Nikolay Kostov's Blog

This December marks a significant milestone in my professional journey – 10 years as a Microsoft Certified Trainer (MCT). Since earning this certification on December 23, 2014, I've been privileged to witness and contribute to the ever-evolving world of technology.

As a software engineer and expert witness in Bulgarian courts, my role as an MCT has been enriched by diverse experiences. Over these years, I've shared knowledge, embraced learning, and watched numerous students and professionals grow.

As my certification approaches its expiry on December 24, 2024, I reflect on a decade of growth, challenges, and achievements. This journey has been more than just a professional endeavor; it has been a vital part of my identity.

I am grateful for the community, experiences, and opportunities that being an MCT has offered. Here's to continuous learning and the next phase of this exciting journey!

For the second consecutive year, I am honored to be recognized as one of the top 100 most influential IT people in Bulgaria (2023).

You can find the complete list of awardees at https://careershow.bg/indexes/it-influencers.

Google released Gemini 3.8 Flash in September and Twelve Labs released Pegasus 1.6 in October, so I ran both, plus Pegasus 1.5 (my current default), on 34 minutes of real surveillance footage (night, infrared, motion-triggered recording, people far from the camera) and scored five configurations against 46 moments I had verified frame by frame. Gemini 3.8 Flash at high resolution wins on accuracy, errors and long clips at about Pegasus' price; Pegasus 1.6 is no more accurate than 1.5, just five times faster; and none of them saw a scuffle between people 90 pixels tall at night.

Five months after standardising on ElevenLabs Scribe v2 for Bulgarian transcription, I re-ran nine speech-to-text engines (ElevenLabs, Speechmatics enhanced and Melia 1, OpenAI gpt-transcribe, gpt-4o-transcribe-diarize and whisper-1, xAI Grok Voice Transcribe 2.0, local Whisper large-v3, and a two-engine combination) on 2 h 41 min of my own varied recordings, scored against hand-corrected transcripts. Scribe v2 still wins on word error rate, recall and speaker attribution; gpt-transcribe silently drops hard passages, whisper-1 loops, and Grok does not support Bulgarian yet.

I won Second Place in the Professionals category of the international DataArt Proggy-Buggy Towel Contest 2026 — an algorithmic programming competition held on 6 June 2026.