THE WORK

Completed high-volume AI data annotation and evaluation work across audio, video, and text projects, helping improve model behavior through structured labeling, quality review, and consistency-focused judgment calls.

Worked inside Multimango while tracking production time through Hubstaff, contributing 485+ hours as of May 2026 across tasks that required careful guideline interpretation, edge-case handling, and sustained attention to detail.

WHAT I DID

• Reviewed multimodal inputs against detailed project-specific annotation guidelines
• Applied labeling rubrics consistently across large batches of production data
• Evaluated ambiguous cases with attention to context, intent, quality, and model usefulness
• Maintained accuracy and pace across long annotation sessions while respecting platform confidentiality requirements

SPECIALIZED ASSIGNMENT

After demonstrating strong performance in general annotation tasks, I was moved into a specialized multilingual annotation track for contributors with language skills beyond standard generalist work.

The work required evaluating language-sensitive content, interpreting project guidelines across different linguistic contexts, and maintaining annotation quality when meaning, tone, or intent depended on language-specific nuance.

PERFORMANCE

Consistently delivering a 2.90 quality score against a 2.00 benchmark.

THE PROJECT

A custom Customer Relationship Management platform built specifically for an online English school to handle student data, scheduling, and reporting.

ARCHITECTURE

The system was engineered as an installable PWA to provide a responsive, native-like mobile experience. Authentication is handled securely via passkeys, backed by a real-time Supabase database.

DATA & I18N

To support a diverse user base, the entire interface implements robust i18n localization. Student management features highly interactive data grids and supports bulk data imports and exports natively through Excel.

INTEGRATIONS

The platform connects directly to the Google Drive API, allowing managers to securely send PDF textbooks to students straight from the dashboard.

THE PROCESS

AI-assisted prototyping drove the rapid development of these complex features. This approach demonstrates how effectively agentic workflows can be used to ship production-ready software for real clients.

THE PROJECT

This custom web platform hosts a 12-module TEFL certification course and was built entirely using Cursor.

THE PROCESS

Prompting AI models helped generate the interface code, connect a Supabase database, and deploy the site to Vercel. The project served as a practical test of using AI to build and ship functional software.

FEATURES

The system handles user accounts, sends automated emails, and generates PDF certificates for students when they finish the course.

CONTENT

The course material includes 12 hours of video and audio lessons, all scripted, recorded, and edited specifically for this platform.

THE PROBLEM

I wanted custom utilities that fit perfectly into my macOS workflow without paying for bloated third-party apps. Instead of learning Swift from scratch, I used Antigravity and Xcode to generate two native macOS widgets.

1. WATER TRACKER

A Python-backed script that hides in the Mac menu bar near the camera notch. It expands on hover or double-click, allowing me to log my daily water intake and sending automated hydration reminders.

2. AUTO LANGUAGE SWITCHER

Since I constantly switch between English and Ukrainian, I generated a secure, local widget that automatically toggles my keyboard language based on context, eliminating the friction of manual switching.

THE TAKEAWAY

I used the Antigravity agent to navigate Xcode and build these utilities. It was a practical exercise in relying on AI tools to ship functional local software in unfamiliar desktop environments.

THE EXPERIMENT

A personal project built purely for fun to test the limits of AI code generation. I used LLMs to help me turn an ancient, jailbroken iPad Mini 2 into a motion-tracking security camera and a custom PWA macro deck.

HOW IT WORKS

When it detects motion, it records a short video clip, converts it to a GIF, and automatically sends it to my Telegram bot. The macro deck side acts as a custom shortcut pad with a virtual keyboard and trackpad.

THE TAKEAWAY

I didn't write this codebase from scratch—I relied heavily on AI prompting to figure out the hardware constraints and write the scripts. It was a practical exercise in testing model capabilities, debugging AI hallucinations, and prompting my way out of dead ends.