Panagiotis Gkilis

AI & Backend Engineer. Creator of ttsproof and trainproof — open-source tools for TTS evaluation and failing-run detection. Trained a 730M-parameter neural speech model. Production RAG + automation.

Thessaloniki, Greece Joined on Aug 1, 2026 bedvibe@bedvibe.studio https://tts.bedvibe.studio/portfolio/

Education

Natural Sciences coursework, Hellenic Open University — 11 of 12 modules, 222 ECTS

Pronouns

He/Him

Work

Founder & Engineer, BedVibe Studio

Panagiotis Gkilis

Aug 31

I Assumed My Retriever Failed at Stage One. The Bigger Failure Was at Stage Three.

#machinelearning #rag #llm #python

3 min read

Panagiotis Gkilis

Aug 31

One Integer Deleted the Stop Token From My Loss. The Curve Never Noticed.

#machinelearning #python #deeplearning #debugging

4 min read

Panagiotis Gkilis

Aug 31

Fourteen Speaker Encoders Heard the Same Voice. Their Error Rates Differed Five-Fold.

#machinelearning #python #datascience #ai

5 min read

Panagiotis Gkilis

Aug 28

Two of My Instruments Were Wrong, in Opposite Directions

#debugging #testing #seo #webdev

7 min read

Panagiotis Gkilis

Aug 25

My Agent Answers From 0.6% of Its Corpus and Reports It Like a Full Read

#rag #ai #architecture #python

5 min read

Panagiotis Gkilis

Aug 25

I Shipped a Feature in May. Nobody Could Reach It Until Yesterday.

#webdev #javascript #testing #architecture

4 min read

Panagiotis Gkilis

Aug 25

My Validation Layer Was Correctly Deleting 16% of My Good Data

#rust #architecture #devops #programming

7 min read

Panagiotis Gkilis

Aug 25

Your TTS Model Sounds Great — Until It Says "GPUB"

#python #machinelearning #testing #ai

4 min read

Panagiotis Gkilis

Aug 25

Why Corrupted Training Data Doesn't Show Up as High Loss

#machinelearning #python #ai #datascience

4 min read

Panagiotis Gkilis

Aug 25

The Same Bug, Four Times, Three of Them Mine

#python #testing #architecture #devops

4 min read

Panagiotis Gkilis

Aug 25

A Hundred Per Cent Success Rate and One Usable Answer

#python #architecture #devops #ai

6 min read

Panagiotis Gkilis

Aug 20

Ten Tests, Written Before the Code. The One That Failed Taught Me the Most

#testing #python #programming #science

4 min read

Panagiotis Gkilis

Aug 17

Speaker embeddings can mistake a 0.7-semitone change for something else

#machinelearning #ai #research #datascience

3 min read

Panagiotis Gkilis

Aug 3

I spent months training a 730M model with no tool that could tell me it was already dead

#machinelearning #python #opensource #mlops

7 min read

DEV Community

Panagiotis Gkilis

Badges

Writing Debut

GitHub Repositories

Skills/Languages

Currently learning

Currently hacking on

Available for

I Assumed My Retriever Failed at Stage One. The Bigger Failure Was at Stage Three.

One Integer Deleted the Stop Token From My Loss. The Curve Never Noticed.

Fourteen Speaker Encoders Heard the Same Voice. Their Error Rates Differed Five-Fold.

Two of My Instruments Were Wrong, in Opposite Directions

My Agent Answers From 0.6% of Its Corpus and Reports It Like a Full Read

I Shipped a Feature in May. Nobody Could Reach It Until Yesterday.

My Validation Layer Was Correctly Deleting 16% of My Good Data

Your TTS Model Sounds Great — Until It Says "GPUB"

Why Corrupted Training Data Doesn't Show Up as High Loss

The Same Bug, Four Times, Three of Them Mine

A Hundred Per Cent Success Rate and One Usable Answer

Ten Tests, Written Before the Code. The One That Failed Taught Me the Most

Speaker embeddings can mistake a 0.7-semitone change for something else

I spent months training a 730M model with no tool that could tell me it was already dead