The Agents Answer at 11 PM: Cyrus Sandoval and the Always-On Sales Team of Malolos

It is 11 PM at a resort in the Philippines, and the front desk is empty. The staff have gone home. The lights in the lobby are low. And on the resort's website, a small chat window blinks awake, because a guest somewhere is still up, phone in hand, trying to decide whether to book.

The guest types a question. Maybe it is in English. Maybe it is in Taglish, that easy everyday blend of Tagalog and English that Filipinos slip into when they are talking to someone they expect to understand them. Either way, the answer comes back in seconds. Not in the morning. Not after someone checks the inbox and apologizes for the delay. In seconds. The conversation continues, question by question, until the guest is guided all the way to a confirmed booking.

There is no night shift. Nobody is sitting in a dark office drinking coffee and waiting for the chat window to blink. The thing doing the answering is an AI sales agent, installed on the resort's site, trained on what the resort offers, and it does not sleep, does not take breaks, and does not care that it is 11 PM.

Now here is the part that matters. That agent was not shipped by a giant software company. It was built and installed by a specialist working remotely from Malolos, Bulacan, a city in the Philippines, as part of a small, focused business that does a few things very well: AI video ad creative, print-on-demand pipelines, AI-built websites, and AI sales agents that answer customers around the clock. His name is Cyrus Sandoval, and this post is about how his work actually functions, piece by piece, because it turns out to be one of the clearest working examples of what happens when AI stops being a demo and starts being a system that real customers talk to every night.

A Remote Worker Since 2013, an AI Specialist Since 2024

Start with the timeline, because it explains a lot. Cyrus has worked remotely since 2013. That is more than a decade of delivering work to clients he mostly does not sit next to, which means his entire career has been built on the thing remote work runs on: systems. When your client is in another city or another country, you cannot tap them on the shoulder. The work has to arrive through a process. The process has to be reliable. If the process breaks, there is no walking down the hall to fix it.

Then, in 2024, he went deep on AI. Not dabbling, not posting about it, but rebuilding his practice around it. Today he describes himself as an AI video and automation specialist for ecommerce brands, and he can start immediately once terms are agreed, which is the kind of sentence a person writes when their calendar is a business and not a hobby. You can read his background, his stack, and his client work on his AI Pro profile, which is where this story starts.

His toolkit is worth listing carefully, because each piece has a job. The automation layer is n8n and Make.com, two platforms that connect apps and services into automated workflows. Underneath sits Airtable, acting as the organizing layer, the place where the information a workflow needs is kept in order. The creative layer is where the newer AI tools live: VEO 3 and Kling AI and Nano Banana for video and image work, and Lovable.dev for building websites. And across everything, ChatGPT and Claude, the general purpose AI assistants that help write, reason, and assemble. Nothing on that list is exotic. That is the point. The interesting thing about Cyrus's work is not that he uses rare tools. It is how he has arranged ordinary tools into machines that keep running after he logs off.

His client and project names, as listed on his profile, give you the shape of the practice: Ruzzoval, his AI sales agent offer; InnQuire PH, which sells his agent to resorts; Balai Kabute Farm, a Filipino food brand; and Ecommercify, another of his AI-built sites. A resort business, a food business, direct-to-consumer ecommerce brands. These are not technology companies. They are ordinary businesses with ordinary problems, and the problems are what his systems are built to eat.

The Pattern: Repetitive Production Becomes a System

Before we look at each machine he has built, it helps to see the pattern they all share, because once you see it, you start seeing it everywhere.

Most small businesses are full of jobs that look like this: the same task, done by hand, over and over, with small variations. A customer asks a question that a hundred customers have asked before. A new product design needs a mockup, and then another design needs the same mockup, and then another. A brand needs a short video ad, and then it needs ten more, each a variation on the same theme. None of these jobs are hard, exactly. Each one is a known sequence of steps. But they eat hours, and the hours they eat belong to people who could be doing the parts of the business that actually need a human.

There are two ways to deal with a repetitive job. The first way is to do it faster. Work harder, hire help, stay up later. The job still passes through human hands every single time, and it stops the moment the hands stop. The second way is to define the job once, carefully, as a system: here is what comes in, here is what happens to it step by step, here is what comes out. Then you let the system do the job from then on. The human's work moves up a level, from performing the task to designing the task and checking its output.

That second way is the whole of Cyrus's business. His profile frames it plainly: across all three of his service lines, the pattern is the same. A repetitive production job that used to eat hours, running as a system instead. The systems differ. One makes video ads. One publishes products. One answers customers. But each one takes a job that used to consume someone's day, every day, and turns it into infrastructure: something the business owns, that runs whether or not anyone is watching.

There is a useful question hidden in this pattern, and it is worth pausing on, because it is the question Cyrus's whole practice answers. The question is not, "Can AI do this task?" For most repetitive tasks, the honest answer in 2026 is yes, more or less. The better question is, "Can this task be handed to a system that a real business can trust at 11 PM, when nobody is supervising it?" That is a much higher bar. It requires the system to know its limits, to be fed the right information, and to be connected to the rest of the business so its answers have consequences in the real world, like an actual booking. Everything Cyrus builds is aimed at that bar.

The Stack, Layer by Layer

It is worth slowing down on the toolkit, because people often copy tools when they should be copying architecture. Cyrus's stack has four layers, and each layer answers a different question.

The first layer is the workflow layer: n8n and Make.com. These are the conductors. A workflow platform watches for a starting event, a new design arriving, a message coming in, a schedule firing, and then carries the work through a sequence of steps, handing data from one service to the next. When Cyrus says a pipeline creates ad creative or publishes products, the conductor is what makes the steps happen in order, every time, without someone clicking through them.

The second layer is the memory layer: Airtable, which his profile names as sitting underneath. Workflows are forgetful by nature; they run and finish. A business needs records that persist: which designs have been processed, which products are live, which customers the agent has talked to. The memory layer is where the system's history lives, in tidy rows that a human can inspect. Skip this layer and your automation is a rumor. Include it and the automation becomes auditable, which is what makes a client comfortable letting it run unsupervised.

The third layer is the generation layer: VEO 3, Kling AI, and Nano Banana for video and images, and Lovable.dev for websites. This is the layer that makes new things: the moving footage in an ad, the look of a page, the face of a store. It is the newest layer and the one that changed most recently, because generation tools crossed a quality threshold where their output stopped being a curiosity and started being sellable. Notice, though, that generation is only one layer of four. In the popular imagination, AI is the generation layer and nothing else. In Cyrus's practice, generation is a component inside systems whose other layers are older, duller, and just as necessary.

The fourth layer is the reasoning layer: ChatGPT and Claude, used across everything. Writing, drafting, answering, assembling: the general purpose models that handle language wherever language needs handling. In the sales agents, this layer is the voice. In the pipelines, it is part of the machinery that turns inputs into sensible outputs.

The lesson of the stack is that none of these tools is the business. The business is the arrangement: conductor, memory, generation, reasoning, wired together around a specific repetitive job for a specific kind of client. Another operator with the same four subscriptions and no arrangement has nothing. Cyrus's decade of remote work shows here, because arranging systems so they run without you in the room is precisely the skill that remote work has always demanded.

Machine One: A Video Ad Pipeline That Starts With a Flat Design

The first machine makes video ads, and the best way to understand it is to follow one job through it, from the moment a client hands something over to the moment a finished ad comes out.

The input is humble: a flat t-shirt design. Not footage. Not a photoshoot. Just the design file, the same flat image a print shop would use. The output, 72 hours later, is a UGC-style lifestyle video ad: the kind of short, casual, person-on-camera video that fills social feeds and sells products precisely because it does not look like a traditional commercial. In between the flat design and the finished ad sits an automated pipeline built on n8n, with the current generation of AI creative tools inside it: VEO 3, Kling AI, and Nano Banana doing the visual generation work.

Think about what that replaces. The traditional way to get a lifestyle video ad for a t-shirt involves finding a model or a creator, shipping them the product, waiting for them to film, reviewing the footage, asking for changes, and editing. It takes days to weeks, costs real money per video, and produces one ad at the end. Brands that advertise seriously do not need one ad. They need a stream of ads, because ads wear out, audiences get bored, and the only way to find the version that works is to make several and test them. The manual process makes that stream expensive and slow. Cyrus's pipeline exists to make the stream cheap and fast: his n8n automation creates video ad creative for direct-to-consumer ecommerce brands in a fraction of the manual time.

The details on his offer page show how productized this is. A client sends the flat design. The turnaround is 72 hours. And here is the detail that tells you he believes in his own method: the offer page itself, the page describing all of this, is one of his Lovable builds. The page selling the AI pipeline was built with AI site tools. The medium is the portfolio. Samples of the output live in his Drive portfolio of AI ad creatives, so a prospective client can look at finished work before paying for anything.

Notice what the pipeline does and does not do. It does not eliminate taste. Someone still has to decide which designs deserve ads, look at the output, and judge whether the video feels like the brand. What it eliminates is the production bottleneck between the decision and the finished ad. The repetitive middle, the filming and assembling and rendering that used to eat the hours, is what got systematized. The human judgment stays at the ends, where it belongs: choosing what to make, and approving what was made.

Machine Two: Print on Demand, With Nobody Touching the Store

The second machine is quieter than the first, and in some ways it is the purest example of the pattern. Print on demand is a simple business model: a design is printed onto a product, a shirt or a mug or a poster, only when a customer orders it. No warehouse of pre-printed stock, no guessing which designs will sell. But between "we have a design" and "the product is live in the store" sits a pile of small, tedious production work. Each design needs mockups: realistic images showing what the finished product looks like. Each mockup needs to be attached to a product listing. Each listing needs to be published, correctly, in the store.

Do that for one design and it is a chore. Do it for fifty designs, across more than one storefront, and it is a job that swallows afternoons whole. It is also, crucially, a job with no surprises in it. The steps are known. The inputs are known. It is repetitive production in its most literal form.

Cyrus's Make.com automation takes that whole pile and runs it as a system. A design goes in one end. A listed, mockup-complete product comes out the other, published to Shopify and Etsy stores automatically, with no one touching the store in between. The automation generates the mockups and handles the publishing. Airtable sits underneath as the organizing layer, keeping track of what exists and what state it is in, because a system that produces things needs a memory, somewhere orderly to record what it has done.

There is something worth understanding about why the memory matters. When a human does the listing work, the record of what has been published lives in that person's head and in the store itself. When a system does it, the system needs its own ledger, or nobody can answer simple questions like: which designs are live, where are they live, and what has not been processed yet? That is the unglamorous backbone of every automation Cyrus builds, and it is the difference between a clever trick and a business process. A trick works once. A process, with a ledger underneath it, works every day and can be checked by anyone.

Put the two machines side by side and the family resemblance is obvious. The video pipeline and the print-on-demand pipeline both take creative production work, the kind that used to require a person at a desk pushing files around for hours, and convert it into flowing systems. One produces ads. The other produces store listings. Both free the humans involved to spend their hours on the parts that cannot be systematized: designing things people want, running the brand, talking to customers who have unusual problems. Which brings us to the third machine, the one that talks back.

Machine Three: Sales Agents That Answer Real Customers

The third line of Cyrus's work is the one this post is named for, and it is the most demanding of the three, because this machine does not handle files. It handles people.

The offer is called Ruzzoval, and it works like this. An AI direct-message sales agent is installed on a store. It is trained on the store's catalog, meaning it knows the products: what they are, what they cost, what questions people ask about them. It is plugged into the store's CRM, the customer relationship system, so the conversations it has are connected to the business's records rather than floating free. Then it goes to work around the clock. It qualifies leads, sorting the casually curious from the genuinely interested. It answers product questions. And it recovers carts, reaching back to shoppers who put something in a cart and drifted away, which is one of the most common ways online sales quietly die.

All of that happens, in his profile's phrase, without adding a single headcount. That phrase is the economics of the whole thing in five words. A store that wants human coverage at 11 PM, on weekends, on holidays, pays for that coverage in salaries and scheduling and management, forever. An agent that covers those hours is a system: built once, trained on the catalog, and then running. On Cyrus's own offer page, the agent even has a name and a visible presence: Helen, greeting visitors from the corner of the page, doing in public exactly what she would do on a client's site.

But an offer page is just an offer. What makes Cyrus's version of this story worth a long post is that the agents are not hypothetical. They are installed, live, on real client sites, answering real customers. His profile is explicit about this: the same agent runs on his client sites. Let us look at the two deployments his profile names, because they show the two ends of what these agents do.

The Resort: Guests Who Message at 11 PM

InnQuire PH is the vehicle that sells Cyrus's agent to resorts, and the problem it solves is stated in one clean image: guests message at 11 PM. Anyone who has traveled knows this is true. People plan trips at night, after work, after dinner, scrolling and dreaming. That is when the questions come. Is there availability? What does the room include? How far is the beach? Can we check in late? A resort that answers those questions tomorrow morning is answering a guest who has already booked somewhere else, somewhere that replied.

With the agent installed, the guest messages at 11 PM, in English or in Taglish, and gets an answer in seconds. The agent replies, keeps the conversation going, and guides the guest to a confirmed booking. The booking part matters. This is not a machine that says "thanks for your message, we will get back to you," which is just a faster way of losing the sale politely. It walks the guest through to an actual reservation. And the InnQuire PH site itself, the site selling this service to resorts, is one of Cyrus's AI builds, which means the sales pitch for the AI agent is itself delivered by an AI-built website with the agent on duty. The proof and the pitch are the same object.

The Farm: Mushroom Bagoong and an Assistant Named Balai

The second deployment is wonderfully unglamorous, which is exactly why it is convincing. Balai Kabute Farm is a Filipino food brand selling mushroom bagoong nationwide. Bagoong is a beloved Filipino condiment, traditionally a fermented paste, and this farm's version is made from mushrooms. This is not a tech startup. It is a food business, selling a regional staple to shoppers across the country, the kind of business that runs on reputation, repeat orders, and questions: how much, how do I order, do you ship to my province, how long does it keep.

The farm's site was built with AI, and it has its own AI assistant, named Balai, answering shoppers. Think about what that means operationally. A small food brand cannot staff a support desk around the clock. Its customers, though, shop around the clock, from phones, at all hours. Before the agent, every question that arrived after hours waited, and some fraction of those shoppers simply bought from someone else. With Balai on duty, the questions get answered when they are asked. The same pattern as the resort, wearing different clothes: the repetitive job (answering known questions about a known catalog) became a system, and the system's hours are all of them.

Ecommercify rounds out the list of his live AI-built sites. Taken together, the client work shows a consistent shape: Cyrus builds the site with AI tools, installs the agent he sells, and connects it so the conversations matter to the business. Build, install, connect. The website is the storefront, the agent is the staff, and the CRM is the memory.

Why 11 PM Is the Most Honest Hour in Ecommerce

Think about when online shopping actually happens. Stores keep office hours in their heads, but customers do not. People shop from bed, from night buses, from the couch after the kids are asleep. They plan trips late, because late is when the day finally gives them a quiet minute. The questions a business receives at 11 PM are not edge cases. For many stores, they are a large share of the most motivated questions of the day, asked by people who are ready to buy and are simply trying to confirm a detail before they do.

Now consider what the traditional store does with that moment. It lets the question sit. An auto-reply, if there is one, promises a response within one business day, which is a polite way of saying the moment will be over. By morning, the guest has booked the resort that answered, the shopper has found the jar of bagoong somewhere with a faster reply, the cart has been forgotten. Nothing dramatic happened. No complaint was filed. The sale just evaporated, silently, the way most online sales are lost: not to a better product, but to a faster answer.

This is why the sales agent matters more than its modest job description suggests. Answering product questions sounds like customer support, a cost center. But at 11 PM, answering a product question is sales, performed at the exact moment of decision. The resort agent that guides a guest to a confirmed booking is not deflecting tickets. It is closing. The farm assistant that answers a shipping question is not reducing workload. It is rescuing revenue that the business was losing every night without ever seeing it leave. When you reframe after-hours answers as after-hours selling, the economics of building the system stop looking like an experiment and start looking like arithmetic.

There is a dignity point here too, for the businesses involved. A small farm or a family resort cannot hire a night team, and should not have to choose between sleeping and selling. The always-on agent is what lets a five-person business respond like a five-hundred-person business during the hours when its customers are awake and its staff should not have to be. That is a genuinely new capability in the history of small business, and it arrived not as a robot in a warehouse but as a chat window that answers politely in two languages.

Live, Not Demos: A Standard Worth Stealing

His profile makes one claim about all of this that deserves to be pulled out and examined, because it is the difference between the AI work that fills social media feeds and the AI work that fills order books. The claim is that these are live systems answering real customers, not demos.

A demo is a performance. Somebody builds a shiny thing, rehearses the exact questions it will be asked, records it succeeding, and posts the video. Demos are not lies, exactly, but they are sheltered. A demo never meets a tired guest typing one-handed in Taglish at 11 PM. A demo never gets the question that is not in the catalog, the customer who changes their mind mid-sentence, the shopper who asks about shipping to a province the store has never shipped to. A demo answers the questions its builder chose. A live system answers the questions strangers choose, and strangers are creative in ways no builder anticipates.

That difference changes what you have to build. A demo needs to be impressive once. A live system needs to be right every day, needs to know the catalog accurately, needs to hand the business conversations it can actually use, and needs to keep working after the builder has gone home. The standard also changes what the work is worth. Businesses do not pay monthly for demos. They pay for the guest who books at 11 PM who would otherwise have booked a competitor, and the shopper whose question got answered before she closed the tab.

The standard cuts the other way as well: it disciplines the builder. When your work will face strangers tonight, you build differently. You care about what the agent does with a question it cannot answer, because a stranger will eventually ask one. You care about the catalog being current, because the agent will quote it with total confidence. You care about the record the conversation leaves, because the business will act on it tomorrow. None of those concerns exist in a demo. All of them are the daily substance of a deployment. This is why experienced operators talk about live systems the way sailors talk about the sea: with respect for the gap between rehearsal conditions and real ones, and with pride that their work holds up in the real ones.

There is a second, quieter honesty marker in Cyrus's setup: his offer page does its selling with the agent visible in the corner, doing the job in front of the prospect. And his video ad service is sold on a page built with his own AI tools, with samples of finished ads available to inspect. The pattern is show, then tell. In a field full of people describing what AI could do for your business, he has arranged his business so that the first thing you encounter is what it already does for his clients.

What an Ecommerce Operator Can Steal From This

You do not need to hire Cyrus, or anyone, to learn from the shape of his work. The shape is the lesson, and it generalizes to almost any small business that sells things online. Here it is, step by step, in the order his own practice demonstrates it.

1. Find the job that repeats, not the job that impresses

The candidates for systematizing are rarely the exciting jobs. They are the jobs someone on your team does for the tenth time this week: building the same kind of mockup, cutting the same kind of ad, answering the same dozen questions. Cyrus's three machines all came from this kind of inventory. Video ads that DTC brands need in a constant stream. Listings that every new design needs. Questions that every store gets every day. If a job arrives on a schedule and follows a pattern, it is a system waiting to be built. If it happens once and needs judgment throughout, keep it human.

2. Your catalog is the curriculum

The reason his sales agents can answer questions is that they were trained on the catalog. That sounds obvious, but it is the step people skip. An agent that knows your products, your prices, your policies, and your delivery areas can be genuinely useful at 11 PM. An agent with generic knowledge and a friendly tone is a liability with good manners. Before automating a conversation, write down what your customers actually ask and make sure the answers exist somewhere the system can learn from. The unglamorous preparation is the product.

3. Connect the system to the business, or it is theater

His agents are plugged into the CRM. His print-on-demand pipeline keeps its ledger in Airtable. His resort agent does not stop at chatting; it guides guests to a confirmed booking. In each case, the automation touches the real machinery of the business: records, listings, reservations. An automation that produces output nobody uses, in a place nobody checks, will quietly rot. Connect the ends. The conversation should create a record. The pipeline should publish to the actual store. The output should land where the business already lives.

4. Sell the outcome, and let the system be the proof

There is a sales lesson in here too, for anyone who builds these systems for others. Cyrus does not lead with model names. He leads with outcomes a business owner can picture: a video ad from your flat design in 72 hours, your products listed without anyone touching the store, your guests answered at 11 PM. And then, when the prospect looks closer, the proof is structural: the page was built with the tools, the agent is sitting in the corner ready to be tested, the samples are there to watch. If you build systems for clients, the most persuasive portfolio is a business of your own that visibly runs on them.

5. Keep human judgment at the ends

Nothing in his practice removes people from the work that needs people. Taste still decides which designs become ads and which ads represent the brand. The businesses still run their resorts and their farm. What moved to the machines is the repetitive middle: production steps and known questions. That is the correct division of labor, and it is worth stating plainly because so much writing about AI gets it backwards, fantasizing about removing the human from the judgment calls while leaving humans doing the data entry. Cyrus's systems do the opposite. The data entry went to the machines. The judgment stayed with the people.

One more honest note. Building a system is real work up front. Someone has to define the steps, prepare the catalog, connect the CRM, check the outputs, and fix the cases the system gets wrong. The payoff is that the work is done once and the benefit repeats every day, including at 11 PM on a holiday weekend. That trade, effort now for reliability forever, is the entire economics of automation, and it is why the operators who think in systems end up with businesses that feel bigger than their headcount. Their stores answer while they sleep. Their listings publish while they work on something else. Their ads arrive on a schedule instead of on a hope. The business stops being a pair of hands and becomes a set of well-tended machines with a person in charge of them, which is a much better job, and a much more valuable business, than the one it replaces.

Why this belongs at Starvard

Starvard collects stories about people who build real things with AI, and Cyrus Sandoval's work is the real thing in its most verifiable form. Not a concept, not a pitch deck, not a demo recorded under laboratory conditions: production systems, running tonight, on the sites of a resort business and a mushroom bagoong farm and direct-to-consumer brands, answering strangers in English and Taglish and turning flat design files into finished ads. His decade of remote work matters to the story, too. He learned systems the way remote workers learn everything, because his livelihood depended on processes that work when nobody is in the room. When AI tools matured, he did not have to change his philosophy. He just got better machines to build with. You can see the full picture, his stack, his services, and his client work, on Cyrus Sandoval's AI Pro profile.

There is also a lesson here that goes beyond ecommerce. The public conversation about AI agents tends to swing between magic and menace, but the working reality looks like Malolos at 11 PM: somebody's guest gets a fast, correct answer, somebody's farm sells another jar, somebody's design becomes an ad while its creator sleeps. Ordinary businesses, ordinary hours made extraordinary by being covered. If you want to meet builders like Cyrus, people quietly putting AI to work in real businesses, they gather in the E2 Discord. Come and see the work for yourself: join the E2 Discord.

Watch

n8n AI Agent Tutorial for Beginners by The NextGen Ai Frontier: a beginner-friendly walkthrough of n8n, the automation platform underneath Cyrus's video pipeline, showing how workflows and AI agent nodes turn a chain of steps into a job that finishes on its own.

Turn Any Image into Viral AI Ads | Consistent Characters & Products Ads Tutorial by Robert’s Tech Toolbox: a hands-on look at turning a single product image into AI video ads with a consistent character and product, the same flat-image-in, finished-ad-out shape as Cyrus's ad pipeline.

ZenTalk Store: Connect Shopify or WooCommerce So the AI Sells For You by ZenTalk: a step-by-step guide to connecting a real store so an AI sales agent can answer live product, stock, and order questions, which is the same basic pattern as the agents answering Cyrus's clients' customers.

Sources

Cyrus Sandoval's AI Pro profile, with his background, automation stack, services, and client list: aipro.fm/cyrus.

Ruzzoval, his AI direct-message sales agent offer, where the agent Helen greets visitors: ruzzoval.com.

Client work named on his profile includes InnQuire PH, which offers his agent to resorts; Balai Kabute Farm, a Filipino food brand selling mushroom bagoong with his assistant Balai on duty; and Ecommercify, another of his live AI-built sites.

Reporting for this post comes from the AI Pro profiles maintained by the E2 Discord community, which document each builder's background, tools, and live client work.