Panagiotis (Panos) Gkilis — Machine Learning Engineer

Machine Learning Engineer — Production ML Systems · MLOps · Speech & Language AI

Thessaloniki, Greece · open to Remote / Athens / Oslo · bedvibe@bedvibe.studio · Greek, English, Norwegian (all fluent) · EU citizen, Norwegian-registered business (ENK)

Everything below is live, running and open to inspection — please look before you decide.

ai.bedvibe.studio/work — start here: everything I have built and published, on one page, with the links into all of it.

tts.bedvibe.studio/portfolio — 36 project write-ups you can open right now: architecture, screenshots, evaluation reports and the failures I found in my own work.

ai.bedvibe.studio/articles — engineering notebook: how each system was debugged, measured and gated, written up in full.

Summary

Six years of end-to-end production ML: models trained from scratch, deployed behind live APIs, and operated for paying customers. Sole engineer of BedVibe Studios — a multilingual speech-AI platform built on a ~730M-parameter TTS model trained on consumer GPUs across 13 languages, running on FastAPI, PostgreSQL/Supabase, Docker and Linux/Nginx with Stripe-billed usage. Thirteen published research records with Zenodo DOIs — eleven publications and two software releases. Four open-source Python libraries published on PyPI and a fifth public on GitHub, plus a real-time telemetry engine in Rust.

How I build

I develop with AI coding agents — Claude Code and Codex, cross-checked against each other through MCP tooling — and treat their output as untrusted until it clears a gate: automated evaluation suites, ASR-validated model QA, and static checks. All four of my open-source libraries are those gates: ttsproof (TTS failure-mode QA), trainproof (training-run linter), spkproof (speaker-verification measurement checks) and notchecked (coverage accounting — what was checked, what could not be, and what was never in scope, as typed states rather than pass/fail). The method is why a LoRA collapse and a live RAG regression were caught before release, not after.

Reviewed and adopted by outside engineers: an IBM Certified Expert IT Architect (21 years at IBM) reviewed the notchecked schema publicly and contributed three corrections, all accepted; the maintainer of SpeechSDK, a commercial API spanning 17+ speech providers, requested an integration, which shipped; trainproof is the schema's first production adopter.

Core Skills

ML & modeling — transformer training from scratch (AR + NAR neural codec LM), LoRA fine-tuning, speaker embeddings (ECAPA-TDNN, WavLM, ResNet), SentencePiece tokenization, Whisper/WhisperX ASR & forced alignment, retrieval/RAG evaluation, LLM-output benchmarking

MLOps & deployment — FastAPI inference APIs, async job queues, release-gating automated evaluation suites (ASR-validated WER/CER), distributed tracing (OpenTelemetry + Langfuse joined on a shared trace id, Jaeger), LLM gateways with per-key quota, token metering, bounded admission and deadline-aware failover, monitoring & production debugging, systemd service management, nginx/TLS reverse proxying, firewall hardening, GPU inference, ONNX export, Docker, Linux, HTTPS/domain routing, CI/CD (GitHub Actions, PyPI trusted publishing)

AI orchestration & data — LangChain/LangGraph stateful validate-retry graphs, CrewAI (evaluated against LangGraph on eight fault scenarios and rejected on measurement), MCP servers and clients with schema-validated responses, RAG pipelines with pluggable vector backends (FAISS IVF/FlatIP, Qdrant HNSW, brute-force baseline) benchmarked on recall@k against an exact index, hybrid retrieval (dense + BM25), schema-enforced output (Pydantic), self-hosted n8n; speech-dataset curation (100+ hours), metadata schema design (1.55M records), custom binary training formats

Backend & languages — Python, SQL, JavaScript, Rust, HTML/CSS · PostgreSQL (schema design, RLS, plpgsql), Supabase, REST, JWT/auth, Stripe & idempotent webhooks, AWS SDK, Deno/TypeScript edge functions

Professional Experience

Founder / Machine Learning Engineer · BedVibe Studios

2020 – Present · Greece / Remote · Norwegian-registered sole proprietorship (ENK)

Model training — trained a ~730M-parameter multilingual neural-codec TTS model from scratch (24-layer AR + NAR transformer over Encodec/DAC tokens, Vocos vocoder): spanning 13 languages, 15 human speakers plus 2 synthetic voices, and 6 emotional states — three independent spans, not a crossed grid — on ~3,000–4,000 GPU-hours on consumer GPUs.
Production deployment — serve it behind FastAPI (20+ REST endpoints) with async queues for long-form generation, 200–400 ms GPU inference per utterance, and Stripe-billed token accounting. Live at tts.bedvibe.studio.
Infrastructure — operate a 10-service backend on a self-managed Linux server: nginx/TLS reverse proxy, each service isolated as a systemd unit, PostgreSQL 16, default-deny ufw firewall.
Evaluation & monitoring — built an ASR-driven QA harness (WER/CER, silence/loop/replay detection, hold-out cases) gating every checkpoint; it caught a total LoRA fine-tune collapse — 60/60 hard failures vs 14/60 baseline — before release.
Data pipelines — engineered a 108,000-sample multilingual corpus with a custom 16k SentencePiece tokenizer, an ECAPA-TDNN speaker-embedding pipeline (4,287 samples, zero failures), and a memory-mapped binary format that removed a 3–5× dataloading bottleneck.
LLM / RAG evaluation — benchmarked 5 retrieval strategies on a 240,767-word corpus against an 80-question gold set (3.3× context recall over naive baseline) and scored 9 external AI systems on separated evidence-grounding vs answer-coverage metrics.
Datasets & compliance — recorded and curated 100+ hours of rights-cleared studio speech under signed talent contracts with GDPR-aware handling; commercial emotional-speech datasets sold to external customers.

Norwegian Language Teacher · globalcitizen

Sep 2013 – Jun 2019 · North Athens

Taught Norwegian to Greek- and English-speaking adults through B2; designed curricula and explained complex material across three languages.

Natural Sciences Teacher · Aristarhos Private High School

Sep 2014 – Jun 2017 · Peraia

Taught chemistry and natural sciences; prepared students for university-track examinations.

Research & Publications

Reconstruction History, Not Synthesis: Benign Processing Moves Frozen Detectors on Genuine Human Speech — Zenodo, Sep 2026 · DOI 10.5281/zenodo.22819222. Thirteen synthetic-speech detectors, declared and hashed before any was scored. Holding speaker, words, performance, room and microphone fixed on real studio recordings, a neural-codec encode–decode — a transformation that generates no speech tokens — moved 9 of 13 detectors past Holm correction, three of them fully prospective at 47 of 47 utterances with rank-biserial exactly −1.000. It is not about codecs: Griffin-Lim, 1984, no neural network, moved 12 of 13. The codec-trained and ASVspoof-era families fail on opposite inputs, so choosing one detector is choosing which false positive to accept. I predicted a codec-trained detector would resist a codec pass; it collapsed hardest in the panel, and that prediction is published. Thirteen of my own claims were withdrawn under three rounds of independent adversarial audit. No mechanism identified, no listening test run. ai.bedvibe.studio/called-synthetic

What a Streaming Decoder Costs: A Five-Gate Benchmark of Twenty Neural Audio Decoders, and a Metric That Failed Its Own Validity Check — Zenodo, Sep 2026 · DOI 10.5281/zenodo.22798415 (v1.0.1). Twenty decoders through five pre-registered gates on identical audio. Only three arms truly stream: an arm qualifies only if stateful chunked decode reproduces full-context output and state is load-bearing, and the three causal FocalCodec configurations pass at state ratios of 3,530–36,740 against a threshold of 10, while fifteen are stateless chunking — the non-causal negative controls, run through the identical code path, sit at exactly 1. The sole pre-registered quality metric then failed the validity instrument built to catch exactly that: Griffin-Lim, zero trained parameters, scores best on mel-cepstral distance in six states of six. The pre-registration declared in advance that this would be a finding about the metric, so it is reported as one and no metric was substituted afterwards. A second measurement sharing no implementation found the same three survivors. Six rounds of independent adversarial audit; no measurement changed between versions. No listening test was run. ai.bedvibe.studio/decoder-benchmark

Continuous Vocal-Attribute Traversals Fall Below the Path Resolution of Current Speech Representations — Zenodo, Sep 2026 · DOI 10.5281/zenodo.22538235. A pre-registered null. Voice conversion and attribute editing draw a straight line between two points in a learned space and treat the coordinates between them as natural intermediate voices — never checked against ground truth, because in every such setting the path is generated by the model being evaluated. I recorded the path instead: a trained singer sliding continuously from clean phonation into rasp, growl or breathiness and back, 58 traversals, every intermediate point a real larynx. Two hypotheses and a validity gate with the power to stop the study were registered before any embedding was computed. The gate failed in both tiers, zero passes across six representations, SNR 0.53–1.85 against a required 3.0. The negative is verified rather than merely observed: speaker discrimination through the identical code path is positive and rises with window length in every encoder. H1 and H2 are unreachable with these instruments, NOT refuted.

Finite-Sample Decision Risk in Unseeded Quantum Compiler Benchmarking — Zenodo, Sep 2026 · DOI 10.5281/zenodo.22310059 (draft preprint, not submitted to a venue). Quantum compiler benchmarks accept or reject changes to production transpilers, but the compilers they measure are stochastic: at the pinned revision of IBM's Benchpress suite the Qiskit gym compiles without passing seed_transpiler, so every gate count is one draw from an unmeasured distribution, while the BQSKit gym in the same repository does seed its compiler. Quantified as finite-sample decision risk — the probability that a verdict from k runs disagrees with the verdict implied by the long-run mean. On bv_n140 over 400 seeds per version across 21 OS processes, Qiskit 1.4.3 → 2.0.0 has a long-run mean of +5.37% (95% CI +4.27–6.50%), while the suite's own three-unseeded-run protocol reports a ≥+10% regression 24.4% of the time; a disjoint 200-seed sample across ten fresh processes reproduces it at 22.6%. Four of my own claims were withdrawn and six defects found in my own apparatus during the audit.

Structure Substitutes for Scale in Cross-Document Temporal Reasoning — Zenodo, Aug 2026 · DOI 10.5281/zenodo.22169643. 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 0/38 at four sizes, refusing 92–100% of the time: the control holds. Given the identical facts as a structured chronology, an 8B model scored 28/38, matching a 14B model given prose — structure substituting for roughly a doubling of parameters. The advantage is gone by 14B, and scrambling the supplied order collapses accuracy to 10.5%, ruling out memorisation. After publishing I audited the scorer behind my own headline: it proved strictly conservative (1,786 items re-scored, 30 gains, 0 losses) but had understated the prose conditions, so my claimed gap over the best prose representation halved. The equivalence held; the gap did not. Corrected in place, not retracted. ai.bedvibe.studio/structure-not-scale

Speaker Encoders Disagree About Who You Are When You Shout: A Content-Matched Benchmark of Fourteen Speaker Encoders Under Expressive Phonation — Zenodo, Aug 2026 · DOI 10.5281/zenodo.22158030. Fourteen encoders over one frozen list of 11,935 trials; eight speakers × the same 1,360 sentences × six phonation states, so content is held constant while delivery varies. Equal error rate spans 0.047–0.233 on identical audio — five-fold, from encoder choice alone — with 29 of 91 pairwise comparisons surviving Holm–Bonferroni. The encoder shipped inside a widely used open-source TTS system is the worst of the fourteen for every speaker and every condition (13/13 comparisons survive correction); the mechanism is impostor-score compression, not degraded genuine scores. Intervals from a vertex bootstrap over the speaker graph, whose correction cost two of my own already-written claims. Ten retractions published alongside the result, including an 11.7 dB effect withdrawn against a 12.09 dB instrument artifact budget. Score tables, analysis code and the full decision record public; every statistic regenerates line for line from a CSV. github.com/Mormolykos/speaker-encoders-disagree

Intra-Speaker Vocal Variation and Speaker-Embedding Displacement: A Matched-Content Replication Across Three Encoder Architectures — Zenodo, Aug 2026 · DOI 10.5281/zenodo.21921958. Four speakers, matched sentences, three encoders (ECAPA-TDNN, ResNet, WavLM); 30/30 speaker × condition × encoder cells negative. Diagnoses differential measurement error in F0 tracking (Fisher p = 2.4×10⁻¹¹); pre-registered, with the adversarial prior-art audit and a withdrawn p = 0.013 result published in full.

The Loss Curve Is Not a Sufficient Statistic: Silent Objective Failures from Sentinel–Class Collisions in Neural Codec Language Models — Zenodo, Aug 2026 · DOI 10.5281/zenodo.21864659. A sentinel–class collision makes masked cross-entropy silently delete an entire class — usually the end-of-sequence token — while the loss curve still descends normally. Reproduced in 40 lines on CPU.

A 2D Linear-Elasticity Finite Element Solver Written from First Principles, Validated Against the Kirsch Stress Concentration Factor — Zenodo, Aug 2026 · DOI 10.5281/zenodo.21892064. Hand-written Delaunay mesher, isoparametric elements and CG solver; no FEA library, no scipy. Recovers the analytic stress concentration factor as 3.00002 against an exact 3 the code never receives. Ten gates pre-registered; nine pass, the failure published.

Selected Production Systems — live and publicly verifiable

ttsproof, trainproof, spkproof & notchecked (open source) — TTS failure-mode QA built from the published framework, benchmarking closed-source engines via a SpeechSDK integration; a training-reliability linter (98 rule IDs, 470 tests, 5 log formats, standard library only) that flagged a real run whose best checkpoint was step 99 while every saved checkpoint came from step 1600+; deterministic F0-contamination and trial-design checks for speaker-verification studies (9 rule IDs, exact tests, zero dependencies); and coverage accounting that reports what a validator could not check as eight typed terminal states instead of silence. github.com/Mormolykos

Grounded Portfolio AI Agent (RAG) — retrieval-augmented Q&A over my work, hardened against 32 adversarial cases (prompt injection, credential invention), shipped with a public evaluation report; the eval suite re-runs on every corpus change and has caught a live model regression. Three vector backends behind one interface — brute force 0.86 ms, FAISS 0.035 ms, Qdrant gRPC 1.3 ms — with the simplest one shipped on measurement, and the whole answer path traced span by span (request p50 12.41 → 2.91 ms once tracing showed a model client being rebuilt on every answer). …/portfolio/agent-evals

basalt — LLM model gateway (open source) — one OpenAI-compatible endpoint over several backends: per-key quota, token metering, bounded admission, deadline-aware retries and failover. Removing backpressure let 32 callers all receive HTTP 200 while 31 arrived after they had given up — a 100% success rate and one usable answer; bounded admission raised that to six. github.com/Mormolykos/basalt

Nine further live systems — BookProof, Audio Drama Studio, Audiobook Studio Editor, Talking Avatars, The Mirelands browser game, a public API tester, a Rust/Axum Amazon Polly service, an MCP mail assistant and a forced-alignment microservice — with screenshots and technical write-ups at tts.bedvibe.studio/portfolio.

Education & Languages

Hellenic Open University — university-level coursework in Natural Sciences: 222 ECTS completed, 11 of 12 required modules (mathematics, physics, chemistry, biology), and six laboratory courses completed where three are required. Final module scheduled September 2026.

Greek (native) · English (fluent, professional technical writing) · Norwegian (fluent) · advanced technical documentation in Greek and English.