Multi-agent pipelines, LLM integrations and the product:
web apps, admin panels, trackers, Telegram bots and mini apps, landing pages.
Free task review
Send two sentences about your task and get acceptance criteria, scope and a delivery date. The material stays with you. If the task is not mine, I say so up front.
Example of a multi-agent pipeline
The process is split into roles: one agent extracts data, another scores it, a third looks for similar cases in the knowledge base. A coordinator assigns the work, and every step can be checked on its own instead of guessing what happened inside one large prompt.
I connect OpenAI and Gemini to your data and your rules: where the model answers freely, where it follows a strict template, and where it must cite a source. Rate limits, retries and fallback behaviour are part of the setup.
I index documents, policies and past conversations. Answers arrive with links to source fragments and a relevance score, so you can see what they are based on and where they might be wrong.
Prompts and pipelines break quietly: answers get worse and nobody notices. I close that gap with a test set and regression checks in CI. A prompt change does not pass if it broke earlier answers.
Full applications built around a process: authentication, roles, database, file upload, export, third-party integrations. Runs in the browser with no install and opens on a phone.
An internal tool for the team: reference data, moderation, bulk operations, role-based access, audit log. What usually lives in spreadsheets and chat threads moves into one screen.
Funnels, deadlines, workload, period comparison, export to a spreadsheet. Metrics are computed server-side so the numbers match for everyone instead of depending on one person’s export.
Contacts, deals, stages, activity history, tasks and a kanban with deadlines. Administrator, manager and contractor roles: each sees only their own work and cannot reach into someone else’s.
Intake, notifications on events and deadlines, announcement funnels, answers from the knowledge base. With caching and logging, so a failure shows exactly what came in and what went out.
An interface inside Telegram: forms, lists, cards, payments. No app install and no separate account, and the data comes from the same backend as the web version.
The bot works with your API, database and third-party services instead of living as a separate toy. One source of truth: a change in the web app shows up in the bot immediately.
Who sees requests, who changes statuses, who exports data. Restrictions are enforced on the server, not hidden behind disabled buttons.
Section order and arguments are built for a specific audience: a therapist, a coffee shop and a payment service need different ones. Proven on five live pages.
A form with notifications, export and spam protection. You see where the person came from and what they wanted, not just “new lead”.
From 320 pixels up, tested in Safari and on a real phone. No heavy builders: the page opens fast on mobile data.
Goals on form submits and clicks, analytics wired up, markup for search. So you can see whether the page works instead of assuming it looks fine.
The app ships as an image and starts with one command on any server. Rebuild and restart by script: the old container keeps serving until the new one is ready.
Reverse proxy, routing by domain, HTTPS with automatic renewal. Plus http to https redirects and security headers.
PostgreSQL with migrations, regular backups, separate staging and production environments. Migrations run by script, not by hand.
Health endpoints for the database and the app, container logs, metrics and alerts on failure. When something breaks you see what broke, not just “the site is down”.
Checks and tests in CI on every commit, image build, automated deploy. A mistake is caught before production, not after.
A README with run instructions, the data schema, environment variables and an update guide. So the project can be handed to another team, not only to me.
Filter by direction. Each row expands into the details.
The system reviews a page piece by piece: usability, structure, conversion points, copy. Several agents with different roles, merged into one report.
Why it mattersAn audit without manual reading. Notes are tied to specific blocks instead of general impressions.
Open on GitHubEvery agent has a role and an area of responsibility. Work moves along the chain from strategy to copy and creatives.
Why it mattersYou can see which step went wrong and replace that step without rebuilding the whole process.
Open on GitHubA plugin for Hermes Agent. It optimises the tool context so the same input produces the same output.
Why it mattersThis is exactly why agents in production behave differently on identical requests. Here the behaviour is reproducible.
Open on GitHubA lead processing pipeline: sources, raw signals, scoring, queue and operator actions behind one API. A separate knowledge search layer that returns its sources.
Why it mattersNo need to scroll through email and spreadsheets to decide what comes first. Operator decisions are recorded in the history.
The stand runs on a separate server and is closed to the public. I show it on a call or send access on request.
Answers are stored and compared between prompt versions. The check runs automatically on every change.
Why it mattersA prompt change does not silently break earlier answers. The break is visible before release.
Open on GitHubFirst the page is parsed, then posts are generated through the OpenAI API from the extracted facts.
Why it mattersFewer revisions after the client reads another round of “innovative solutions”.
Open on GitHubA project and task tracker: different data visibility for three roles, kanban, change history, notifications, search that handles Cyrillic properly.
Why it mattersThe team sees tasks, deadlines and workload in one place. Zero external dependencies: only the Node standard library.
Open the live standA habit tracker: one-tap check-in, current and best streak, daily heatmap, archive with history. Undoing a check-in recalculates the stats.
Why it mattersYou can see where a habit breaks. The streak does not break until the day is over, so the stats do not lie.
Open the live standAn order stage tracker: a fixed pipeline with transition history, contractor assignment, reports on the funnel, stage time and workload.
Why it mattersOverdue is measured by stage deadline rather than the whole order, so it shows up in advance.
Open the live standA therapist and coach, a coffee shop with online ordering, a payment gateway for LatAm, a marketing agency, a motorcycle school in St Petersburg.
Why it mattersA fast way to test demand: a page, a form, real requests. The different niches show that structure follows the audience.
Four steps. Each one gives you something you can touch.
I read your description and ask the questions people usually skip.
You get: acceptance criteria, scope, date
A working version, not a mockup: it stores data, computes and displays.
You get: a live link
I collect notes into a list and work through it; anything debatable we discuss.
You get: a list of changes and decisions
I hand over the repository with access and show you how to deploy it yourself.
You get: code, README, schema, one month of support
Lower bounds. The exact price comes after the task review.
from €700
A single agent, bot or pipeline, or an LLM integration into an existing process.
5 to 10 days
from €1 500
Beyond the agent you also need an interface, a database, an admin panel and team access.
2 to 4 weeks
Always included: the task review, acceptance criteria, tests, deployment, run instructions and revisions within the agreed criteria. Domain, server and paid APIs are billed separately, and I show the real amounts before we start.
You do. I hand over the repository with access and rights on your account or organisation. I do not use licences that stop you from doing what you want with it.
You can, and that is how I work. The difference is verification: a model produces the first plausible answer, while the actual work is finding edge cases, covering them with tests and making the behaviour reproducible.
Revisions within the agreed criteria are included. If the prototype shows the idea does not work, we admit it and change the approach instead of pushing it to handover.
By scope, not by hours and not by guesswork. After the review you see the breakdown: what is included, what is not, and how long it takes. If the scope changes mid-project, you hear about it before I start on it.
One month is included: I fix anything that broke because of me and answer questions about running it. After that, hourly or by subscription, agreed in advance.
Yes. I sign an NDA before the review if you need one. Commercial and NDA work I show on request and never publish.
I only use the data you hand over for the prototype and never reuse it in other projects. After handover the access is revoked and no copies stay with me.
Yes, and it is sensible. The first step is often a single pipeline or a single screen: we test the approach on a live prototype and expand from there. Less risk on both sides.
I can start this week. I give the delivery date after the review and never move it quietly: if the date slips, you hear about it immediately.
Describe the task in your own words, no technical terms needed. You get acceptance criteria, scope, an estimate and a date. The material stays with you even if we never work together.
