Ilya Yaverbaum

Tel Aviv, Israel

Ilya Yaverbaum

AI/ML Engineer

AI/ML Engineer with hands-on experience in deep learning, data science, agentic AI and LLMs — taking ideas end-to-end from prototype to production. Unit 8200 alumnus with a strong background in statistics, data analysis and applied research, and a track record of building production-grade ML algorithms and refining them for accuracy and efficiency.

Download CVGitHubLinkedInEmail
  • Python
  • C
  • C++
  • Java
  • SQL
  • LLM
  • LLM Agents
  • RAG
  • Tool-Calling
  • MCP
  • LangChain
  • LangGraph
  • OpenAI API
  • Embeddings
  • Vector DBs
  • ChromaDB
  • FAISS
  • Claude Code
  • Anthropic API
  • Gemini
  • Groq
  • Ollama
  • Multi-Agent Systems
  • LLM Evaluation
  • LLM-as-Judge
  • Structured Outputs
  • Prompt Caching
  • BM25 / Hybrid Search
  • Vision-Language Models
  • Local / Self-Hosted LLMs
  • LiteLLM
  • Instructor
  • Sentence-Transformers
  • Offline Evals
  • Graded / Rubric Evaluation
  • Deterministic Guardrails
  • Tracing & Observability
  • Cost Telemetry
  • Regression Testing
  • Provider Abstraction
  • Model Routing
  • Cost Optimization
  • Test Doubles / Fakes
  • Offline-First Testing
  • Release Gating
  • PyTorch
  • TensorFlow
  • Hugging Face
  • scikit-learn
  • Transformers
  • Fine-Tuning
  • Active Learning
  • GPU Training
  • Human-in-the-Loop
  • Data-Centric ML
  • Uncertainty Sampling
  • Mislabel Detection
  • Model Deployment
  • Model Evaluation
  • Experiment Tracking
  • Hyperparameter Optimization
  • MLflow
  • Optuna
  • Hydra
  • Docker
  • CI/CD
  • Git
  • Azure
  • pytest
  • Vercel
  • Render
  • GitHub Actions
  • Data Labeling Pipelines
  • Model Monitoring
  • Error Analysis
  • Train/Serve Skew
  • Automated Labeling
  • NumPy
  • Pandas
  • SciPy
  • Spark
  • Dash
  • Time-Series Analysis
  • Forecasting
  • Anomaly Detection
  • Statistical Analysis
  • Signal Denoising
  • Outlier Detection
  • Data Enrichment
  • FastAPI
  • Django
  • REST API
  • Async Python
  • asyncio
  • MongoDB
  • Redis
  • Pydantic
  • SQLAlchemy
  • PostgreSQL
  • SQLite
  • Streamlit
  • Concurrency
  • API Versioning
  • Scheduling
  • Telegram Bot API
  • Flask
  • API Integration
  • Response Caching
  • Signal Processing
  • Radar Technology
  • RF Systems
  • Real-Time Systems
  • Algorithm Development
  • Radar Signal Processing
  • Equipment Validation
  • Field Data Collection
  • Hebrew — Native
  • English — Fluent
  • Russian — Native

Experience

ML Engineer

DRS RADA Technologies

2025 — Present

  • Led development of data labeling workflows using Active Learning, cutting analysis time and manual labeling workload by ~50% while surfacing mislabeled samples in previously annotated data.
  • Built performance-analysis tooling that exposed lab-to-production inference gaps and mapped model overfit and underfit regions, catching ~3 algorithmic bugs per model release before deployment.
  • Built end-to-end ML pipelines from prototype to production. Trained and fine-tuned transformer-based models on GPU.
  • MLflow
  • TensorFlow
  • PyTorch
  • Dash
  • SciPy
  • Optuna
  • Hydra
  • Active Learning
  • Transformers
  • Fine-Tuning
  • GPU Training
  • Python
  • Data Labeling Pipelines
  • Model Evaluation
  • Experiment Tracking
  • Uncertainty Sampling
  • Mislabel Detection

Data Science Analyst

DRS RADA Technologies

2021 — 2025

  • Developed time-series analysis, forecasting and anomaly-detection algorithms; signal noise and outlier reduction cut both training and inference time and raised detection accuracy.
  • Automated data enrichment and labeling by integrating external APIs with radar tracking systems, recovering thousands of previously missed tracks.
  • Applied statistical analysis and machine learning to improve system reliability under real-world conditions.
  • Time-Series Analysis
  • Forecasting
  • Anomaly Detection
  • Statistical Analysis
  • Python
  • NumPy
  • Pandas
  • SciPy
  • API Integration
  • SQL
  • scikit-learn
  • Signal Denoising
  • Outlier Detection
  • Data Enrichment
  • Automated Labeling

Field Engineer

DRS RADA Technologies

2018 — 2021

  • Conducted field experiments and radar tests to validate system performance in real-world environments.
  • Ran customer training, technical support and system demonstrations through deployment.
  • Radar Systems
  • Field Testing
  • System Validation
  • Customer Training
  • Technical Support
  • Field Data Collection
  • Test Planning & Reporting
  • System Deployment

RF Engineer

Intelligence Corps, Unit 8200

2015 — 2018

  • Responsible for RF system design, including transmission and reception planning, and lab testing with Network Analyzers and Spectrum Analyzers.
  • Led product integration, equipment validation and implementation across laboratory and field environments.
  • RF Design
  • Signal Processing
  • Network Analyzers
  • Spectrum Analyzers
  • Product Integration
  • Lab Testing
  • Transmission Planning
  • Equipment Validation

Projects

Production sensor ML at DRS RADA, and three independent LLM systems — public end to end.

  • Active-learning loop: a model trains on the labeled data pool, identifies uncertain samples in the unlabeled pool, and requests human review, which feeds new labels back into training.

    Active Learning — Labeling Pipeline

    ML Engineer @ DRS RADA Technologies

    Labeling large volumes of sensor data is the bottleneck in every production ML cycle, and the labels that already exist are not necessarily right. I led development of labeling workflows built around Active Learning, so the pipeline ranks which samples are worth a human's attention instead of working the queue in order — and runs the same scoring pass over already-annotated data, where disagreement between model and label surfaces likely mislabels.

    • Uncertainty-based sample ranking
    • Mislabel detection across prior annotations
    • In production · ~50% less manual labeling
    • Active Learning
    • PyTorch
    • MLflow
    • Model Evaluation
    • Python
    • Uncertainty Sampling
    • Mislabel Detection
    • Data Labeling Pipelines
    • Human-in-the-Loop
  • Lab versus production model behaviour side by side: clean converging accuracy curves in the lab, degraded and unstable curves in deployment, with a heatmap magnified to locate the failing regions.

    Performance Analysis — Overfit Mapping

    ML Engineer @ DRS RADA Technologies

    Models that look strong in the lab do not always hold up at inference time, and an aggregate accuracy number hides where they fail. I built analytical tooling that compares lab and production behaviour and maps where a model overfits and underfits across its input space, turning a single metric into a picture of which regions are actually trustworthy — wired into the release path so the check runs before deployment, not after.

    • Lab-vs-production inference comparison
    • Overfit / underfit region mapping
    • Gates every release · ~3 bugs caught per model
    • Model Evaluation
    • Dash
    • SciPy
    • Optuna
    • Hydra
    • Experiment Tracking
    • Error Analysis
    • Train/Serve Skew
    • Release Gating
    • Model Monitoring
  • The two costs of unrefined sensor data: a noisy time-series stream lengthening the training cycle, and anomalies hidden inside that noise reducing detection accuracy.

    Time-Series Anomaly Detection & Enrichment

    Data Science Analyst @ DRS RADA Technologies

    Sensor time-series arrives noisy, and the noise costs twice — once in training time, once in detection accuracy. I developed time-series analysis, forecasting and anomaly-detection algorithms with signal-noise and outlier reduction ahead of training, and separately automated data enrichment and labeling by integrating external APIs with tracking systems to recover signal the pipeline had been dropping.

    • Forecasting + anomaly detection on sensor streams
    • Noise and outlier reduction ahead of training
    • In production · thousands of tracks recovered
    • Time-Series Analysis
    • Forecasting
    • Anomaly Detection
    • Statistical Analysis
    • NumPy
    • Pandas
    • Signal Denoising
    • Outlier Detection
    • API Integration
    • Data Enrichment
    • SciPy

Independent projects

  • Northbeam Support Copilot

    Independent project

    A tool-calling support agent for a fictional radar company, built so a support engineer can ask in plain language and get an answer that shows its work. Twelve read-only tools in three layers — records, calculation, retrieval — run in a bounded loop; their real results become the evidence for one structured call. The interesting decision is the refusal layer: where the right answer is enumerable, deterministic rules decide the outcome rather than the model, so 62% of declines are settled in code and the answer can say which rule fired.

    • 12 read-only tools across records, calculation and hybrid BM25 + embedding policy search
    • Deterministic refusal layer — 62% of declines decided in code, not by the model
    • 196 offline tests · 77 graded evals · trace-based KPI dashboard
    Read the code →
    • Python
    • LLM Agents
    • Tool-Calling
    • LiteLLM
    • Instructor
    • Structured Outputs
    • RAG
    • BM25 / Hybrid Search
    • Sentence-Transformers
    • Deterministic Guardrails
    • Offline Evals
    • Graded / Rubric Evaluation
    • Tracing & Observability
    • KPI Dashboard
    • Provider Abstraction
    • Flask
    • pytest

  • Beat the ASPP

    Independent project

    Multi-agent equity research that decomposes the problem the way a research desk does. Three specialist agents — technical, fundamental, sentiment — run in parallel, each with its own tools and Pydantic-enforced output, then a strategist agent reconciles them into one verdict. The decision worth stealing is two-stage write-then-extract: Opus writes the narrative, Sonnet extracts the JSON, which holds the quality and costs roughly five times less than asking one model for both.

    • Three analyst agents in parallel; a strategist reconciles one defensible verdict
    • Opus writes the prose, Sonnet extracts the JSON — same quality, ~5× cheaper
    • Provider abstraction across Anthropic / Groq / Ollama, with per-call cost telemetry
    • Research tooling, not investment advice — no verdict here is a recommendation to trade
    Read the code →
    • Multi-Agent Systems
    • Structured Outputs
    • Pydantic
    • Provider Abstraction
    • Anthropic API
    • Groq
    • Ollama
    • Prompt Caching
    • Cost Telemetry
    • Concurrency
    • SQLAlchemy
    • PostgreSQL
    • Streamlit
    • Scheduling
    • Telegram Bot API

  • Menu Alchemist

    Independent project

    Photograph a drink or a menu and get the reconstructed recipe — ingredients, steps, glassware, estimated ABV — plus a shopping list of only what you don't already own. The split is the point: the vision model does extraction and nothing else, while pantry detection, package sizing, serving math and retailer matching are pure Python, so the part that has to be right every time is unit-tested rather than prompted.

    • Pluggable vision providers behind one interface — Gemini, local Ollama, or a fake for tests
    • Vision model extracts; deterministic Python decides what actually reaches the buy list
    • 86 tests that need no network and no API key · installable PWA · deployed on Render
    Read the code →
    • FastAPI
    • REST API
    • API Versioning
    • Gemini
    • Ollama
    • Vision-Language Models
    • Vision Extraction
    • Deterministic Guardrails
    • Provider Abstraction
    • Test Doubles / Fakes
    • pytest
    • Pydantic
    • SQLAlchemy
    • Response Caching
    • PWA
    • Vanilla JS
    • Render

Skills

From RF and signal processing to production ML pipelines and evaluated LLM systems — the stack behind the work above.

01

Programming Languages

  • Python
  • C
  • C++
  • Java
  • SQL
02

Agentic & Generative AI

  • LLM
  • LLM Agents
  • RAG
  • Tool-Calling
  • MCP
  • LangChain
  • LangGraph
  • OpenAI API
  • Embeddings
  • Vector DBs
  • ChromaDB
  • FAISS
  • Claude Code
  • Anthropic API
  • Gemini
  • Groq
  • Ollama
  • Multi-Agent Systems
  • LLM Evaluation
  • LLM-as-Judge
  • Structured Outputs
  • Prompt Caching
  • BM25 / Hybrid Search
  • Vision-Language Models
  • Local / Self-Hosted LLMs
  • LiteLLM
  • Instructor
  • Sentence-Transformers
03

LLM Reliability & Evaluation

  • Offline Evals
  • Graded / Rubric Evaluation
  • Deterministic Guardrails
  • Tracing & Observability
  • Cost Telemetry
  • Regression Testing
  • Provider Abstraction
  • Model Routing
  • Cost Optimization
  • Test Doubles / Fakes
  • Offline-First Testing
  • Release Gating
04

ML & Deep Learning

  • PyTorch
  • TensorFlow
  • Hugging Face
  • scikit-learn
  • Transformers
  • Fine-Tuning
  • Active Learning
  • GPU Training
  • Human-in-the-Loop
  • Data-Centric ML
  • Uncertainty Sampling
  • Mislabel Detection
05

ML Engineering & MLOps

  • Model Deployment
  • Model Evaluation
  • Experiment Tracking
  • Hyperparameter Optimization
  • MLflow
  • Optuna
  • Hydra
  • Docker
  • CI/CD
  • Git
  • Azure
  • pytest
  • Vercel
  • Render
  • GitHub Actions
  • Data Labeling Pipelines
  • Model Monitoring
  • Error Analysis
  • Train/Serve Skew
  • Automated Labeling
06

Data & Analysis

  • NumPy
  • Pandas
  • SciPy
  • Spark
  • Dash
  • Time-Series Analysis
  • Forecasting
  • Anomaly Detection
  • Statistical Analysis
  • Signal Denoising
  • Outlier Detection
  • Data Enrichment
07

Backend & Data Stores

  • FastAPI
  • Django
  • REST API
  • Async Python
  • asyncio
  • MongoDB
  • Redis
  • Pydantic
  • SQLAlchemy
  • PostgreSQL
  • SQLite
  • Streamlit
  • Concurrency
  • API Versioning
  • Scheduling
  • Telegram Bot API
  • Flask
  • API Integration
  • Response Caching
08

Signal Processing & RF

  • Signal Processing
  • Radar Technology
  • RF Systems
  • Real-Time Systems
  • Algorithm Development
  • Radar Signal Processing
  • Equipment Validation
  • Field Data Collection
09

Languages

  • Hebrew — Native
  • English — Fluent
  • Russian — Native

Education

2020 — 2023

B.Sc. Computer Science

Holon Institute of Technology (HIT)

Studied alongside the Field Engineer and Data Science roles at DRS RADA.

2013 — 2015

Associate Degree, Electronics and Computers

Ort Hermelin College

Technological Track of the Intelligence Corps — electronics and computer systems.