We build AI agents, websites, software and automation. Every claim on this site has a commit behind it.

58 repositories. One client engine: 2,245 commits in 71 days, 239 production Postgres tables, 54 edge functions, 157 CI test guards. The AI receptionist is one product at $497 — the rest of the stack is the company.

What we build

• AI agents — A voice or chat agent that answers on your data — built, deployed and operated. From $497, published range $497 one-off · $4,950–$9,950 build · $1,997–$4,997/mo run.
• Websites — A site that search engines rank and answer engines quote — not just a nice-looking brochure. From $8,000, published range $8,000–$15,000 build · $300–$1,060/mo care plan.
• Custom software — Production software that ships — portals, MVPs, internal tools and native apps you own outright. From $35,000, published range $35,000–$150,000 per engagement.
• Automation — The copy-paste job between your systems, deleted — and monitored so you know when it breaks. From $5,000, published range $5,000–$50,000+ to build · $3,000–$20,000/mo to operate.
• AI search — Be the source the answer is built from — measured, not asserted. From $1,000 audit, published range $1,000 audit (credited) · retainer sized to scope.
• AI consulting — A ranked, costed answer to 'where does AI actually help us' — from people who also ship the build. From $1,000, published range $1,000 audit · $5,000–$30,000/mo fractional CAIO.

The receipts, and the command behind each one

• 58 Repositories — Git remotes under the org
• 2,245 Commits on one client engine, in 71 days — git log since 13 June 2026
• 239 Production Postgres tables on that engine — information_schema, production
• 157 CI test guards blocking a bad deploy — 25 GitHub Actions workflows
• 2,245 Commits — Since 13 June 2026 — 71 days
• ~487,000 Lines of code — Tracked files, excluding vendor
• 1,378 Static pages generated — Build output
• 456 SQL migrations — Ordered, replayable, in-repo
• 239 Production Postgres tables — Live schema
• 56 Supabase edge functions — Deployed
• 35 pg_cron scheduled jobs — Running on schedule
• 25 GitHub Actions workflows — 157 test guards across them

One client engine, in detail

• Commits —
• SQL migrations —
• Production Postgres tables —
• Supabase edge functions —
• CI guards and verifiers —
• Static pages generated —

Selected work you can check

• UD Growth Engine — A full revenue operating system — lead capture, dispatch, quoting, marketing and reporting — built as one Postgres-first platform for our own roofing company. 2,245 commits in 71 days (31.6/day) (git rev-list --count HEAD).
• Sovereign — A Tauri 2 desktop application that keeps its data, its secrets and its audit trail on the operator's own machine rather than in someone else's cloud. 38,857 lines of Rust across 115 modules (find src-tauri/src -name '*.rs' -exec wc -l {} + | tail -1).
• Coast — A native iOS companion for people on GLP-1 medication, with real subscription infrastructure and a hardcoded safety guard that runs before any language model call. 7 crisis categories intercepted before the LLM (ios/Coast/Services/Coach/RedFlagGuard.swift — enum RedFlagKind).
• Last Signal — A mobile survivor game built in Unity 6 with the Universal Render Pipeline, a Supabase backend and complete store-listing packaging for both app stores. 465 C# scripts (177,390 lines) under Assets (find Assets -name '*.cs' | wc -l).
• CurrentAds.net — The website and lead system for CurrentAds, our sister performance-advertising agency — server-rendered React 19 running as a single Cloudflare Worker on D1. 74 file-based routes on one Worker (find src/routes -name '*.tsx' | wc -l).
• Whisper — A real-time conversational voice platform where speech-to-text, a language model and speech synthesis complete a full turn in under a second. 700–1000ms measured warm turn, end to end (ARCHITECTURE.md line 74 — 'measured live, warm').

What does each line cost, against what the market charges for it?

Six lines. Each is priced against the published 2026 market band for buying that specific thing, read off first-party vendor and industry pricing pages. On 2 of the 6 our published ceiling sits below the bottom of that band. On 3 it overlaps and we are not cheaper, which each of those lines says plainly. On 1 nobody publishes a comparable band at all and we publish the method instead of inventing one. A supplier claiming to undercut the market on all six things it sells is either lying or bad at four of them.

The 21-day answered phone — AI agents

The call is answered in two rings at 9pm on a Sunday, the caller is qualified, the job is booked into your calendar, and the record is in your CRM before you have read the notification. For: A service business where the phone ringing out is a job going to a competitor, and where nobody is going to hire a night-shift receptionist to fix it. Build — market $13,000 – $40,000, ours $497 – $9,950 (under the market). Monthly — market $3,000 – $20,000 / mo, ours $1,997 – $4,997 / mo (at the market). Our published ceiling is below the bottom of the market band for the same components. Not for: Anyone wanting a chatbot bolted onto a website with no connection to a backend. That is a $99/month product, not a build, and we will tell you so on the call.

What is actually in 21-day answered phone?

• Conversation design and the agent itself — The greeting, the qualification path, the booking flow, the transfer rules and every failure path — built against your real call recordings rather than a script we imagined. Published market band for buying this on its own: $8,000 – $25,000 (A time-boxed test of whether the approach works at all, ending in a go or no-go.). Method: /tools/ai-agency-cost.
• Function calling into the systems you already run — Customer lookup, availability, booking, job creation and the text confirmation — wired natively, not by a human reading a transcript the next morning. Published market band for buying this on its own: $5,000 – $15,000 (One process, mapped, built, instrumented and handed over.). Method: /tools/automation-roi.
• An evaluation suite that runs before launch and after every change — A test suite for the agent itself, so a prompt edit cannot quietly break the booking path and a model version change underneath you cannot go unnoticed. Not sold separately by anyone in this market. No published market band exists because it is not a line item this category sells. Ask three suppliers what happens to their agent when the model version changes underneath it.
• Weekly tuning against what callers actually said — Someone reading the transcripts, finding the question the agent fumbled, and shipping the fix — not a support inbox that waits for you to notice. Published market band for buying this on its own: $3,000 – $20,000 / mo (Operating the automations: monitoring, exception handling, changes as systems shift.). Method: /tools/automation-roi.
• An hourly health watch on the number — The line is called on a schedule and the result is recorded. You find out the agent stopped answering from us, not from a customer. Not sold separately by anyone in this market. Nobody publishes a price for this because almost nobody does it. It is the difference between an agent and an agent you can leave alone.

The site an answer engine can quote — Websites

When someone asks an assistant who to call for what you do, your page is structured, dated and specific enough to be the source the answer is built from — and you can see it happening in your analytics. For: A business whose buyers have started asking a model instead of a search box, and whose current site is a brochure the model cannot parse, cite or attribute. Build — market $8,000 – $15,000, ours $8,000 – $15,000 (at the market). Monthly — market $300 – $800 / mo, ours $300 – $1,060 / mo (at the market). We are not cheaper here and do not pretend to be. A custom small-business site costs what it costs. What differs is that the AI-search structure and the build-time assertions are inside the price rather than a retainer sold to you afterwards. Not for: A single landing page for a campaign. That is an afternoon, not an engagement, and we will tell you to build it yourself.

What is actually in site an answer engine can quote?

• The build itself, as real software — Typed, tested, prerendered, fast on a phone on a bad connection. Core Web Vitals and accessibility are acceptance criteria in the scope, not a report you get afterwards. Published market band for buying this on its own: $8,000 – $15,000 (Designed rather than themed, custom components, real content structure and a CMS.). Method: /tools/website-cost-estimator.
• An entity-linked schema graph and answer-first architecture — Cross-referenced @id nodes rather than scattered snippets, question-shaped headings, and blocks that still make sense when a model quotes one out of context. Not sold separately by anyone in this market. Too new for anyone to publish a defensible band. We publish the method and the measurement instead of a price, and the citation baseline is a separate $1,000 audit you can buy on its own.
• Build-time assertions on every page — Unique title, unique h1, self-referencing canonical, description inside 140–160 characters — enforced by a build that throws rather than a checklist someone runs. This site is 1,015 pages under those rules. Not sold separately by anyone in this market. Not a line item anyone sells. It is the reason a programmatic site is still clean at a thousand pages instead of filing several hundred duplicates in Search Console.
• The care plan and the 90-day refresh cadence — Dependency and security updates, uptime watch, and a dated content refresh loop your team can actually run — because an undated page is a page an answer engine discounts. Published market band for buying this on its own: $300 – $800 / mo (Keeping it running: monitoring, small changes, a support path.). Method: /tools/ai-agency-cost.

The system you own outright — Custom software

The portal, the MVP or the internal tool exists, it is in a repository with your name on it, and the day you decide to take it in-house nothing about that is a negotiation. For: A team whose business now depends on something no off-the-shelf tool does, and who have worked out that renting it forever from a vendor is the more expensive option. Build — market $50,000 – $120,000, ours $35,000 – $150,000 (at the market). Our band sits inside the published market band rather than under it, and on a build this size a supplier quoting well under the band is telling you something about the scope they are planning to deliver. Not for: Fixed-price megaprojects with a frozen twelve-month specification. Nobody delivers those well, including us, and we would rather say so before the contract than during it.

What is actually in system you own outright?

• The platform build — Product surface, data model, auth, integrations and the deploy path. Native mobile alongside it when the product genuinely needs it rather than because it sounds better in a scope. Published market band for buying this on its own: $50,000 – $120,000 (A first sellable version of a product: auth, billing, core loop, nothing else.). Method: /tools/website-cost-estimator.
• A migration-first database with row-level security audited as a deliverable — Every schema change is a replayable migration, and the access rules are tested rather than assumed. On our own flagship engine that is 456 migrations against 239 production tables. Not sold separately by anyone in this market. Sold nowhere as a line item, and the reason is uncomfortable: most builds treat the security model as configuration rather than as something with tests against it.
• CI that blocks a bad deploy — Test guards that fail the pipeline rather than filing a warning nobody reads. You inherit the guards along with the code. Not sold separately by anyone in this market. Quoted as 'best practice' by everyone and delivered by fewer. The count is checkable on our repositories rather than asserted here.

The retyping, deleted — Automation

The job moves from intake to dispatch to invoice to follow-up without a person shepherding it, it tells you when it breaks, and there is a human escape hatch on every path it can take. For: An operator who can name the exact moment each week when somebody copies a number out of one system and types it into another, and who has stopped finding it funny. Build — market $6,500 – $17,500, ours $5,000 – $50,000 (at the market). Monthly — market $3,000 – $20,000 / mo, ours $3,000 – $20,000 / mo (at the market). We publish the same band the market publishes, because we read it off the same tables. What is inside ours that is frequently not inside theirs is the process map and the escape hatch — the parts that decide whether the thing still works in six months. Not for: Processes nobody has written down yet, and workflows that change fundamentally every month. Those need a decision, not a pipeline.

What is actually in retyping, deleted?

• A written map of the process, before anything is built — Every step, owner, trigger and exception on one page, agreed before a line of code. Automating an undefined process encodes the confusion and makes it faster. Published market band for buying this on its own: $1,500 – $2,500 / day (A small senior team with real delivery capacity behind the advice.). Method: /tools/ai-consultant-rates.
• The workflow itself, self-hosted — n8n on infrastructure you control, scheduled jobs that run without supervision, and serverless functions where the real logic lives. No per-task pricing on a platform that owns your pipeline. Published market band for buying this on its own: $5,000 – $15,000 (One process, mapped, built, instrumented and handed over.). Method: /tools/automation-roi.
• Idempotency, retries and a human escape hatch on every path — The pipeline can be run twice without doing the thing twice, it recovers from a provider being down, and any branch a person needs to interrupt can be interrupted. Not sold separately by anyone in this market. Not a line item, and the omission is where most automation projects fail six weeks after handover rather than during the build.
• Operate and improve — Monitoring on the pipeline itself, not just on the systems either side of it, plus the changes as your process changes. Month to month. Published market band for buying this on its own: $3,000 – $20,000 / mo (Operating the automations: monitoring, exception handling, changes as systems shift.). Method: /tools/automation-roi.

The citation baseline — AI search

You have a measured baseline of where you are cited today, an entity graph that makes you legible to a model, and a quarterly report showing the direction of travel rather than an assurance. For: A brand that suspects it is invisible when a model answers a question in its category, and wants to know before spending anything on fixing it. Build — market no published band, ours $1,000 (no published band). Monthly — market no published band, ours $300 – $1,060 / mo (no published band). This is the one line with no market comparison at all, and pretending otherwise would be the exact behaviour the rest of this page is arguing against. What we publish instead is the method, so you can judge the work before buying it. Not for: Anyone wanting guaranteed placement in an AI answer. There is no such product, and suppliers selling it are selling a guess.

What is actually in citation baseline?

• A citation baseline you can argue with — The questions your buyers actually ask a model, run, recorded and attributed — with the method published so you can rerun it yourself and get the same answer. Not sold separately by anyone in this market. The category is roughly a year old and nobody publishes a band worth quoting. Rather than invent one, we productized the audit at a published price and credit it against any work that follows.
• An entity and schema graph, and answer-first restructuring — The pages that matter rebuilt so a model can attribute them: cross-referenced entities, question-shaped headings, and a comparison programme that concedes where a competitor genuinely wins. Not sold separately by anyone in this market. Same reason. The nearest published figure in this space is that 97% of llms.txt files received zero AI-bot requests in May 2026 — which is a finding about the tactic, not a price.
• A 90-day refresh cadence and a quarterly citation report — Visible dateModified on the pages that carry the claims, AI-referrer traffic split out as its own analytics channel, and a report each quarter you can hold against the baseline. Not sold separately by anyone in this market. There is no market band for this and we will not manufacture one. The audit is $1,000, credited, and everything after it is quoted from what the audit found.

The costed roadmap — AI consulting

A working document with real prices in it: what to do first, what it costs, what to leave alone, and the explicit do-not-do list — not an inventory of everything technically possible. For: An owner or exec being told to 'do something about AI' who needs a ranked, priced answer they can take to a board — from someone who also ships the build and therefore has to live with the estimate. Build — market $1,500 – $2,500, ours $1,000 (under the market). Monthly — market $5,000 – $30,000 / mo, ours $5,000 – $30,000 / mo (at the market). Our published ceiling is below the bottom of the market band for the same components. Not for: Teams looking to validate a decision already made. We will tell you if we think it is wrong, which is the whole point of paying for the opinion.

What is actually in costed roadmap?

• The audit: ranked opportunities, costed, with the do-not-do list — A build-versus-buy call on each item, a data and systems readiness assessment, a measurement plan defined up front, and the honest list of things not worth doing. Credited in full against any build that follows. Published market band for buying this on its own: $1,500 – $2,500 / day (A small senior team with real delivery capacity behind the advice.). Method: /tools/ai-consultant-rates.
• Fractional ownership of the AI agenda, if you want it — Strategy, vendor selection, governance sized to your business, and internal capability building — advisory or embedded. Month to month, and it ends when you no longer need it. Published market band for buying this on its own: $5,000 – $30,000 / mo (Part-time senior ownership of the AI agenda, including vendor selection and governance.). Method: /tools/ai-consultant-rates.

Where those market bands come from

Market bands are the observed 2026 ranges published in our own benchmark tables, read off first-party vendor and industry pricing pages on 23 August 2026, and each row links to the method behind its band. Our bands are read from the same price tables published on the service pages. Neither side is typed by hand, and the verdict on each line is computed by comparing them.

Why would this go badly, and what stops it?

• "I paid an agency before and got a login and a slide deck." — That is the median outcome in this market, and nothing on a sales page distinguishes the suppliers who do it from the ones who do not. What removes it: A fixed written scope before any money moves, work done in a repository you can watch, and 30 days to take the build fee back if what ships is not what the scope said.
• "I will have no idea whether it is actually working." — Most AI deployments are evaluated by vibes, and the supplier is the only one holding the transcripts. What removes it: Evaluation suites that run before launch and after every change, and every transcript, run and outcome in your own dashboard rather than ours.
• "It will sound like a robot and embarrass me in front of my customers." — It will if nobody budgets for latency and barge-in, and plenty of builds do not. What removes it: You hear it and sign off on the script before it takes a single real call, and the latency budget is written into the scope rather than discovered afterwards.
• "I will be stuck with you." — Most of this category is built so that leaving means starting over, and that is a commercial decision the supplier made on purpose. What removes it: Source, prompts, schemas and transcripts are exportable on day one, on a month-to-month agreement with no minimum term.
• "I do not have time to project-manage this." — Every implementation that fails, fails because it quietly became the owner's second job. What removes it: You are asked for decisions, not deliverables. On the agent lane that is three touchpoints total; on a platform build it is a weekly half hour with someone who can actually decide.
• "The price will double once you are inside." — Change orders are how this industry makes its margin back, and the number on the proposal is often not the number on the invoice. What removes it: The ladder is published from $99 to $150,000, the scope document is the contract, and anything outside it is quoted before it is built.
• "This will take half a year." — On a mid-market programme it does — the published market band for one starts at a six-month minimum. On a production platform, honestly, it should. What removes it: On the agent lane, 21 days or the build fee comes back. On the lines where six months is the real answer we say six months, and the scope carries dated milestones you can hold us to.

What are you actually risking?

• Live in 21 days, or the build fee comes back. Measured from the kickoff call, against the scope document we both signed. Not from when you return the questionnaire, not from when procurement clears — from kickoff. The boundary: This one is the agent lane. A platform build is measured in months and the scope carries dated milestones instead; promising 21 days on a $150,000 system would be a lie told to win a deal. The exception on every lane is an integration you did not tell us about at kickoff.
• 30 days to change your mind. If what ships does not match the spec we agreed, we refund the full build fee and you keep the artefacts. No clawback argument, no attribution dispute. The boundary: The build fee. Retainer months already delivered are not refunded.
• You own it from day one. Source, prompts, schemas and transcripts export on request from the first day of the build. Month to month, no minimum term, no exit fee. The boundary: Third-party platform accounts stay on their own vendors' terms — we can hand you the keys, we cannot rewrite Twilio's contract.
• We will tell you not to buy. Each of the six lines publishes who it is wrong for, and those lists are real. If your volume is under a few hundred minutes a month, a $49/mo packaged receptionist is the correct purchase. If a configured off-the-shelf tool solves it, we would rather lose the build than sell you one you did not need. The boundary: This is not modesty. A client who should not have bought is a refund, a bad reference and a support burden, and we would rather lose the deal on the call.

Who should not buy this?

Five situations where we are the wrong call, published because a supplier who says yes to everyone has told you nothing by saying yes to you. Buyers who want a guaranteed revenue number. We guarantee delivery, dates and ownership. We will not guarantee an outcome we have no instrumentation to measure. Anyone who needs it live this week. A done-for-you build is measured in weeks, and on the software line in months, and we will not pretend otherwise to win the deal. Teams with an in-house engineer who wants to own the build. You will move faster than us and you should. Projects with no decision-maker available weekly. Custom software dies from unanswered questions faster than from hard problems. Regulated deployments needing signed BAAs on day one. Say so on the call and we will tell you honestly whether the timeline survives it.

Why every claim carries a commit

This site used to run case studies with invented customers and invented percentages. They were deleted. What replaced them is a portfolio of systems that exist in repositories, where each figure is published next to the shell command that produces it. If a number here cannot be checked, it should not be here.

How an engagement runs

A 30-minute scoping call with the engineer who would do the build, then a written scope with a number and a date. Builds go live in 21 business days from kickoff. Pricing is published rather than quoted on request, service is month-to-month with no notice period, and the build fee is refunded if the work is not live by the committed date.

The receptionist is one product, not the company

An AI receptionist at $497 is the cheapest way to prove the rest of the stack works, which is why it stays on the front page. It is one line item out of six. The websites, the custom software, the automation programmes and the AI-search work are the bulk of what actually gets built here.

What does TrainYourAgent actually build?

Six things: AI agents, websites, custom software, automation, AI search, and consulting. The AI receptionist is one productized agent inside the first line, not the whole company. Every line ships on the same terms — a written scope, a fixed date, and source you own.

How do I know you can build what you say you can build?

Because the counts are public. 58 repositories. One client engine with 2,245 commits in 71 days, 239 production Postgres tables, 54 edge functions, 40 scheduled jobs, and 157 CI test guards. This site is 1,015 prerendered pages, each with a unique title, a unique H1, and a self-referencing canonical.

How much does an AI receptionist cost?

$497 for the productized build, live in 21 days. Self-serve Agent Builder is $99/month if you would rather build it yourself. Larger multi-location and multi-system deployments are scoped individually and quoted in writing before you pay anything. Full breakdown on /pricing.

How long does a build take?

21 days from kickoff to live, on every service line. That covers scope, build, evaluation, and launch. If we miss the date on anything other than an integration you did not disclose at kickoff, the build fee comes back. Larger programmes are staged into 21-day increments.

What happens if I want to leave?

You take everything. Source code, prompts, database schemas, migrations, and transcripts export on day one and stay exportable. There is no proprietary runtime you have to keep paying for. Commercially there is a 30-day money-back on the build fee and no multi-year contract.

Do you use no-code tools or do you write software?

We write software. The evidence is the shape of the work: 456 ordered SQL migrations, 25 GitHub Actions workflows, 157 test guards, 54 edge functions, four native mobile apps, and a Unity 6 project with 465 C# scripts. None of that is expressible in a no-code builder.

Can you run our advertising too?

No — and that is deliberate. TrainYourAgent builds; a separate company, CurrentAds, runs performance marketing. TrainYourAgent owns a 40% stake in it. They are separate companies with separate contracts, and we will refer you rather than quietly quote you for both.

What do you not do?

We do not publish client logos, testimonials, satisfaction scores, or uptime SLAs, because we will not put a number on this site we cannot show you the source of. We also do not take a build we think is the wrong first move — the 30-day window exists so that is cheap for you to test.

Where to start

Read the work — 11 systems with 87 checkable figures. Read the services — six lines with published prices. Open a calculator and run your own numbers. Or book the scoping call. Last updated 23 August 2026.