DESIGNER × DEVELOPER × MAKER
Design. Code. Create.
I am Zlata. I build software — a soil-mapping demo for farm fields, a ComfyUI editor with 183 GitHub stars, a parametric cabinet tool that keeps plan and elevation in sync — and I design kitchens. I work from Cape Girardeau, Missouri, remotely across the US, and I am open to AI roles and commissioned work.
Selected projects
Sites and interfaces
A bookmark manager built around spatial memory instead of folders. Saved links stay findable because you remember where they were, not what you filed them under.
Selling the plumbing, not the demo.
AI automation for small businesses in southeast Missouri. Published prices, self-hosted models, and a site that says what it costs.
Small businesses get sold a demo. A chatbot that works in the sales call, then meets a real customer with a real edge case and nobody owns it. The model is rented, the prompts live in someone else’s notes, and the invoice arrives monthly whether it works or not.
Built the practice around the unglamorous half. Self-hosted models instead of per-seat APIs, agent systems for the work that actually repeats, document and workflow automation for the paperwork nobody wants. Then hosting, security and maintenance — because a tool that dies quietly is worse than no tool.
The site is live with the prices on it, plus writing on the three automations trades businesses actually ask for: missed-call-to-text, quoting and estimating, and scheduling follow-up. One person who can answer for the whole stack.
A storefront you can stand inside.
A spatial web framework — with a working furniture showroom built on it. Four view modes over one data model.
Product pages flatten everything into a photo, a price and a cart, then ask the customer to imagine the piece in a room. 3D configurators exist, but they are a viewport bolted onto a page: one product, no sense of place.
Built the space as the interface. A scene graph of cards laid out in three dimensions, category nodes that pull forward when you click them, and a pluggable view system — universe, flat, XR and VR — all reading the same content model. Drag to orbit, scroll to zoom, click to descend.
The demo runs. Aurel & Oak: ten pieces across five categories with a live cart, navigable from an overhead universe view down to a single piece, and the same content re-renders as a flat 2D map when you would rather look at it that way.
The universe you are standing in. A Three.js scene with a three-level card hierarchy, camera orbit, and content that lives on the cards themselves rather than in a panel.
Tools and systems
One model, many views.
A browser-first parametric kitchen design workbench. Runs on the device — no server, no account.
Kitchen design software is a drawing program. Move a cabinet in the plan and the elevation, the schedule and the 3D view are now wrong, because they were never the same thing. Someone retypes the correction three times and hopes.
Built it around one parametric document. Plan, 3D scene, properties and exports all derive from the same validated model, so a change lands everywhere at once. A fixed CAD shell keeps the ribbon, catalog, both views and save state on screen together.
Phase B2 runs: an immutable catalog of base, wall and tall families with typed construction and finish options, grid and wall snapping with a wall/face/offset constraint that survives a drag, undo per committed action, autosave with a previous-good copy, and validated JSON round-trips.
One flat rate for a field that was never flat.
Satellite soil intelligence for row crops. A working demo — open it and drive the map.
Fields get one flat fertilizer rate because a field looks uniform from the road. It never is. In the demo field, nitrogen runs 36 to 57 lb/ac — same bag, every acre.
Built it as a nutrient map: soil samples in, hex grid out, five layers, a value per cell, quartile zone breaks, satellite and terrain underneath. MapLibre over live Esri tiles.
It runs. Three demo fields, 170–198 cells each; zones export as CSV or GeoJSON for a variable-rate spreader. Load your own samples — parsed in the browser, nothing leaves the machine.
A crayon shop that had to behave like a real store.
Custom crayon sets and monthly clubs — accounts, subscriptions and a checkout that fails on purpose. Built end to end in an afternoon with five parallel agents.
A demo store is easy to fake and hard to make honest. Anyone can render product cards; the tell is whether the boring parts work — a cart that holds a configured line, a subscription that survives a reload, a declined card that saves nothing.
Split the build into five workstreams with disjoint file ownership so they could run in parallel without collision. Schema and seed data first, then storefront, then cart and checkout, then the account area, with the brand system mapping the seams.
It shipped and it runs. The decline path returns 402 and leaves the cart intact, which is the detail most demo checkouts skip. Two integration bugs surfaced on first deploy — a router with no base path and a doubled API prefix — and both were visible in a browser within a minute of going live.
The deadline that quietly costs you the deal.
A dependency-graph simulator for residential transactions. Transaction software tells you when something is due; this shows what breaks when it slips.
A deal is tracked as a list of dates, but it behaves like a chain. Move the appraisal and the closing disclosure, the walkthrough and the closing all move with it — and the buyer rate lock does not. That coupling is where money is lost, and no checklist shows it.
Modelled the transaction as a dependency graph with finish-to-start and start-to-start edges, lag, and anchors that cannot move. A critical-path pass computes float per milestone, and every drag emits a diff so the cascade is calculated rather than narrated.
Dragging the appraisal 25 days late lands closing on day 46 — eleven days past the contract target — and expires the rate lock one day before closing, which surfaces as sixty-two dollars a day of carry risk at the live mortgage rate. The number is arithmetic from a real feed, not a placeholder.
The edit loop was in the wrong app.
A layered image editor that runs inside ComfyUI. Built solo, with AI assistance. 183 stars on GitHub, MIT licensed.
ComfyUI generates fast and iterates slow. Repaint a hand and you export the image, open a paint app, re-import it, re-run. Every version you liked comes back flattened — the layer you actually wanted is gone.
Put the editor inside the tab. Two nodes, sessions bound to the node, a real layer stack, non-destructive per-layer masks, brush, bucket, marquee, blend modes, and a prompt dock that remembers what you typed. Output pushes back in as a new layer.
v1.0 went fully local — no Flask server, no second browser tab, no split workflow. It is in other people’s pipelines now, and they file issues about it.
Testing what everyone else assumes.
A from-scratch research environment for small mixture-of-experts language models. Quiet for now — open source later.
Everyone building with AI uses the same handful of models and inherits their assumptions. The questions worth asking — how experts should route, whether backprop is the only learning rule worth using — sit behind compute nobody can afford.
Built a lab instead of a wrapper. One spike per idea, each proving itself before the next builds on it: custom routing, alternative learning rules, a dashboard measuring perplexity and routing behaviour on every run — trained locally, from scratch. I direct the research; my team of AI agents does much of the building.
Four spikes through to working models. Nothing released yet. It stays quiet until it is worth using rather than worth admiring, and then it goes open source.
A self-hosted agent system running a team of specialised assistants — orchestration, scheduling, memory, and tool access, all on hardware I control.
Coloring books and print
Personalized coloring books sold on Etsy — a child’s name on the cover, their interests in the pages. Twenty-odd titles and counting, each one a full book rather than a few loose pages. The part that took the work was making forty pages look like one book instead of forty separate images.
Residential design
Full white shaker kitchen: uppers, lowers, island, granite counters, tile backsplash, appliance integration. I measured it, drew the elevations, priced it, ordered it, and stayed on it through install. The hard part was never the design. It was the wall that was neither plumb nor square, and a fridge opening that had to grow two inches without moving the doorway beside it.
Cherry cabinetry with a quartz peninsula and island seating. The layout came from watching how the family actually moved through the room: garage entry, groceries down, prep at the island, plates through to the dining room. Everything else followed that path. Cherry is unforgiving — one door hung out of sequence and the grain reads wrong across the whole run.
Measure, design, quote, order, install. Ordering is where people lose months. Cabinets have one lead time, appliances have another, and the countertop cannot be templated until the cabinets are set and level. Get that sequence wrong and the kitchen sits in boxes in your garage for six weeks while you pay rent on a storage unit.
Finished installs and working drawings from the shop years. Drag to scroll through. These are jobs that got measured, drawn, ordered and fitted — not mood boards.
Designer, systems builder, optimizer
Start here
I point at the problem first, then work toward the solution. That is most of the job. If I think you are wrong about something I will explain why — and if explaining does not land, I will build the thing in front of you so you can see it. People assume I understand everything on the page. I do not. I know where to look and how to find the answer, which is a different skill and the more useful one.
I ordered an entire kitchen in the wrong colour. The fix was buying the right colour for the whole $10,000 order out of my own pocket. What changed: nothing gets ordered now until I have held the finish and door style against a physical sample. Expensive once, cheap forever.
AI should be more optimized than scaled. Everyone is racing to make models bigger, and most of the problems I get handed would be solved by making a smaller model do one thing properly. That is why I run models on my own hardware instead of renting a bigger one.
TypeScript, React and Three.js on the front. Python and Node behind it. Local models when the data should not leave the building. Fusion 360, Blender and 2020 Design when it has to be physical. I pick whatever stays maintainable after I hand it over — no framework of the month.
No straight line
At fifteen I wanted to write software. Instead I co-founded a cabinet shop in my early twenties — six people, three of us drawing and selling, three building — then design and sales, then online pickup at Walmart (yes, really), then the Menards kitchen and appliance desk, then AI systems. The software caught up with me eventually. The thread the whole way: I keep getting handed a mess and asked to make it work.
2018 — A B Cabinets: six people, three designers, three builders. Good cabinets, a few expensive mistakes. 2022 — Karter’s Cabinet Depot; 200+ project files from first measurement to install. 2024 — online pickup, then the Menards kitchen and appliance desk. 2025 — InnovAIte. Different tools every time, same job.
Small talk. Estimating how long anything will take. Remembering that not everyone wants a lecture about drawer slides. I am good at the work and mediocre at the parts around it — which is why I would rather you tell me the problem than the solution.
What's on the bench
Three things at once: this site, a parametric cabinet design tool, and a small agent team that runs on my own hardware. The last one is the interesting one — mostly because nobody has a good answer for what happens when your tools start doing the work.
Nova — a from-scratch mixture-of-experts language model lab running on a 6 GB laptop GPU. Nobody asks about it, which is part of why I like it. If an interviewer ever asks what I am most proud of, this is the answer.
Open to work — full-time, contract, or a one-off fix. Honestly: for the right price I am available for most things. Two years out I would rather be earning from something I own than selling hours, which is the plan and also not a plan yet. A friend would tell you I am an amazing addition to any team. Email reaches me faster than the phone, and I am available by appointment most days.
Start a conversation
The direct line
zlata.salyukova@gmail.com · 573-450-2522 — tell me what you are trying to make and I will tell you honestly whether I am the right person for it. Open to AI roles and commissioned work; available by appointment.
Tell me the shape
The best briefs describe the problem, not the solution. What is broken, who it affects, and what has already been tried. If you arrive with the solution already chosen, I will still ask what it is supposed to fix — that question is where the useful work starts.
A short call, a written scope with a fixed price, then weekly demos of something real. No surprise invoices, no long silence, no handover document nobody reads. You own the code and the files at the end, and I will tell you what will break first.
Where else to find me
InnovAIte Solutions for consulting. Etsy for the coloring books. Email zlata.salyukova@gmail.com or call 573-450-2522 — the fastest route to an actual answer.
Notes, ideas, stories
2026-11-19
Why I quote in person, not by email
A quote is not a price, it is a list of decisions. If I have not stood in the room, I am only quoting someone else's drawing of it.
I have quoted a lot of kitchens. I have never sent a first number without standing in the room first. That is not a boast about customer service. It costs me work. People email a floor plan, ask for a price, and go quiet when I say I need to come out and measure. Some of them are talking to three other people who will answer by email within the hour, and the one who answers fastest usually gets the first real conversation, which is not the same as getting the job but is a lot closer to it than silence. I still don't do it. ## A quote is not a price A quote is a list of decisions wearing a number as a disguise. Where the sink goes decides where the drain is. The drain decides whether the dishwasher sits left or right of it. The dishwasher decides where the run ends. The run ends at the wall, and the wall is never quite where the drawing says it is. Every one of those decisions is cheap to change on paper and expensive to change after the boxes are built. A quote that skips them is not a quote. It is a guess with a footer on it. By email, you are not pricing the room. You are pricing someone else's drawing of the room, made by a person who was also not in the room. ## Four things the room tells you None of these are in a PDF, and all four change the number. **The ceiling line.** A kitchen that looks level on a plan can drop three inches along the window wall. That changes the upper cabinet heights, which changes the door sizes, which changes the order. I have watched a perfectly good design die at the ceiling. **The door swing.** A refrigerator door needs roughly ninety degrees plus room for a hand, and a wall oven door needs somewhere to go when it is open. On a plan, a swing is an arc. In a room, it is a decision about whether two people can both be in the kitchen. **The window reveal.** A sill that projects, a blind that needs to roll somewhere, a radiator under the glass. Each one eats depth you thought you had. You cannot see any of it from above. **The route in.** A thirty-six inch refrigerator box does not go around a newel post, up four steps, and through a doorway that lost two inches to trim in 1974. If it cannot get in, the design is wrong no matter how good the elevation looks. I have also learned to look at the floor. Not to measure it — to see whether anyone has already tried to level it, and what they used. ## Email makes quotes comparable, which is the problem When every quote arrives as a total on a PDF, the only thing a client can compare is the total. So they compare the total. The cheaper one wins a job it may not be able to finish, and the difference reappears three months later as change orders, and by then the number is not a number. It is an argument. In the room, I can show my work. This is why the tall unit is fifteen inches and not eighteen. This is why there are two filler panels at the end of the run. This is why the corner is a blind corner and not a lazy susan. None of that fits in a cell on a spreadsheet, and all of it is the difference between a price and an explanation. There is a second effect. When someone watches me measure, they stop shopping and start deciding. A quote by email invites comparison; a visit invites a decision. That is not a sales trick, it is just what happens when the thing becomes specific. ## The email quotes I do send I am not absolute about this. Repeat clients, identical units in a block, someone replacing a kitchen with the same kitchen — those get priced from a drawing, because I already know the room and so do they. When I do it, I write the assumptions down. Priced from your drawing dated the fourteenth. Assumes a ninety-six inch ceiling, no services inside the west wall, existing floor level within a quarter inch. If any of those are wrong, the number moves. That paragraph exists because I have been bitten by not writing it. The most expensive mistake of my career began with a finish approved from a screen instead of held against a door in the actual light of the actual room. The screen was not lying. It was just answering a smaller question than the one I asked it. ## What it costs me I lose quotes to people who answer faster. I have made peace with that, mostly. The ones I win arrive with fewer surprises. The person has already seen the room being read out loud, so the first week of the job is building cabinets instead of building agreement. And the awkward conversations — the wall that has to move, the window that has to lose an inch of sill — happen while they are still cheap, which is the only time they are ever cheap. ## The rule No number without a visit. If I genuinely cannot visit, I say so, and I call what I send an estimate rather than a quote, and I list what I assumed. Then the client can decide whether the assumption is worth the risk, instead of finding out at installation. A quote is a promise about a room. I would rather make that promise standing in it.
2026-11-18
Six employees, no process
I co-owned a cabinet shop at twenty-two with six people on the floor and no system for anything. What that taught me about process, and why it has nothing to do with software.
I was twenty-two and a part-owner of a cabinet shop with six people on the floor. I had no idea what I was doing, and neither did the shop. We were good at the work. That is the trap. When the work is good, nobody notices that there is no system underneath it — until the week when two jobs land on the same install date, the finish order hasn't arrived, and the person who knows where the hardware is happens to be off. ## Everything lived in somebody's head Job status was in my head. The order list was in a notebook that lived in a drawer. Which supplier had which lead time was tribal knowledge, and it was wrong about a third of the time because lead times move and nobody updates the notebook. The shop ran because six people were paying close attention. That works until anyone is tired, sick, or on holiday — which is to say, most weeks. Here is the thing about a shop like that: it looks calm from the outside. Jobs go out. Clients are happy. Nobody is shouting. The failure is invisible because the slack is being absorbed by people, and people don't report that they are the buffer. They just get quieter, and then one day somebody hands in notice and takes a whole job's worth of knowledge out of the building with them. ## What we tried, in order First a whiteboard. It worked for about three weeks, until the whiteboard became another thing to maintain and someone erased a job to make room for a note about a delivery. Then a spreadsheet. Better, but only if somebody updated it, and the person with the most reason to update it was always the busiest person in the building. Then a rule, which is the only one that stuck: no job moves to the next stage without a name and a date against it. Not status, not a colour, not a feeling. A person and a date. It sounds bureaucratic. It was the opposite of bureaucratic. It took about ninety seconds per job and it ended the "I thought you were doing that" conversation, which used to happen weekly and cost hours. ## The mistake I made for the first year I treated every one of those fixes as a communication problem. The whiteboard was going to fix communication. The spreadsheet was going to fix communication. I kept buying better surfaces to write the same missing information on. The problem was never the surface. It was that no job had an owner. Not a team, not a stage, not "the shop" — a person whose name was on it. Once a name was attached, the communication sorted itself out, because there was someone who would notice. That is the mistake I see small businesses make constantly, including mine: they go looking for a tool to fix a problem that is actually about who is answerable. A tool cannot assign responsibility. It can only record it after a human has decided. ## Process is not paperwork People hear "process" and picture forms. What process actually is: the part of the work that still happens correctly when you are tired. A shop with no process is not a flexible shop. It is a shop held together by whoever is paying the most attention that day, and it degrades quietly every time that person is unavailable. The test I use now is simple. Pick any job in the building and ask: if the person who usually does this step were off sick tomorrow, would it still happen, and would anyone know it had been missed? If the answer is no, that is not a personality failing of the person. That is a hole in the shop. ## Why this matters if you want to automate anything You cannot automate a process that only exists in someone's head. You can only codify chaos, and codified chaos is worse than the original — it now fails silently, at machine speed, without anyone noticing. So when someone asks me to automate their quoting, my first question is not what tools they use. It's: show me the last quote you wrote, step by step, including the parts you did from memory. Almost every time, the answer reveals that the process isn't one process. It's three different versions depending on the day, the customer, and how busy the shop is. Automating any of those versions would be automating the wrong one. The tell is always the same: when the person describes the process, they say "usually" and "normally" and "if it's a big one". Those words are where the exceptions live, and the exceptions are the work. ## What I do instead One week of notes. Not a workshop, not a discovery engagement — a notebook and the discipline to write the job down as it happens, every step, including the stupid ones. At the end of the week you have something no consultant can sell you: the actual process, with all its exceptions, in your own handwriting. Then — and only then — you can decide what to remove, what to simplify, and what to automate. Half the time the answer is that three steps disappear and nothing needs building at all. That is not a disappointing outcome. That is the cheapest possible fix, found before anyone wrote a line of code.
2026-11-17
The same kitchen, drawn four times
A kitchen gets drawn four times before it is built, and every redraw is a chance to introduce a typo nobody checks. So I am building a tool that draws it once.
A kitchen gets drawn four times before anyone cuts a board. Once in the design software, where the client approves the look. Once as a shop drawing, where the millwork actually gets dimensioned. Once on the order sheet, where the parts get specified. Once as a render, because the client needs to see something that looks like a photograph. Four documents. One kitchen. And nobody checks any of them against the others. ## The problem is not the drawing Designers talk about this as a rendering problem or a workflow problem. It isn't. It's a re-keying problem. The same cabinet run is retyped into four systems by a human being, usually at the end of a long day, usually under time pressure. The first drawing is thoughtful. The fourth is transcription. Transcription is where errors live. Not because anyone is careless, but because retyping the same number is not work the brain treats as important. It's the reason a cut list can disagree with a plan that everyone already approved, and nobody notices until the shop floor does. I have watched a $10,000 order go wrong because a finish was approved against the wrong evidence. That one was mine. This one is more common and less dramatic: the parts are right, but the list is wrong, and the mistake surfaces at the saw. ## What I'm building One model. Everything else generated from it. Change the model, the plan changes, the elevations change, the cut list changes, the render changes. No re-keying, because there is nothing to re-key into — the documents are outputs, not copies. The stack is unglamorous on purpose: Python for the geometry and the cut list logic, Three.js for the 3D view, running locally. It works on my machine, which is the only claim I'm making today. ## The render was never the hard part Everyone assumes the 3D is the difficult bit. It isn't. The 3D is the part that impresses clients and takes a weekend. The cut list is the hard part. A cut list is not a list of parts. It's a list of parts *under constraints*: sheet yield, grain direction, edge banding, kerf, the order the shop prefers to cut in, and which offcuts are worth keeping. Two cabinets that look identical in a render can have completely different cut lists, because one has a grain direction that forces a different sheet layout. Get that wrong and the model is a toy. Get it right and the shop stops double-checking your paperwork, which is the actual product. ## What this is not It is not a product. It is not for sale. There is no pricing page and no waitlist. It is also not finished. Right now it's held together by hope and one very long function I'm embarrassed by, and I'd rather show that than pretend otherwise. What it is: the tool I wanted when I was designing kitchens and retyping the same run for the fourth time at 7pm. ## If you do this by hand I want the edge cases I haven't hit. Which constraint bites first in your shop — grain direction, sheet yield, or lead time on a specific door style? Where does your cut list disagree with your drawing most often? What does the shop do when the plan and the parts list argue? If you have a workflow you'd defend, I'd like to hear it. That is not a pitch. I'm building the thing and I'd rather build it against a real constraint than a guessed one.
2026-11-16
I ordered a $10,000 kitchen in the wrong colour
I paid to replace the whole order myself. What that taught me about where mistakes are cheap to catch — and where they stop being recoverable.
The truck arrived on a Tuesday with twelve cabinets, two islands, and a full run of uppers in a colour the client had never actually seen. She had seen a swatch. She had seen a render. She had not seen a door. I had — in the shop, under fluorescent light, propped against a grey wall, where it read as a warm off-white. Her kitchen faces north. In that room, at eight in the morning, the same finish goes green-grey. ## Nobody did anything wrong except me The showroom lighting was normal. The swatch was accurate. The render was accurate. The order was correct against everything I had checked it against. I had checked it against the wrong things. Approval was mine. That is the whole story, and it is the part worth keeping: not that a colour was misjudged, but that I signed off on a decision while holding evidence that could not possibly show me what I needed to see. ## What it cost Ten thousand dollars to replace the order. I paid it personally. Not the supplier, not insurance, not the manufacturer — me. I could have argued. There was a defensible case that the swatch was misleading, that the lighting was misleading, that a reasonable professional would have made the same call. Arguing would have cost the client another three weeks and cost me the relationship, and the relationship was worth more than the cabinets. So I replaced them, installed them, and wrote the number down where I would see it again. ## The rule that came out of it Nothing gets ordered now until I have held a physical sample in the real room, in the real light, at the time of day the client will actually be standing there. Not a screen. Not a render. Not a swatch book. The thing itself. That rule has cost me maybe two days of driving over the years. It has saved considerably more than that. But the rule is the small lesson. The bigger one is the question underneath it. ## Where is the cheapest place to catch this? Every mistake has a price, and the price depends entirely on *when* you catch it. In a kitchen: - A line on a drawing: cheap. Erase it. - A cut list: expensive. Everything downstream is now wrong. - A delivered cabinet: brutal. You are eating ten thousand dollars and a client's patience. The error was identical at every stage. The cost was not. In software it maps exactly: - A type error: seconds. - A failing test: minutes. - A code review comment: an hour. - A customer finding it at two in the morning: days, and trust. Same bug. Wildly different bills, depending on where it surfaces. So the question I ask before building anything is not "how do I avoid mistakes?" You will not. The question is: **where does this get caught, and can I move that point earlier?** ## How it changed what I build This is the reason I build parametric tools now instead of drawing the same kitchen four times. When the model is the source of truth and the plan, the elevation, the cut list and the render are all generated from it, a change propagates instead of being retyped. Four documents become one decision. The catch point moves from the shop floor back to the model, where a fix takes seconds. That is not a technology argument. It is the same argument as holding a door against a north-facing window: move the moment of discovery to the cheapest possible place. ## What to do with this Find the most expensive step in your own work — the one where being wrong hurts most. Then walk backwards from it and ask where that same error could have been caught instead. Then put a check there. A person, a form, a second pair of eyes, a sample held in the actual room. It is unglamorous and it will feel like slowing down. It is also the cheapest thing you will ever do. I still think about that Tuesday. Not because of the money — because it taught me the only question I ask before building anything: where does this get caught, and can I move that point earlier?
Journal archive
Every post, in one place.
The archive — newest first. Only the latest five live in the 3D universe.
Browse every post. Test cards will be deleted once real writing starts.
Email zlata.salyukova@gmail.com or call 573-450-2522.
Taking on select projects: AI automation, custom software, and kitchen or bath design. Small enough to move fast, experienced enough to know what breaks. Available by appointment most days — email reaches me faster than the phone.
Cape Girardeau, Missouri. I work remotely across the US and travel when a project genuinely needs someone in the room — site measurements and installs, mostly. The shop years were all within an hour of here, which is why I know which suppliers actually deliver on time.