Agents in the repo, the ticket queue, and the bill

Product teams still run on a queue. Linear, Jira, or a GitHub project board holds the work. A ticket has an owner, a sprint, and a customer quote pasted from Intercom or Zendesk. Standup is a walk down that board. Cyriac Zeh and Nidrosoft start there, because a model that cannot read the ticket and write a draft comment the owner accepts is a demo. We sit in the grooming, the on-call rotation, and the Friday invoice review before we talk about agents.

Support and success sit next to engineering. A conversation lands in the inbox. Someone tags it, someone searches Notion for the last workaround, and someone opens a ticket if the customer is loud enough. The reply that should have been a saved reply gets rewritten from memory. We map the path from first message to ticket to release note. If that path is three tools and a Slack mention, the first build is a draft that stays in those three tools, not a new portal the support lead will ignore.

Engineering already has AI in the editor. Copilot or an equivalent writes a function. A reviewer still has to catch the wrong API. The useful work for us is the loop around the editor: evals for a feature that calls a model, a prompt that lives in a repo instead of a chat history, and a bill that tells you which endpoint spent the money. Nidrosoft will not replace your code host. We will put a draft eval, a draft changelog, or a draft incident summary where the team already looks.

Founders and heads of product ask for fewer surprise invoices and fewer tickets that rot. They want a brief for a vendor or a hire that does not take an afternoon. They want renewals that do not take down staging because a domain sat on a personal card. They want a design system the next surface can reuse. Those are the jobs TokenTra, Nerlude, NanoBrief, and Omnira UI already cover in public. Custom work starts when the queue, the eval, or the inbox needs a path those products do not ship.

Most teams keep prompts in a Notion page, a Slack thread, or a file that one engineer owns. The production prompt is a cousin of that file. When output quality drops, nobody can say which version shipped. Nidrosoft asks to see the file, the deploy path, and the last time someone changed a sentence. If those three disagree, we stop the feature conversation and put the prompt in the repo with a version the review can see. A model swap without that file is how you ship a regression and call it an experiment.

Evals start as a Sheet with a column for input, a column for expected, and a column for a human score. Someone runs twenty rows by hand after a deploy. Then the sheet rots. We treat that sheet as the spec. We will help you move it next to the code, run it in CI, and keep a human score on the rows that money or customers touch. Automated graders help on format and retrieval. They do not get the last word on a refund email or a medical-adjacent answer, even if your product is not healthcare.

Support macros and agent tools need the same cases. A tool that looks up a customer should fail closed when the id is wrong. A tool that drafts a reply should never send. We write the cases for those tools before we wire them. Teams that skip the cases discover the failure in a public thread. Nidrosoft would rather look slow in week one than explain a sent message in week four. The owner of the inbox remains the person who hits send.

Protectron exists for teams that need EU AI Act classification, drafted technical docs, and checklists by risk level. If you ship an agent into Europe, that paperwork is part of the product. We will not invent a legal opinion. We will point at Protectron and at counsel, and we will keep an audit trail on the agent if that is in scope. A software company that only sells in one US state may not need that path. Say so early so we do not scope a filing you will never submit.

TokenTra is live at tokentra.io. You connect OpenAI, Anthropic, Google, Azure, and Bedrock and see spend in one dashboard. Setup is minutes, not a quarter. The point is attribution. A finance lead can finally say which team, project, or feature burned the month. Budget alerts fire at 50, 80, and 100 percent. Keys can disable when a budget is gone. Chargeback reports go to the people who own the feature. Nidrosoft built this because invoices without names create arguments, not savings.

Waste shows up as a playground key on production traffic, a retry loop, a fat prompt that stuffs a whole corpus, or a model that is too large for a classification. TokenTra's public result on this site is that teams cut 10 to 30 percent of wasted spend after they can see the waste. We will not put that range in your board deck until your providers are connected. Cyriac Zeh will show the dashboard on a live account, not a slide with a fake sparkline.

Custom work around TokenTra looks like routing rules you already believe in. Cheap model for classification, expensive model for the last rewrite, cache for the repeated question. Those rules belong in your repo with an owner. TokenTra tells you whether the rule saved money. It does not write the product. If your only request is a cheaper invoice and you will not change a prompt or a route, buy a reserved commitment from the vendor and skip us.

Agencies and multi-product studios have the same bill with worse names. Each client, each environment, each experiment needs a tag. We have done that tagging work next to Nerlude, because the same team that cannot name a token also cannot name which AWS account holds the key. The two problems travel together. A software company that wants only TokenTra can start there. A company that lost a key in Slack last quarter should look at both.

Nerlude is live at nerlude.com. Founders and small platform teams use it to see domains, hosting, databases, payments, and credentials in one place. Alerts land 30, 14, and 7 days before a renewal. Credentials sit encrypted, split by environment. Contractors get time-limited access instead of a forever key in a DM. Project notes live next to the services so the next hire is not hunting a registrar login. Nidrosoft ships this because outages from a lapsed domain are boring and expensive.

The mess we walk into is predictable. The marketing site is on one registrar. The app is on another. Staging shares a key with production because someone copied a .env two years ago. The Stripe webhook secret is in a pinned Slack message. A former contractor still has the Vercel seat. We do not lecture. We list the services, the owners, and the dates. Then we move secrets into Nerlude or into the vault you already pay for, with the same alerts either way.

Cost tracking per project matters once you have more than one product. A studio with a consumer app, a B2B dashboard, and a forgotten experiment will feel this. Nerlude shows burn next to the services. TokenTra shows the model bill. Together they answer the question finance asks at month-end. We still need a human to cancel the experiment. The software only makes the line visible in time to cancel it.

Larger engineering orgs already have Datadog, a CMDB, and a security team. Nerlude is the wrong buy if you want another enterprise console. It is the right buy when a 30 to 80 person company has outgrown a spreadsheet and has not grown into a platform org. Custom Nidrosoft work starts when your renewals live in NetSuite or a procurement queue and you need the alert to open a ticket, not only an email.

NanoBrief is live at nanobrief.io. A short industry questionnaire becomes a shareable brief in under two minutes, against the three-plus hours people spend rewriting the same doc. Teams export PDF or Word, regenerate a section, and keep version history. Product, design, and agency leads use it when the kickoff keeps slipping because nobody wrote the brief. Nidrosoft will not claim it writes your PRD for a regulated market. It writes a first brief you can share without an account on the viewer side.

Custom brief work starts when your template has fields NanoBrief does not ship. A security review checklist with SOC questionnaire rows. A marketplace listing form. A partner integration one-pager that legal already blessed. We take those templates, keep the questions short, and leave the risky paragraphs for a person. The public two-minute figure is for the shipped questionnaires. Your template may be slower if counsel added six required fields. We will time it on your questions in your Notion or Confluence page, not on ours.

Omnira UI is the open-source design system under the Omnira products. If you are building an operator console, an agent inbox, or a review queue, you can start from those patterns instead of inventing another table, another approval chip, and another empty state. Nidrosoft still designs the flow for your objects. We will not drop a theme on a broken process and call it a product. Cyriac Zeh treats the system as a way to ship the next surface without redrawing the same button.

Omnira Dental is the live vertical that uses this thinking in a clinic. Omnira Real Estate is in build. Software companies that want an agentic console for their own industry can look at those products as proof we ship operator software, then ask for a build in their domain. We will not slap a dental recall flow onto a SaaS admin and call it done. The objects change. The review-queue habit does not.

Embed means we join the standup, read the last ten incidents, and watch a support lead clear the inbox. We clone the repo we are allowed to see. We sit on the invoice review if finance will have us. Nidrosoft does this from San Diego on your accounts. We do not ask for a week of workshops before we have seen a ticket. The method is the same one on the how-we-work page. Software companies just have more logs.

Diagnose means we cost the recurring work. Hours on changelog writing. Hours on evals run by hand. Hours hunting a key. Dollars on tokens that nobody can name. We score feasibility against your stack. A team on GitHub Actions and Linear gets a different first build than a team on a locked-down Jira with no API access. If IT cannot grant a token, we write that as a blocker and we do not pretend the roadmap is free.

Scope means you buy a result. Draft comments on tickets that match a label. An eval suite that fails the deploy. A TokenTra view your finance lead will open. A Nerlude project for the products that already exist. We write what the agent may do and what it must queue. Sending email, charging a card, and deleting data stay behind a person. If a founder wants an agent with production write access and no review, we end the scoping call.

Build and Own happen with an engineer on your side in the repo. We pair, we leave tests, we leave runbooks. Training is a sprint on your tools, not a slide about transformers. When we step back, on-call stays yours. We remain available for the work you name. There is no handover theater because the code never lived in a private Nidrosoft org you cannot see. Ask for that access in writing on day one.

An AI audit for a software company covers the queue, the prompts, the eval sheet, the token invoice, and the renewal calendar. We interview an engineer, a support lead, and whoever pays the cards. Cyriac Zeh will say on the first call if you only want a tools tour. Nidrosoft sells a ranked list of where automation returns time or money. We do not sell a new architecture for its own sake.

We write each opportunity next to the system it touches. Linear. GitHub. Intercom. The OpenAI project. The registrar. That list keeps the roadmap honest. A build that needs four new vendors is usually the wrong first build. We would rather ship a draft changelog from merged tickets than a platform that replaces your queue. If the audit finds that ChatGPT seats are enough for the writing you do, we will say so and leave.

Security and tenancy go on the map. Where prompts log customer data. Which keys can hit production. Whether a contractor still has a seat. We are not your pentest firm. We will refuse a build that dumps tickets into a consumer chat with no agreement. EU hosting and residency are available when your policy needs them. Azure, AWS, Google Cloud, Cloudflare, or the private infrastructure you already run. Pick the one your security lead will sign.

Email cyriac@nidrosoft.com with the stack and the pain. Linear or Jira, GitHub, Intercom, and which model vendors you pay. Tickets, tokens, renewals, or briefs. Two weeks is the usual audit window. If you already know you want TokenTra or Nerlude as products, say that in the first line. Product adoption and a custom build are different scopes and different access lists. Mixing them without saying so wastes a week of IT tickets for a repo we never needed.

A first build we take is a draft on the ticket. A labeled Linear or Jira issue gets a suggested repro, a suggested test, or a suggested reply, posted as a comment the owner already sees in the queue. The owner edits or dismisses. We log the model and the prompt version on that comment so a later incident review can see what shipped. Teams that want the bot to close tickets will not like us. Closed tickets without a person are how you close the wrong customer and lose the thread in Intercom.

A second build we take is evals in CI. The Google Sheet or CSV of cases becomes a GitHub Actions job that fails the pull request when format or retrieval breaks. A human score remains on the rows you mark, usually refunds, account deletion, or anything that names a customer. We have done enough model work to know a green check on a similarity score is a weak ship decision. Your release owner still clicks merge. The job only stops the silent regression that used to hide until Friday Slack.

A third build we take is spend and renewals wired to the people who can act. TokenTra alerts to Slack or email. Nerlude alerts to the owner of the domain. A custom ticket in Linear when a budget crosses 80 percent. Finance sees a chargeback. Engineering sees a key to rotate. Nobody sees a PDF at month-end as the first signal. That is the whole point of putting a name on the dollar.

Transformation is the longer path: audit, two or three builds, training for the team that will own the prompts, then the next squad. We need a decision maker, a repo we can open, and a cloud we can bill to you. Rent-an-engineer exists when you want a senior person embedded monthly instead of a fixed scope. Both paths still end with your team on the pager. Nidrosoft will not keep production keys after you can run the job.

Corporate AI training for a software company is a focused sprint. Leadership gets the ranked map and the gates. Engineers bring a live prompt, a live ticket type, or a live invoice. We write the first eval or the first draft comment together. The playbook they keep is their own repo path, their own labels, and their own budget alerts. A generic workshop with public examples is a different product. Say which one you want.

We refuse training that is only a tour of chat tools. If the room cannot open Linear or the billing dashboard, the afternoon will not change the sprint. Buy a course. Come back when a manager will grant access. Cyriac Zeh has run rooms in San Diego and on livestreams for operators who brought their own work. The same rule applies here. Your ticket or your bill has to be in the room.

Designers and PMs get NanoBrief and Omnira UI in that sprint when the gap is the spec or the surface. Engineers get TokenTra and the eval job when the gap is money or regressions. Mixing both rooms without a shared backlog produces two playbooks that never meet. We will split the day if we have to. We will not pretend a single lecture covers a support lead and a staff engineer.

After training, the first build should already be obvious. A labeled ticket type, an eval job, or a TokenTra alert with a feature name. If it is not obvious, the audit was the missing buy and the room only collected opinions. Nidrosoft would rather sell the two-week audit than sell a second workshop that repeats the first afternoon. Email cyriac@nidrosoft.com if you are choosing between them. We will pick one and say which access we need on Monday.

Hire a strategy consultancy if the board wants a market map and no one will open the GitHub org or the billing account. Nidrosoft is a poor fit for that mandate. Cyriac Zeh will say so on the first call. We are also a poor fit if you only want ChatGPT seats rolled out with a policy PDF and an Okta group. TokenTra can wait until the seats exist and the invoice hurts. Until then, buy the seats from the vendor and skip a studio invoice.

Do not book us without a decision maker. A staff engineer with curiosity and no budget will get a kind decline. A vendor who wants us white-labeled without access to your Linear queue will get the same. We work with teams, often 30 to 250 people, who can name an owner, grant a repo token, and pick a repeating workflow. Smaller teams fit when one build pays for itself before the next hire. Larger orgs start in one squad, usually support or the team that owns the model bill.

We will not take a scope that gives an agent production writes without a queue. Deleting customers, charging cards, emailing a list from Customer.io or Mailchimp, or merging to main without a human is outside what we ship. Other firms will demo an autonomous closer in a staging project. You will own the incident when it hits production. Nidrosoft would rather lose the deal than keep a write key after go-live.

If you need a consumer app in dating, social, food, or travel, look at Rizzer, Night Owl, Nourishr, and Guidera in the public archive. Those are Nidrosoft studio products with their own stores and accounts. They are proof we ship mobile, not a white-label menu. We can build in that shape when the scope, the brand, and the App Store listing are yours. We will not reskin them, swap a logo, and sell them as your roadmap in a pitch deck.

Write cyriac@nidrosoft.com with the tracker you use, the models you pay, and the pain. Linear or Jira, OpenAI or Anthropic, tickets, evals, tokens, or renewals. Cyriac Zeh reads that mail. Attach a redacted invoice or a screenshot of the board if the bill or the queue is the fire. Book the public calendar if a decision maker can join the same week. If the note is a partnership deck with no system named, we will decline in one reply.

The first live process should show up in a tool you already open every day. A comment on a ticket. A failing eval job on the pull request. A Slack or email alert with a feature name on the token line. Nidrosoft has 125 products shipped and 16 public systems. TokenTra, Nerlude, NanoBrief, and Omnira UI are the ones software teams should look at first. The rest of the archive is context so you can see we ship. It is not a menu you must buy in one contract.

You own the code on day one. Hosting stays on Azure, AWS, Google Cloud, Cloudflare, or the private infrastructure you already run. EU residency is available when a customer contract or a security questionnaire requires it. Audit trails stay on in the repo and in the billing dashboard. After we leave, your engineer should be able to rotate a key in Nerlude or the vault and change a prompt without a San Diego call. If they cannot, we did not finish the Own step.

Funds that back you may want the venture capital page for dealflow and LP packs. Plants and clinics in your customer base have manufacturing and healthcare pages with different objects and the same gates. Read how we work for Embed, Diagnose, Scope, Build, Own. Read about for Cyriac Zeh and the studio. Use contact for the calendar or cyriac@nidrosoft.com. The work starts when someone who can approve repo and billing access is on the thread.