Open to a new role · can start in two weeks
Senior Full-Stack AI Engineer TypeScript, React, Angular, Node.js

I build AI products. And I test the models behind them.

Ten years building web products in TypeScript.
Since 2025, AI work too: Lumigo, OpenAI's AI Lab on contract, and my own LLM job-search bot in production.

Warsaw (CET) · EU work-authorised · remote in Europe · B2B through my own company · English: full professional

Mikhail Kolatau on a balcony at sunset, holding a small black and white dog
OpenAI AI Lab
Aug → Oct 2026. I wrote software-engineering evaluation tasks for OpenAI's models. Promoted to reviewer a week in.
5 LLM providers
in the trace and evaluation UI I owned at Lumigo: OpenAI, Anthropic, Bedrock, Gemini, Vertex
My own LLM bot
in production since June 2026. Its scoring is measured on 80 labelled real postings.
10 years
of commercial TypeScript, frontend and full-stack. Angular since 2016, React at ICE, Kaseya and Pynea.

Experience

2016 → now

Contract work through Digis

Feb 2026 → now · Digis is a contracting agency

OpenAI AI Lab (opens in new tab) · AI Evaluation Expert and Domain Reviewer

Aug → Oct 2026 · alongside Oreamnos

I wrote hard software research tasks, ran the model on them and graded its answers against a rubric.

  • Three tasks delivered. Each came with a rubric of 60 to 97 items, every item backed by a source.
  • Promoted to reviewer of other experts' tasks a week in. Five submissions reviewed.
More about the OpenAI AI Lab work
  • I ran the model on every task to find where it really failed, and wrote up each failure with the source that disproves it.
  • Built tooling for the work: five Claude Code skills and a rubric linter that runs on every edit.
LLM evalsrubric designClaude Code

Oreamnos (opens in new tab) · Senior Full-Stack Engineer

Feb 2026 → now

Four Angular products and a Spring Boot service for German compliance teams.

  • Main frontend author of two products: document management and a whistleblowing platform.
  • Wrote the Playwright E2E suite, 169 tests in 31 files.
More about the Oreamnos work
  • Put a Tiptap rich-text editor (React 19) inside the Angular app, with support for collaborative editing. Built a reporting library with drag-and-drop dashboards on GridStack and ECharts.
  • E-invoice validation product, merged to staging. I built the Angular 21 frontend on my own layer of 31 components, plus part of the Spring Boot backend (multi-tenancy, sign-up with email verification, rate limiting).
  • Second-largest contributor to the shared design system of 80 components, used by three product apps. Led its upgrade from Angular 20 to 21 in a day.
Angular 21NxNgRx SignalsPrimeNGTailwind 4TiptapPlaywrightSpring BootKeycloakWCAG AA

Pynea (opens in new tab) · Senior Software Engineer

Feb → Apr 2026 · fixed-scope contract, alongside Oreamnos

Enterprise HR software, London-based. Next.js 15 and TypeScript, a monorepo of 17 packages.

  • Led the redesign of the Settings area, 8+ screens, from Figma to production.
  • Built interview scheduling on GraphQL, plus roles and permissions for the multi-tenant product.
More about Pynea
  • Fixed keyboard focus in the shared dialog system: a focus trap that blocked inputs, and focus handling in nested drawers and dialogs.
Next.js 15React 18GraphQL / ApolloZustandRadixStorybookTurborepo

Lumigo (opens in new tab) · Senior Software Developer

Feb → Nov 2025 · ended when the company was acquired

Monitoring for cloud and LLM apps. A frontend monorepo of 100+ packages.

  • Main author of the AI / LLM part of the product UI.
  • Shipped the LLM evaluation screens and wrote, on my own, the module that aggregates their scores.
  • Built the LLM trace view for five providers, with token cost per call and OpenTelemetry spans.
More about Lumigo
  • The five providers: OpenAI, Anthropic, Amazon Bedrock, Gemini, Vertex.
  • Second most active contributor to the frontend over the period.
  • Main author of Logs Live Tail, real-time log streaming over GraphQL subscriptions. Built the LLM traces table on AG Grid with server-side loading.
  • Ran releases every two weeks and live customer demos, mentored engineers, led code reviews.
Angular 19NxRxJSAG GridCytoscape.jsHighchartsCypressfeature flags

Kaseya (opens in new tab) (Pulseway (opens in new tab)) · Software Development Engineer

Aug 2021 → Feb 2025

Remote monitoring software for IT teams. Led the team and ran hiring interviews.

  • Moved the app from Angular 12 to 18: build time 272 s → 99 s, known vulnerabilities 65 → 3.
  • Built the Security page (two-factor enforcement, IP allow-list) and the shared search and filter bar.
More about Kaseya
  • Moved 30% of components from legacy LESS to Tailwind and added React modules to the Angular app through Module Federation.
  • White-label build on a feature branch: the brand name became a setting across the app's views and code, 2,170 unused files removed.
AngularAngularJSNGXSNgRxModule FederationReact widgetsTailwind

ICE Mortgage Technology (opens in new tab) · Senior Software Development Engineer

Mar 2020 → Aug 2021

A large US mortgage platform, part of Intercontinental Exchange. Angular with React modules inside.

  • Designed the layer that lets the Angular and React parts of the product talk to each other.
  • Cut page load time from about 4.5 s to 2.7 s.
AngularReactReduxmicro-frontendsJestGitLab CI
Earlier: Strateg Group, Full-Stack Developer (Angular 2+, PHP), 2016–2020

A start-up in Minsk. I chose Angular 2+ as the main framework, built the backend on Symfony and Laravel, and took features from requirements to release.

Angular 2+SymfonyLaravel

Projects

My own projects. The specs, decisions and measurements are mine. Claude Code, an AI coding agent, writes most of the code.

GigaWork (opens in new tab)

in production since June 2026

A Telegram bot that finds jobs for me. It pulls postings from 18 sources, drops the wrong ones with simple rules before any model call, and has a model score the rest.

37,702
postings collected by October 2026
47 of 60
good postings (ones I applied to or the bot sent me) pass the scorer I shipped. The old one, re-run, passes 35.

One thing I caught by measuring. The bot was paying a model to score postings that had nothing to do with my work, nurses and lawyers among them: 1,365 of 4,845 paid scores in three weeks came out near zero. I added a free check on the job title that runs before any paid call.

More about GigaWork
  • I test scoring on 80 real postings labelled by outcome. After a model change I re-ran them: scores fell about 8 points and the live threshold of 60 still held.
  • Posting text reaches the model wrapped as data, as a defence against prompt injection. The agents that read it get only allow-listed tools.
  • The job queue is plain Postgres. Workers take rows with an atomic claim, with no separate queue service.
  • It also built CVs and cover letters on demand from a structured record of my whole career, with the facts pinned so nothing got invented.
  • If one model provider fails, the next one takes over. A failed provider is parked for its retry window (a circuit breaker).
  • 11 LLM-judge agents check each generated CV, and two can block it. The judges are not yet checked against human labels.
  • An e-mail application goes out only after I press Send, and never twice.
  • I measured the Batch API at 2.3 times the price of cached calls ($0.0047 against $0.0020 per posting), so I dropped it before launch.
TypeScriptNext.jsNode.jsPostgres (Neon)SQLDrizzleClaudeGeminiZodVitest

GigaMixer (opens in new tab)

live

Mixing a multitrack recording in the browser. You type what is wrong ("the vocal is buried") and a model turns it into a checked list of changes.

  • A cheap model answers first. If the answer fails validation or the call fails, it retries once on a stronger model.
  • Guardrails live in code, not in the prompt: the server clamps every change the model proposes to an allowed range.
More about GigaMixer
  • A test set of 150 labelled phrases with pass thresholds, including attempts to trick the model.
  • A Python worker separates instruments from a recording with local models.
Next.jsAnthropic SDKtool callingshadcn/uiDemucsSupabaseCloudflare R2

big-s-studio (opens in new tab)

live

The site of my recording studio. Artists leave comments right on the audio waveform, and I update the site from a Telegram bot. It has its own design system.

  • 63 design tokens (colours, fonts, spacing) from one JSON source, generated as code for web, iOS, Android and Windows.
  • The build fails if text contrast drops below WCAG AA, the accessibility standard.
Next.js 16React 19Server ComponentsWeb Audio APIdesign tokensDrizzle

bandcamp-daily (opens in new tab)

runs daily

Picks one new release a day from my Bandcamp collection. A small audio model compares how releases sound, and that raised ranking quality (MRR) from 0.118 to 0.200. I tuned and tested on the same data, so the number is optimistic.

TypeScriptNode.js 20ONNXembeddingsGitHub Actions

More projects

Bandcogs (opens in new tab)An app for bands. Next.js, NestJS, React Native with Expo, Supabase.
in progress
agent-harness (opens in new tab)Safety rules for coding agents. Every rule has a test that proves it blocks what it should.
public repo
ctaio-jobfit (opens in new tab)Matches a CV to 406 open postings with one model call.
live demo
under-big-black-sun (opens in new tab)A photo zine in English and Polish. Payload CMS, Next.js 15, Postgres.
live
SpeakNow (opens in new tab)Video recording in the browser that adapts quality to connection speed. Angular 21.
public repo
Manyas E2E (opens in new tab)Playwright tests for a survey in 10 languages on three devices. Client work.
public repo
MixFeedback (opens in new tab)Real-time comments on a track. NestJS, Angular 19. A study project.
public repo
quizMy study tool. A model grades spoken answers about code against a rubric.
private repo
tg-summary-botSummarises long Telegram chats with Gemini. Python.
runs locally

Skills and stack

Most used first.

AI / LLM

structured outputstool callingLLM evalsClaude APIprompt cachingLLM observability
11 more AI and LLM skills
LLM-as-judgemodel routing and fallbackcheap model first, stronger one on failurecost and latency trackingprompt-injection defencesagents with least privilegebrowser agentshuman approval before irreversible stepsembeddings, cosine similarityWhisper speech-to-textGemini

Frontend

TypeScriptJavaScriptAngular 2 → 21React 18–19Next.js 15–16SSR, App RouterRxJSTailwind
10 more frontend skills
HTML, CSSAngular SignalsServer ComponentsReact Native, ExpoAngularJSSCSSLESSstyled-componentsCSS ModulesWeb Components

State and data

NgRx SignalsTanStack QueryGraphQLRESTZod
10 more state and data skills
NGXSNgRxElfZustandReduxApolloReact Hook FormOpenAPIWebSocketAWS Amplify subscriptions

UI and design systems

design tokensaccessibility, WCAG AAFigma to codeStorybookPrimeNG
15 more UI skills
Angular MaterialRadixMantineshadcn/uiStyle DictionaryAtomic DesignTiptapAG GridCytoscape.jsEChartsHighchartsRechartsvisxGridStacki18n, Transloco

Backend

Node.jsPostgreSQL, SQLDrizzleTelegram botsqueues in Postgres
9 more backend skills
NestJSExpressPayload CMSSupabaseNeonUpstash RedisSpring BootSymfonyLaravel

Node and Postgres in production: only in my own products. At employers my backend work was Spring Boot and PHP.

Architecture

Nx monorepoModule Federationmicro-frontendsTurborepomulti-tenancyroles and permissionsfeature flagslarge version migrations

Testing

VitestPlaywrightCypressJestJasmineReact Testing LibraryMSWunit, integration, E2Evisual regression

Tooling

Claude CodeGitVercelCI/CD: GitHub Actions, GitLab CI, CircleCIDockerViteWebpackSentryPostHogKeycloak, OIDC

How I work with AI agents

The setup is in my repos. I can show it on a call.

ChatGPT in my daily work since 2022, Cursor at Lumigo, Claude Code now. I write the spec and a failing test first. One agent writes the code, a second one reviews it and cannot change files.

The four rules
Spec before any behaviour changethe edit is blocked while the spec has open questions
Separate agents write, review and verifydiagnosis runs on a stronger model
Automatic guardsno secrets in commits, lint must pass, the turn cannot end while tests are red
Every guard is testedwith one case it must allow and one it must block

Outside of code

Audio engineer

I record, mix and master bands from Warsaw's DIY hardcore-punk scene at big-s-studio (opens in new tab). I also write my own audio analysis tools in Python.

Trained as a quantum electronics engineer: BSc, BSUIR (Minsk), 2018 · English: full professional · Russian: native · Polish: A2

Contact