The backend — and the AI — businesses run on.

DEMRA — an in-house AI data analyst

An AI agent that lets anyone ask a half-terabyte database questions in plain English.

DEMRA plans its own steps, writes and validates ClickHouse SQL, runs statistical analysis in Python, and narrates an answer anyone can read. It's the third iteration of an idea I kept alive: my first attempt — a self-hosted model fed raw data — burned ~$2,000 a month before proving the approach wrong, and the third-party agent that replaced it was abandoned by its vendor. So I built the real thing in-house, partly on my own time, tuned its SQL generation with prompt evaluations, and it has served production since — at API-call prices.

• Python
• Gemini
• ClickHouse
• LLM agents
• Prompt evaluation

Clara — the software that runs an online school

One platform runs an entire online school — and its owner, not a vendor, holds the keys.

An online English school runs its whole business on software I designed, built, and deployed solo: consultations, enrollments, lessons, subscriptions, and revenue in one place. On top sits an AI layer that turns each bilingual (Russian↔English) consultation into a structured read — the client's goals, action items, the student's actual mistakes — for about $0.04 a session, and an MCP server that answers plain-language questions about the whole school. All of it runs on infrastructure the owner controls: their own database, storage, and login — software they own, not rent.

• Node.js
• React
• MySQL
• Deepgram
• OpenRouter
• MCP
• Docker

AI invoice processing, Xero-integrated

Replaced manual invoice entry with an AI reader — for under €2 a month.

When the proof-of-concept's maintainer left the company, everyone expected a months-long rebuild. I shipped it inside the main product in about two weeks: drop in an invoice PDF or photo, an LLM reads it into structured fields you confirm in a click, and it syncs straight to Xero. The transcription engine underneath became a shared building block the rest of the team now reuses for their own AI features.

• Gemini
• OCR
• PHP
• Xero
• API integration

Analytics over half a terabyte

Nearly five years on one high-traffic ad-tech platform — building the statistics its users live in.

I build the dashboard's statistics features — the views where users filter, group, and sort a half-terabyte ClickHouse dataset by any dimension they can think of — on an analytics layer the team has grown over years. Behind it, I've worked across the asynchronous AWS pipeline of queues and serverless processors that ingests the platform's millions of events. That's the quiet value of long tenure on one complex system: I've seen most of the ways it can fail, and fixed my share of them.

• ClickHouse
• SQL optimization
• AWS Lambda
• SQS
• DynamoDB

Multi-chain transaction scanner

Tens of millions of dollars in token sales, tracked across seven blockchains — and verified.

I took over a scanner tracking partner token sales across six or seven blockchains — one that dropped batches of transactions and failed overnight, with everyone pinging me by morning. Hardening it took iterations: reading raw transaction data block by block, extracting the real amounts because the stated ones couldn't be trusted, and reconciling each day's totals against independent explorers, where any discrepancy past 5% meant a problem to chase down. By the time partner sales peaked in the tens of millions, the scanner had become the one thing nobody had to stay up watching.

• Node.js
• Ethereum
• Solana
• QuickNode
• AWS Lambda