Fișier CV încărcat

Module 3

To detect anomalies at the third level of the catalog, a DeBERTa v3 + DeepOD module was built. In this design, DeBERTa v3 serves as the semantic encoder, outperforming other LLMs in capturing subtle lexical nuances (brand, model, form factor), while DeepOD—a neural framework that implements DevNet—performs fine‑grained anomaly detection in texts that are lexically hard to moderate.

Such a combination delivers the following results:

• Higher anomaly‑detection accuracy for the catalog’s third level.
  • Homogeneous vocabulary — DeBERTa forms a tight “normality cluster”, and DeepOD flags even minor semantic deviations.
  • Reduced overfitting — the model learns genuine sub‑category patterns instead of memorizing top‑level noise.
• Controlled computing costs despite the large L3 data volume.
  • DevNet’s lightweight MLP architecture is CPU‑friendly, enabling millions of predictions without a dedicated GPU farm.
  • The project scales linearly: DeepOD keeps operating expenses in check, so adding new product SKUs only increases the number of vectors, not the model’s complexity.
• Rapid adaptation to new products and promotions.
  • Fast retraining thanks to a low entry threshold: the weakly‑supervised mode lets us label data quickly and follow market trends.
  • Encoder invariance → stable embeddings: DeBERTa does not “drift,” so retraining never breaks previous predictions.

Stack: PyTorch, PySpark, Pandas, NumPy, PyArrow, HuggingFace Transformers, Ray Tune, Apache Airflow, MLflow, DeepOD, Prometheus + Grafana, Docker.

Houlsby Adapters proof‑of‑concept (1–3 % of weights):

Detachable modules can be plugged in within hours to support new analytical tasks—e.g., attribute validation—without pausing the core service, turning the solution into a flexible multi‑task platform for content‑quality control.

Stack: PyTorch, Hugging Face Transformers, Hugging Face PEFT (LoRA), Adapter‑Transformers, PySpark, scikit‑learn, pandas / NumPy, Hugging Face Datasets, ONNX Runtime GPU + Optimum, BatchScheduler, Docker, MLflow, TensorBoard, Evidently AI, pytest.

Module 4

Developed a universal long‑text preprocessing module for the product catalog. The module automatically splits each description into overlapping fragments, forms a single [CLS] vector for the product, and stores all vectors in a FAISS index—so every downstream model (anomaly detector, classifier, search engine) receives a complete and consistent representation of the item.

Key engineering steps and benefits:

• Sliding‑window segmentation of 510 / 50 tokens and mean pooling of [CLS] vectors give a unified document representation. Instead of truncating text, the module cuts descriptions into overlapping fragments and stitches them back together, enabling the model to read 100 % of the content. Validation showed ≈ +9 pp improvement in anomaly‑detector recall—fewer missed errors, fewer returns, and fewer customer complaints.
• A custom DocumentDataset and dynamic collate_fn (batch‑max‑padding + attention mask) assemble batches without redundant copies. The in‑house DocumentDataset + collate_fn fill batches directly in memory, avoiding extra copying and serialization. Result: the same server processes 40 % more product cards per shift, reducing infrastructure costs.
• Normalized embeddings are indexed in FAISS (IVF) for millisecond‑level k‑NN search. Product embeddings are stored in the FAISS index; nearest‑neighbour queries take only milliseconds. An analyst immediately sees how a suspicious item differs from normal ones and can decide without delay → moderation is almost real‑time, improving user experience.

Stack: Faiss (IVF), Hydra, PySpark, HuggingFace Transformers / Tokenizers, Kubernetes, NumPy, pandas, PyTorch, Hugging Face Tokenizers, Docker.

Module 5

Deployed an end‑to‑end monitoring module for NLP‑model quality: the core framework tools (TensorFlow / PyTorch) automatically log ROC AUC, Precision, and a custom label‑free heuristic during training and almost in real time; a library‑based drift‑tracking mechanism sits on top of these logs, checks fresh values every ≤ 60 s, and sends notifications. Thanks to this, degradations are detected before a model reaches the version catalog.

Key engineering steps and benefits:

• TensorBoard hooks that compute ROC AUC, Precision, and the custom label‑free heuristic during training and immediately after inference.
  • The metrics ROC AUC, Precision, and the custom heuristic are calculated at every training run and practically in real‑time.
• 60‑second drift callback that reads the latest TensorBoard logs and sends a web notification when metrics deviate.
  • An alert fires if metrics drift is detected in under 60 seconds.
• Cut‑off gate with automatic threshold selection along the ROC curve (Youden J criterion).
  • Transparent metric cut‑offs enable clear Go/No‑Go decisions for releases and allow proactive retraining planning.

Stack: TensorBoard, tensorboardX, scikit‑learn, PyTorch, Matplotlib, argparse, PyTest, NumPy, Pandas, Docker, asyncio + requests, logging.

Module 6

Created a module that automates metric monitoring and model updates in the production NLP pipeline—from auto‑versioning to a safe blue‑green switch—so the business always sees trustworthy KPIs and can ship improvements with zero downtime.

What was done — key engineering steps:

• Unique timestamp + SHA‑256 auto‑versioning with a parallel TorchScript dump — each model version is captured as a “serial number,” making it easy to locate, compare, or roll back—no lost files and no chaos in production.
• Fast sanity gate on 50 edge cases in torch.no_grad() — an automatic “express exam” filters out faulty weights in seconds, ensuring a degraded model never reaches users and dashboards stay honest.
• Atomic blue‑green switch of the current_model symlink — traffic is routed to the new model instantly with zero downtime; if anything goes wrong, rollback is a single command.

Stack: PyTorch, TorchScript, pytest, hashlib + json, Bash, syslog, Cron, argparse, Docker.