Elena Raikova

Senior AI and Machine Learning Engineer / Technical Lead

13 years in production software engineering, including 5 years in end-to-end AI and ML systems.

Classical ML, deep learning, LLM/RAG evaluation, demand forecasting, agentic AI.

Download PDF CV Contact

Experience

DataArt logo

DataArt — Senior AI/ML Engineer / Technical Lead

2013 – Present · Remote / Buenos Aires
Most recent projects

Failure-attribution evaluation for RAG & agents

  • Reusable evaluation framework that turns a pass/fail score into a diagnosis — pinpointing which stage failed with transition matrices across runs
  • Scored semantic quality with an LLM judge calibrated against human labels
  • Verified diagnoses causally by perturbing the suspected component and checking whether the failure changed as expected

Demand forecasting + causal price optimization

  • Forecasted demand across 140+ parking facilities × 4 targets; benchmarked Chronos-2, Chronos + LightGBM, Prophet, LSTM, DeepAR, TFT. Reduced MAPE from ~50% to < 15% on a 30-day horizon
  • Calendar/event features, lag/rolling features, cross-facility features, booking curves, missing-data handling, leakage-safe time splits
  • Multivariate, global training across facilities and targets, with transfer to cold-start markets
  • Price elasticity under endogeneity via IV/2SLS, regression discontinuity, control function, and panel fixed effects
  • Revenue-maximizing price from a monotone-constrained LightGBM demand model pinned to the causal elasticity; served via FastAPI over reproducible DVC pipelines.
Case study dataart.com

Agentic AI workflow automation

  • Agentic workflow automation for a manual royalty-approval process (20k requests/year): LLM-assisted agents embedded in a Power Automate workflow to extract, normalize, and validate Excel/PDF requests, with decision traces shown in Teams adaptive cards.
  • Separated deterministic business rules from LLM-assisted decisions, with confidence-based abstention, human approval, sanctions checks, subscription/price-book validation, idempotent processing, malformed-document handling, and full audit trails

Enterprise helpdesk Teams chatbot

  • MS Teams chatbot for a 6k-user corporate knowledge base: answered user questions and routed requests to the correct helpdesk category using classification methods
  • Built a one-pass ModernBERT decision classifier that scores arbitrary answer sets as parallel isolated branches; redesigned the attention mask to prevent cross-option leakage, improving accuracy from ~55% to ~85%.
  • Hybrid RAG over ~54K multilingual Confluence docs indexed in Qdrant, GLiNER entity scoping
  • Jira Service Management integration: opens tickets from chat, syncs status/comments to the Teams thread
Case study dataart.com

Sensor-data lakehouse with real-time ML

  • Implemented and evaluated anomaly-detection and forecasting models on sensor data within a Kafka/Dagster/Dremio/S3/MinIO platform
  • Benchmarked Chronos-2 against linear and Random Forest baselines

Earlier engineering (2013–2022). NLP for trading-form automation and CV parsing/matching; a carrier-grade Ethernet test & diagnostics platform with a Java engine and Python device connector over SSH/TL1/SOAP; Django CRM and an Open edX LMS. Stack spanned Python, Java, JavaScript, Scala, SQL, Docker, Kubernetes.

Independent research and writing

  • LLMs: rule induction and LLM-guided program search; calibrated decisions with RL (RLCR); post-training with SFT, LoRA/QLoRA, RLHF/RLAIF, DPO, GRPO; attention variants (MQA, GQA, MLA); agent architectures and temporal grounding
  • Deep learning: mechanistic interpretability of GPT-2 small (sparse autoencoders, activation patching, induction heads, steering); knowledge distillation on CIFAR-10; an AlphaZero-style Go player from scratch
  • ML systems: how GPUs run ML and matrix multiplication; LLM inference on a T4 GPU (prefill vs decode); from a PyTorch call to a CUDA kernel
  • Retrieval: embeddings from LSA and word2vec to BERT, E5 and BGE-M3; chunking strategies; text clustering with UMAP and HDBSCAN
  • Forecasting & causal inference: ARIMA/SARIMAX, LSTM, TCN, anomaly detection; endogeneity, IV/2SLS, regression discontinuity, panel fixed effects, pricing elasticity

Personal projects

  • Sharada typed decisions about text in one forward pass
    pip install sharada
  • torchrolling pandas-style rolling statistics for PyTorch, fast on GPU
    pip install torchrolling
  • tidyenv typed environment variables for Python
    pip install tidyenv

Skills

Languages
  • Python
  • Java
  • JavaScript
  • Scala
  • SQL
ML / DL
  • PyTorch
  • scikit-learn
  • deep learning
  • classical ML
  • model evaluation
  • hyperparameter optimization
  • mechanistic interpretability (SAEs, activation patching)
  • knowledge distillation
LLM / GenAI
  • RAG
  • agents
  • LLM-as-a-judge
  • fine-tuning
  • LoRA SFT (MLX)
  • rejection-sampling fine-tuning
  • test-time training
  • RL with calibration rewards
  • program synthesis
NLP
  • spaCy
  • GLiNER
  • NLTK
  • entity extraction
  • classification
  • document intelligence
Forecasting
  • Chronos/Chronos-2
  • LSTM
  • DeepAR
  • quantile forecasting
  • booking curves
  • price optimization
  • what-if
Environment
  • Power Automate
  • Dynamics 365
  • Dataverse
  • Azure AI Document Intelligence / AI Search
  • MS Teams
  • Jira Service Management
  • Confluence
  • Kafka
  • Dagster/Airflow
  • Dremio
  • DVC
  • PostgreSQL
  • MongoDB
  • Qdrant
  • Snowflake
  • Redis
  • FastAPI
  • Streamlit
  • Docker/Podman
  • Kubernetes
  • MinIO
  • Terraform
  • GitHub Actions
  • Sentry
  • GPU/CUDA
  • Kaggle
Cloud
  • Azure
  • AWS

Education

M.Sc., Applied Mathematics, Informatics and Mechanics
Voronezh State University, Russia