Skip to content
View stelioszach03's full-sized avatar
🏠
Working from home
🏠
Working from home

Highlights

  • Pro

Block or report stelioszach03

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
stelioszach03/README.md

Stelios Zacharioudakis

ML engineer in Athens. BSc Computer Science, NKUA — expected Jun 2026.

Head Engineer since Jun 2022 at the Paphos Medical Association, building asklepiosmed.org — the platform for a 480-doctor member registry, on an Express + PostgreSQL + React 18 + iOS monorepo.

Live demos: SMT-verified reasoning · PHI redaction · graph fraud scoring · subway anomaly detection

Project Measured result Evidence in repo
TrustQueryNet 83.5% ± 0.6 on HAM10000 over 5 seeds, ECE 0.0445 — and 56.9% under external shift to ISIC 2019 per-seed CSVs, paired bootstrap CIs
EuroSAT benchmark LoRA reaches 98.22% training 0.36% of parameters (310,292 / 86,108,948), 3.7× faster than full fine-tuning saved notebook cell outputs
Speculative decoding lab 1.39× aggregate latency speedup on A100 (1.46× on math prompts); the Transformers assisted path was slower than baseline in all 9 cells paper/data/*.csv
MTA-Scan recall@20 1.00 vs 0.625 for the best of three baselines, at the highest false-alarm rate (0.030), on 216 rows / 3 incidents docs/generated/replay/metrics.json
LimitForge RLS 92.57% line coverage (536 / 579), 79.55% branch coverage.xml

The research repos are solo, unpublished manuscripts — not peer-reviewed. Where a result rests on synthetic data or a small replay set, the repo says so.

ML — PyTorch · HuggingFace · vLLM + EAGLE-3 · PEFT/LoRA · Z3 · LightGBM · River · spaCy · CUDA / A100 Systems — Python · FastAPI · Node/Express · React · PostgreSQL · Redis · Kafka · Docker · GitHub Actions · Prometheus · gRPC

stelioszach.com · LinkedIn · stelios@stelioszach.com

Pinned Loading

  1. colab-speculative-decoding-speed-lab colab-speculative-decoding-speed-lab Public

    Speculative decoding on A100 80GB: 1.39x aggregate with vLLM+EAGLE-3; Transformers-assisted was slower.

    Jupyter Notebook

  2. graph-fraud-command-center graph-fraud-command-center Public

    Real-time graph fraud scoring, 14 topology features. Local load test: 457 req/s, p95 3.27 ms, 0 errors.

    Python

  3. TrustQueryNet TrustQueryNet Public

    Noise-robust ConvNeXt-Tiny on HAM10000: 83.5% over 5 seeds, ECE 0.0445, 56.9% under shift to ISIC 2019.

    Python

  4. llm-smt-verifiable-reasoning llm-smt-verifiable-reasoning Public

    LLM + Z3 SMT reasoning with conflict-directed search, evaluated over 7,500 compute-matched runs.

    Python