Panagiotis (Panos) Gkilis, machine learning engineer, independent researcher, founder of BedVibe Studios and author of The Mirelands

Panagiotis (Panos) Gkilis

Machine Learning Engineer • Independent Researcher • Founder • Author

I am a machine learning engineer and independent researcher, and the founder of BedVibe Studios. I train speech models from scratch and run them in production, and I publish open-access research on the failure mode that is hardest to catch — the one that returns a plausible success instead of an error. A growing record of citable open-access research with DOIs, four open-source verification libraries on PyPI, and a 730-million-parameter text-to-speech model serving live traffic. I also write dark fantasy, as Panos Gkilis.

⚙️

Engineering & Architecture

Panagiotis is the founder of BedVibe Studios and the sole engineer behind BedVibe TTS — a 730-million-parameter multilingual neural codec text-to-speech model he trained from scratch on consumer GPUs and runs behind a live, billed API. His technical focus is machine learning, speech synthesis and retrieval systems. He maintains four open-source verification libraries on PyPI, designs and operates production AI systems end to end, and publishes open-access research reports and preprints with DOIs.

The through-line of all of it is measurement. A fault that returns an error is one you fix that afternoon; a fault that returns a plausible success is one you ship. Every system here carries an instrument built to catch the second kind — an evaluation gate that can refuse a deployment, a linter that fails a training run before the GPU hours are spent, a coverage layer that reports what it could not check instead of staying silent.

That instinct is not improvised — it comes out of the laboratory. Years of natural-sciences coursework and bench work train a particular habit: an experiment is not finished when it produces a number, but when the error on that number has been characterised, bounded and stated. Systematic versus random error, propagation through a calculation, the difference between precision and accuracy, and the discipline of declaring what a measurement cannot resolve — those are laboratory reflexes, and they are exactly what makes an evaluation gate worth trusting. Every experiment and every validation in this work is built to hold the margin of error as tight as it can honestly be held, and to report it plainly where it cannot be tightened further.

Education: university-level coursework in Natural Sciences, Hellenic Open University — 11 of 12 required modules and all laboratory courses completed, 222 ECTS, with the final module in progress.

🔬

Research & Engineering

Panagiotis publishes single-author, open-access research on how machine-learning systems fail quietly: training objectives that corrupt while the loss curve still descends, speech synthesis that passes automated scoring and fails a human, retrieval that returns confident answers the evidence does not support, and speaker embeddings that reject a genuine voice after an imperceptible change. Every record below is released with its code or data, and several of them are the direct source of the libraries in the next section.

Published, citable open-access records with DOIs — preprints, technical reports and software deposits, all single-author. Zenodo is the canonical deposit for every one of them; ORCID and OpenAIRE index from there. The complete and current list is on Zenodo; the records below are the ones described here. This page identifies the person; the Engineering Portfolio documents the individual projects, and the engineering notebook carries the write-ups.

Publication / Record Type Finding DOI
Structure Substitutes for Scale in Cross-Document Temporal Reasoning Preprint · Aug 2026 Thirty-eight questions ask whether one event precedes another when the two are narrated in different documents and share no character, place or causal link — the ordering exists only across the corpus, never inside it. Given the passages, every model scored 0 of 38 at four sizes, refusing 92–100% of the time: the control works. Given the identical facts as a structured chronology, an 8B model scored 28/38, matching a 14B model given prose. The advantage is gone by 14B, and scrambling the supplied order collapses accuracy to 10.5%, which rules out memorisation. He then audited his own scorer and published the correction: the equivalence held, his claimed gap over the best prose representation halved. 10.5281/zenodo.22169643
Speaker Encoders Disagree About Who You Are When You Shout: A Content-Matched Benchmark of Fourteen Speaker Encoders Under Expressive Phonation Technical Report · Aug 2026 Fourteen encoders across three architecture families, scored on 11,935 identical trials frozen before any model was loaded, over eight speakers producing the same 1,360 sentences in six phonatory and emotional states. Error rates span 0.047 to 0.233 on identical audio — a fivefold range, and 7.4-fold on whispered speech. Bootstrap intervals resample speakers and rebuild both sides of every draw, with Holm–Bonferroni control over a family of 91 comparisons declared in advance. 10.5281/zenodo.22158030
The Loss Curve Is Not a Sufficient Statistic: Silent Objective Failures from Sentinel–Class Collisions in Neural Codec Language Models Preprint · Aug 2026 A sentinel–class collision makes masked cross-entropy silently delete an entire class while the loss curve descends normally, so no curve-based diagnostic can detect it. Reproduced in 40 lines on CPU. The two detectors it defines ship in trainproof. 10.5281/zenodo.21864658
Intra-Speaker Vocal Variation and Speaker-Embedding Displacement: A Matched-Content Replication Across Three Encoder Architectures Preprint · Aug 2026 One speaker 0.7 semitones below normal — imperceptible — lost 0.238 of verification score across all eight sentences, in ECAPA-TDNN, ResNet and WavLM alike. Pre-registered, with an adversarial prior-art audit that killed its own headline claim and was published anyway. 10.5281/zenodo.21921957
A 2D Linear-Elasticity Finite Element Solver Written from First Principles, Validated Against the Kirsch Stress Concentration Factor Software · Aug 2026 Hand-written Delaunay mesher, isoparametric elements and a conjugate-gradient solver, with no FEA library and no scipy. Recovers the analytic Kirsch factor as 3.00002 against an exact 3 the code never receives. Ten gates pre-registered; nine pass and the failure is published. 10.5281/zenodo.21892063
BedVibe-TTS: An Engineering Report on a Multilingual Neural Codec Language Model with Multi-Axis Conditioning Technical Report · 2026 Architecture, data pipeline, training diagnostics and ablations for a from-scratch multilingual neural codec TTS system with speaker, emotion and trait conditioning, plus a reproducibility finding on attention-backend × precision interactions across GPU generations. 10.5281/zenodo.19781414
An Automated Failure-Mode QA Framework for Neural Text-to-Speech Systems Technical Report · Jun 2026 130 edge cases across 3 voices (390 samples) with zero structural audio defects. Blinded human validation of the 42 ASR-uncertain cases (controls 15/15) split 45/55 between real failures and ASR false negatives — establishing that uncertain cases must be quarantined for review, not auto-judged. The methodology behind ttsproof. 10.5281/zenodo.20757552
Diagnosing Hierarchical Retrieval Failure in Long-Document RAG Technical Report · Jun 2026 A stage-level ablation isolating where long-document retrieval actually fails: first-stage chapter-selection error compounding (27/80) and neighbour-expansion dilution (32/80). Oracle-chapter routing recovered recall to 0.7844, quantifying how much of the loss is routing rather than ranking. 10.5281/zenodo.20692450
LongBook Verifier: Evidence-Grounded Evaluation of Long-Document Retrieval and AI-Generated Answers Software · Jun 2026 Retrieval baselines, model-output scoring and a 240,767-word stress-test benchmark, released as a complete research package. The research behind the live BookProof product. 10.5281/zenodo.20513115
ORCID: 0009-0007-3805-170X  ·  Zenodo: all deposited records
Each DOI above is a Zenodo concept DOI: it always resolves to the most recent version of that record, so a citation made today does not go stale when a revision is deposited.
🎙️

Data, Tooling & Production

The training data behind BedVibe-TTS was not licensed or scraped. It was recorded, produced and curated in-house — which meant building the apparatus first, then running a production operation around it.

What was built Why it had to be built
A purpose-built recording tool
Unity · C# · Rust binary
Off-the-shelf recording software captures takes; it does not enforce a corpus design. This tool drives a performer through a fixed script, holds the emotional state constant across a session, and writes data already aligned to the training format — so curation is a property of capture rather than a cleanup pass afterwards.
A parallel emotional corpus 1,360 identical phrases recorded in six emotional states per speaker — neutral, happy, angry, whisper, shouting, scared. Because every speaker reads the same script, emotion and speaker identity can be separated cleanly. Most public emotional corpora cannot support that comparison.
Continuous attribute traversals Recordings that move smoothly between vocal qualities — clean to raspy, to growl, to breathy, to nasal — captured on purpose. Almost nothing public holds the trajectory between voice qualities rather than only the endpoints.
A production operation Professional voice actors and actresses engaged under contract, directed through sessions, with signed consent and clear commercial licensing held by the studio. Rights on the corpus are unambiguous, which is where most speech datasets fail.
Rust across the stack The speech synthesis service (axum), a real-time voice-activity detection front end, a production Amazon Polly server, and a real-time aerospace telemetry engine — systems chosen for Rust where latency and correctness under load mattered more than iteration speed.
A complete game A browser card game built on its own engine, set in The Mirelands and tied to the saga's lore — designed, engineered and shipped solo, not assembled from a template.

The pattern is the same one that runs through the research: when the instrument required did not exist, it was built and then held to a measurable standard. Dataset details and licensing · For voice providers

🧪

Open-Source Verification Libraries

Four libraries, all published to PyPI, all installable today. They exist because the same failure kept recurring across unrelated systems: a check that reports success without having checked anything. Each library is a deterministic instrument — same input, same verdict, no model call in the judgement path — and three of the four are the direct implementation of a published research record above.

Library What it refuses to let through Grounded in
trainproof
pip install trainproof
A linter for ML training runs. Deterministic PASS / WARN / FAIL verdicts on datasets, tokenizers and training logs, run before the GPU hours are spent. It carries the two sentinel-collision detectors, so a corrupted objective is caught at pre-flight rather than at evaluation. Loss-curve preprint (21864659)
ttsproof
pip install ttsproof
Automated failure-mode QA for text-to-speech: structural audio checks, equivalence-aware WER/CER that does not punish a correct homograph, and ASR-uncertainty quarantine — cases the scorer cannot adjudicate are held for human review instead of being silently marked pass or fail. TTS QA report (20757553)
spkproof
pip install spkproof
Deterministic checks for speaker-verification studies. Detects F0-tracking contamination — the differential measurement error that breaks pitch trackers in one direction only and silently biases a result. It measures false rejection of genuine speakers, which is the failure mode most verification work leaves unexamined. Speaker-drift preprint (21921958)
notchecked
pip install notchecked
Coverage accounting for validators: what was checked, what could not be checked, and what was never reached. It closes the gap that makes every other gate untrustworthy — a validator that inspected 0.6% of its corpus and reported green. Absence of a finding is reported as absence, never as a pass. Built from the failure it names

The discipline runs to the research itself: predictions are pre-registered before a run, adversarial prior-art audits are published even when they kill the headline claim, a result at p = 0.013 was withdrawn because two defensible corrections pointed in opposite directions, and a rate-limited extraction that lost 119 of 259 calls was discarded and re-run rather than patched, because the survivors were a biased sample and not merely a smaller one.

📚

Authorship & Publications

Panos Gkilis is the name under which he publishes fiction and language-learning titles. The bibliography spans language education reference works — now in three languages — and a long-form dark fantasy saga:

  • Norwegian Grammar: A Complete A1–C1 Reference 600+ verbs, pronunciation and advanced usage. Written from fourteen years of teaching the language.
  • Норвезька граматика · Gramatyka Norweska The same A1–C1 reference published for Ukrainian and Polish learners — not translations of a summary, but the full grammar.
  • Norwegian Exercise and Puzzle Books Language exercises and puzzle collections for self-study.
  • The Mirelands Saga Dark Fantasy Series (6 Books Published, 10-Book Roadmap) Red Moon Rises · The Ancient Observatory · The First Horn · The Drowned Reach

Every title is available as ebook and paperback, and the saga is produced in parallel as full-cast dramatized audio. Browse the catalogue.

🌌

Worldbuilding & The Mirelands

The Mirelands belongs to a larger fictional universe. The setting of the story is the continent of The Mirelands, which exists on the planet Kthonia. Six books in this world have already been published, with a final goal of ten books. Dramatized audio drama adaptations have been produced for the first five books; the sixth audio drama is currently in development, seeking actresses and actors for voice roles. The long-term goal is a multimedia franchise spanning books, games, audio dramas, tabletop projects, and film collaborations.

It opens as PG-13 epic fantasy with coming-of-age romance and darkens into darker fantasy and cosmic horror from Book Five onward. The first book begins with an orc patrol and elven scouts colliding at the marsh edge as the Red Moon rises out of cycle, and survival turns enemies into reluctant allies.

The world is written under a fixed canon: a locked timeline, a rule for how myths form inside the setting, a constructed Elven language, and a celestial-event system that governs when things like the Red Moon can occur. Continuity is checked against that record rather than remembered — the same discipline as the engineering work, applied to fiction.

The franchise hub · Kthonia · Reading order · Dramatized audio

🔗

Official Sources & Profiles

The following links represent the verified, official public sources for Panagiotis (Panos) Gkilis' professional, creative, and corporate activities:

Portrait & press use. The portrait on this page — © Panagiotis Gkilis — may be reproduced without charge in editorial and journalistic coverage, conference and speaker listings, podcast and interview announcements, and academic or press material referring to this work, provided it is credited to BedVibe Studios and is not altered beyond cropping and colour correction. Commercial, promotional and any other use requires written permission: bedvibe@bedvibe.studio.

Related Pages

BedVibe Studios Engineering Portfolio ML Engineering Profile Engineering Notebook Author Page

Collaboration & Opportunities

Panagiotis is open to remote engineering roles, global relocation opportunities, and strategic creative collaborations. He seeks to connect with professionals interested in scaling AI-driven production pipelines, as well as creative partners looking to expand the dark fantasy world of The Mirelands into tabletop games, collectable figures, and film/media adaptations.