Mykyta Bulatnikov
AI Engineer & Founder
Prague, Czech Republic · m.bulatnikov@icloud.com · linkedin.com/in/mykyta-bulatnikov · github.com/razornne · skriptly.io
Summary
I build Skriptly, an AI transcription and dictation product now in private beta, and I am the only person who owns data and AI at a proptech platform. I decide what to build and how it should behave, use AI tooling to ship it fast, and measure the result: evaluation sets, cost per hour of audio, and checks on every number a language model states.
Looking for an early-stage startup where I can own data and AI end to end. Always glad to talk with people building applied AI and with investors who follow the space.
Experience
AI transcription and dictation. Web app, Windows app, subscription plans. Solo: product, engineering, pricing, billing, support. In private beta with a limited group of testers.
- Built it from idea to private beta: speech-to-text in 60+ languages with speaker recognition, summaries and action items, a hold-to-talk dictation app for Windows, and Free, Pro and Team plans with Stripe billing.
- Rebuilt the engine twice. A hosted API first, then an open-source pipeline (faster-whisper, pyannote, an LLM correction pass) on serverless GPUs, then Soniox after a five-language benchmark: 2.4% against 3.7% word error rate, no dropped words against 1.3%, and about four times cheaper per audio hour.
- Works from unit economics: an hour of call costs about $0.30 to process end to end. Switched off two features after measuring that they cost more than they earned (GPU-generated titles, a live transcript that doubled the cost of an hour).
- Dictation built for latency: a Rust app that streams audio straight to the speech provider with a short-lived key, pastes LLM-cleaned text in about a second, and learns each user's vocabulary from their own edits.
- Handles 3–4 hour recordings with silence-aware chunking, parallel GPU workers and speaker re-identification across chunks. Quality is measured, not felt: an evaluation set tracks word error rate and deleted words.
Prague. Sole owner of data and AI, reporting to the CEO. Part-time intern Jun–Sep 2025, full-time since Oct 2025.
- Built the analytics platform from zero: PostgreSQL into BigQuery, a dbt warehouse of 100+ models, and a metric registry of 150+ metrics defined once in git, so a dashboard card without a definition fails the build.
- Replaced Looker Studio with four custom Next.js dashboards served from static exports of the dbt marts. Cut reporting latency from about a day to 15 minutes.
- Built ML prototypes for building risk, churn and payment behaviour, validated out of time with point-in-time features.
- Built an MCP server over the metric layer and an LLM analytics agent that answers through it, with every number it states checked against the query it came from.
- Built a proactive analyst agent that works as a member of the team: it watches all metrics around the clock and flags anomalies without being asked.
- Audited production data instead of trusting dashboards: found a collection-rate metric that was off by several times, exclusion logic duplicated across views, and a device ID being counted as a person.
- Built monitoring and alerting on Cloud Functions and Slack, and cut its alert volume by about two thirds after auditing a week of its own output.
Open-source projects
lease-validator — an agentic RAG engine that checks lease agreements against YAML rules and returns no verdict it cannot quote the document for. Violation F1 0.976 against a 0.692 keyword baseline, 100% violation recall, on 180 labelled pairs. github.com/razornne/lease-validator
metrics-mcp — a semantic layer an LLM agent has to go through: metrics defined once in YAML, SQL built from dbt, and every number the agent states verified against its source query. The guard is tested against deliberately lying stubs. github.com/razornne/metrics-mcp
Skills
Applied AI: Anthropic, Gemini and speech-to-text APIs, MCP servers, RAG and agent loops, evaluation sets, faster-whisper, pyannote, embeddings and pgvector, self-hosted open-weight models
Engineering: Python, TypeScript, Next.js, Rust (Tauri), Modal, GCP (Cloud Functions, Cloud Scheduler), GitHub Actions, Stripe, Git
Data: SQL, BigQuery, PostgreSQL, dbt, dimensional modelling, metric governance, product analytics, anomaly detection
Education
Languages and work authorization
Ukrainian and Russian (native), English (C1), Czech (A2). Based in Prague. Available as an employee, or as a contractor (Czech sole-trader registration takes about a week). Open to remote, hybrid and Prague-based roles.