opinion: that discipline is on the professional services page. Real estate transfers the clock. Eviction Wizard holds statutory dates the way a questionnaire holds a buyer deadline. Automotive transfers the night job: Nexuvo sends outreach while the lot is closed, with a human on the buy. We do not buy cars with your treasury.

E-commerce transfers chargebacks and inbox promises, which are money-adjacent in a different costume. We have no flagship storefront, and we say so there. Healthcare transfers the record habit from Omnira Dental: a write that changes a chart or a claim needs a person. The object changes. The refusal to let a conversation be the ledger stays. If your “FinTech” project is actually a clinic billing agent, say that and we will send you to the healthcare page first.

Write cyriac@nidrosoft.com with a redacted invoice, a key list, or the last EU or buyer questionnaire you answered. Name the person who can disable a key and the person who can approve a money-adjacent write. Cyriac Zeh reads that mail from San Diego. We will tell you whether the next step is TokenTra, Protectron, a two-week audit, or a no because you need a bank or a law firm. We do not need a pitch deck about the future of money.

If you already want the ranked map, book the audit on /services/audits. Two weeks, your accounts, your people, a build order you can take elsewhere. If you already know the bill is the fire, say so in the first line. If the packet is the fire, say that. If an agent is already refunding in production, say that in the first line so we can talk about stopping it before we talk about chrome.

A workshop is available when finance, platform, and counsel do not share a story. Corporate AI training at Nidrosoft uses your invoice and your questionnaire, not a toy bank. Leadership sees what a meter can show and what a person still has to sign. The playbook stays. A workshop is enough when alignment is the problem. It is not enough when the keys are already unnamed and the agent is already writing. We will say which side you are on.

Nidrosoft is twelve-plus years of product work in a studio that still answers mail. 125 products. 16 public systems. The FinTech-shaped proof is TokenTra (10 to 30 percent of wasted spend cut) and Protectron (EU AI Act drafts with a human reviewer). We are not a bank. If that mix matches the desk you run, write us. If it does not, use the sibling pages or go back to the invoice. Both are honest endings.