Pavel Khaustov

Task

Build a marketplace from zero: match a person with the right coach, run the session, take the money, pay the coach — and stay inside the law while doing it.

Solution

Ran the entire project alone: research and design, then frontend, backend, database — 33 models. AI matching on vector embeddings reads what the user is actually asking for. Sessions run as in-browser video with background blur and noise suppression; group events have their own catalog, seat purchase and waitlists. Scheduling covers availability windows and exceptions; money covers packages, balances, promo codes, payouts and refunds across several acquiring providers. Plus the compliance layer marketplaces usually discover too late: published executor details with checksum-validated tax IDs.

Task

Sales and campaign data sat in an LMS that publishes no API for mailings. The numbers the company steered by were assembled by hand, and nobody could say which campaign actually brought the money.

Solution

Worked out the protocol and wrote our own access layer, then wired the data into the corporate portal and pushed payments back into analytics as offline conversions — so revenue could finally be attributed to a specific campaign. Supplier invoices are read straight out of the mailbox by a vision model and filed into the expense register. This is the part of the job that has no documentation: the system you must integrate with was never designed to be integrated with.

Task

Handle inbound sales conversations without a human on the other side.

Solution

An autonomous agent wired into the backend: it holds the conversation, checks inventory over API, and carries the customer through to the order. It runs the funnel with no operator involved and brings in $3,000–4,000 a month.

Task

Sales calls were reviewed by hand, which meant most of them were never reviewed at all.

Solution

A pipeline that transcribes call audio, analyzes the conversation with an LLM against the script and the actual outcome, then returns structured metrics for management and specific feedback for each operator. Weekly summaries flow into the company AI base, where call quality sits next to the sales funnel — so a drop in one can be read against the other.

Task

Every AI session started from zero: re-explaining the projects, pasting context by hand, losing decisions made last week.

Solution

A task manager and notes base with its own Model Context Protocol server, so the assistant reads and writes the database directly instead of being spoon-fed context. Notes arrive by voice from Telegram and a model structures them; "ask the base" answers semantically across everything written. Kanban with realtime updates on top. Integration between AI and internal systems at the protocol level, not through ready-made plugins.

Task

Repetitive outbound calls and routine operational chores eating the team's day.

Solution

Voice robots that make the calls and act on the answers, plus Telegram bots for data collection, notifications and internal routines. The boring half of operations runs without anyone touching it.

Task

Get useful answers about household money instead of typing receipts into a spreadsheet.

Solution

Receipts are caught straight out of the mailbox from four different senders, deduplicated by Message-ID and parsed line by line into product groups — so "where does the grocery budget actually go" has an answer. On top: planning with subscriptions and credit schedules, month-over-month analytics, and a RAG advisor over PostgreSQL with pgvector. Personal project, and the place where I test RAG patterns before taking them to client work.

Task

Full observability across the cluster.

Solution

Prometheus Operator scrapes metrics, Grafana visualizes, Loki + Promtail aggregates logs.