Ehud Vaknin

Hi, I'm Ehud Vaknin

Junior Software Engineer who enjoys developing full-stack applications and creating MLOps pipelines. Experienced in building and deploying a production-scale SaaS platform end-to-end using TypeScript, Next.js, React, Supabase, Python, and FastAPI.

Download CV Download MLOps CV
Profile

About Me

I'm a fourth-year B.Sc. Software Engineering student (specialized in Data Science, GPA: 92) with hands-on experience as a Software Engineer at Mind Artists Journey, where I directed end-to-end development of a production-scale B2B SaaS platform - from architectural design and database schemas to frontend UI and AI integrations.

I build across the full stack using TypeScript, Next.js, React, Supabase, Python, and FastAPI, with experience in AI integrations (OpenAI, Anthropic, Gemini), CI/CD pipelines, Docker, Redis, and Vercel deployments.

I'm actively seeking a full-time junior software development position to apply my technical expertise, contribute to collaborative projects, and continue developing skills in real-world environments.

Hebrew - Native | English - Fluent

Education

B.Sc. in Software Engineering

2022 - 2026

Sami Shamoon College of Engineering

Specialization: Data Science GPA: 92

Work Experience

Software Engineer

2025 - 2026

Mind Artists Journey

Led development of a production-scale B2B SaaS platform for coaches and business owners, owning the system end-to-end - architecture, database schemas, REST APIs, frontend, AI integrations, and CI/CD.

Leadership

  • Recruited and managed an offshore developer across time zones - scoping R&D tasks, reviewing PRs on GitHub, and coordinating release cadence.

AI & Backend

  • Designed and shipped AI-powered lead engagement and nurturing systems, integrating OpenAI, Anthropic, and Gemini for content and transcript workflows, with memory-based automation for context retention.
  • Engineered an embeddings-based RAG pipeline and modular knowledge architecture deploying brand-specific context across five AI surfaces - powering semantic search and a real-time in-product AI coach with OpenAI streaming.
  • Implemented secure GoHighLevel CRM integrations with signature-validated webhooks and idempotent contact sync; cut AI messaging latency via server-side RPC aggregation and React Server Components (RSC) prefetching.

Learning Management System

  • Architected an LMS with a hierarchical content tree, automated enrollment policies, and a rule-based milestone engine with auto-generated certificates.
  • Integrated Mux and YouTube for video analytics and progress tracking; used Supabase Realtime for live session state management and gamification.

Infrastructure & Reliability

  • Built CI/CD pipelines with parallelized Playwright E2E tests; profiled artifact upload/download steps and removed bottlenecks, cutting total pipeline time from 20 to 12 minutes.
  • Managed multi-environment deployments (production + preview) on Vercel and Supabase.

Machine Learning & MLOps

Full ML lifecycle - from raw data and feature engineering to trained, evaluated, and deployed models. Computer vision, anomaly detection, and LLM-based RAG systems in production.

90%
mAP@0.5 · YOLO
170K
Curated images
Multi-LLM
RAG in production
Israel Housing Dashboard - microservices data pipeline
Data Pipeline · MLOps 2025

Israel Housing Dashboard

End-to-end microservices stack for collecting, enriching, and serving Israeli real-estate data - orchestrated with Docker Compose. Five dedicated services, a MongoDB feature store with ~88 enriched features, seven trained ML models behind a multi-model serving API, and an interactive Next.js dashboard with a zoom-aware deck.gl map over 270K+ records.

Data Pipeline · ETL

  • Playwright scrapers across 5 sources: Govmap (Rashut), odata.org.il, Tax Authority, Madlan, CBS
  • ETL: raw_records → features_enriched (~88 features) with 2dsphere + H3 indices in MongoDB
  • OSM spatial feature engineering: distances to schools / parks / water, POI counts at multi-radii, land-use ratios
  • Macro-economic enrichment: CPI, prime rate, GDP growth, USD/ILS, real interest rate, housing-CPI gap

ML Serving

  • 7 trained models: LightGBM × 4 variants, CatBoost, Stacked × 2 (R² + MAE tracked per model)
  • FastAPI + LRU artifact cache · lazy loading · per-model and multi-model comparison endpoints
  • Live champion switching via env var - no service rebuild required
  • Reproducible runs: model.joblib + metrics.json + run_metadata.json per artifact

Frontend & Map

  • Next.js 16 + React 19 · MapLibre GL + deck.gl 9 · TanStack Query · Zustand · Recharts
  • Zoom-aware H3 clustering: res 5 (~8.5 km) → 7 → 8 → individual points at high zoom
  • Mongo $group aggregation on indexed H3 field - millisecond response over 270K records
  • Pages: live map, stats (timeseries / YoY heatmaps / histograms), property detail, AI prediction playground

Infrastructure & Geocoding

  • Docker Compose orchestration across 5 services + Makefile shortcuts for ETL / health / champion
  • Address-level geocoding via Govmap - idempotent, resumable, ~63K-address cache covering ~295K records (4.6× hit ratio)
  • 2dsphere + text + H3 indices · MongoDB Atlas-compatible · async Motor driver throughout
  • Multi-environment .env · health checks · streamlit QA tool
Docker Compose FastAPI MongoDB LightGBM CatBoost XGBoost scikit-learn pandas geopandas Playwright H3 joblib Next.js MapLibre deck.gl
View on GitHub
Bird vs. Drone Detection - YOLO bounding box around a drone in flight Computer Vision
B.Sc. Final Project ★ Completed with Honors
2026

Bird vs. Drone Detection (YOLO)

Custom Ultralytics YOLO object detector for airspace monitoring - distinguishing birds from drones in real time on video streams, targeting low-latency edge deployment.

  • Curated & labeled a ~170K-image dataset (53% drone / 47% bird) with 80/10/10 train/val/test split
  • Generated 12K synthetic drone images in Unreal Engine 5.4 to expand coverage of rare viewpoints
  • Achieved mAP@0.5 of 0.90 (mAP@0.5:0.95 = 0.66), precision 0.95, recall 0.82 on held-out validation
  • Real-time inference on video streams for edge deployment
PyTorch Ultralytics YOLO Python Computer Vision Unreal Engine 5
ML Pipeline
Research Project 2025

Acoustic Emission ML Pipeline

End-to-end ML pipeline on raw acoustic-emission sensor data - from custom parser to clustering and anomaly detection, with a reproducible workflow for benchmarking new algorithms.

  • Custom parser for proprietary .TXT instrument files with header & merged-cell edge-case handling
  • 19-feature signal representation scaled with RobustScaler
  • K-Means clustering (k selected via silhouette analysis, PCA-inspected)
  • Anomaly-detection benchmarking: Isolation Forest, LOF, and AutoEncoder with ROC-AUC
  • Reproducible workflow: hyperparameter grid search, serialized artifacts (joblib), JSON run metadata
scikit-learn pandas NumPy AutoEncoder joblib Python

Production MLOps · Mind Artists Journey

2025 - 2026

Shipped production AI systems on a live B2B SaaS platform - owning the MLOps lifecycle end-to-end, from model orchestration to CI/CD and multi-environment deployments.

AI Systems

  • Modular "skills" framework orchestrating OpenAI, Anthropic, and Gemini for lead nurturing
  • Embeddings-based RAG pipeline powering semantic search and an in-product AI coach
  • OpenAI streaming inference with memory-based automation for context retention

Infrastructure

  • CI/CD with parallelized E2E tests - cut pipeline time 20 → 12 min
  • Dockerized services across production and preview environments
  • Multi-environment deployments on Vercel and Supabase
OpenAI Anthropic Gemini RAG Embeddings Docker GitHub Actions Vercel Supabase
See also in Projects: PulseBoard (real-time event pipeline with anomaly highlights over ClickHouse).

Technical Stack

Languages & Frameworks

TypeScript TypeScript
Next.js Next.js
React React
Python Python
C C
C++ C++
C# C#
Java Java
SQL SQL
.NET .NET

ML & Data Science

PyTorch PyTorch
scikit-learn scikit-learn
pandas pandas
NumPy NumPy
Ultralytics Ultralytics

Infrastructure & Tools

FastAPI FastAPI
Supabase Supabase
Vercel Vercel
Docker Docker
Stripe Stripe
Mux
Redis Redis
Playwright Playwright
GitHub GitHub
MongoDB MongoDB

Selected Projects

PulseBoard - self-hosted product analytics platform with live event firehose
Data & Backend 2026

PulseBoard

A self-contained, end-to-end product-analytics platform - a miniature Mixpanel / PostHog - built as a 6-stage pipeline (SDK → FastAPI ingestion → Postgres outbox → Redpanda → ClickHouse sink → Redis pub/sub → Next.js dashboard). Three synthetic demo apps emit live events through the real SDK so the full stack is exercised on every cold start.

  • Crash-safe Python SDK with disk-buffered batching and at-least-once delivery
  • Transactional outbox + Redpanda relay for exactly-once-ish ingestion
  • Sub-second live firehose via Redis pub/sub and Server-Sent Events
  • Cohort retention heatmaps, anomaly highlights, and per-app KPIs over ClickHouse
  • One-command local stack via Docker Compose with healthchecks and 90-day seeder
Python FastAPI Next.js TypeScript ClickHouse Postgres Redpanda / Kafka Redis SSE Tailwind Docker Compose
View on GitHub
n8n WhatsApp workflow automation

Whatsapp Order Chatbot Assistant

Exposes REST APIs for managing orders and items, featuring a Hebrew intent detection utility. Includes a Redis-based chat memory router for chat session state and utilizes a MySQL database via SQLAlchemy.

  • Python 3 & FastAPI web backend
  • RESTful API endpoints for order and item management
  • MySQL + SQLAlchemy for persistent storage
  • Redis for fast, stateful chat memory
  • AI agents for conversation and automation flows
  • Natural language processing and text classification (Hebrew intent detection)
  • Smart features: upsell suggestions, dynamic item linking
View on GitHub
Project 2

Full-stack Students Requests System

A full-stack web application built with FastAPI and React, featuring a modern architecture and comprehensive testing suite.

  • RESTful API backend with FastAPI
  • Database integration with SQLAlchemy & Alembic
  • Comprehensive test coverage (Pytest)
  • CI/CD pipeline with Azure DevOps
  • Email and AI service integrations
View on GitHub
Space Invaders Game Screenshot

Space Invaders Game

  • Developed using Pygame library
  • Object-oriented Python design
  • Custom game logic, powerups, graphics, and UI
Try it on GitHub
eLibrary Project

eLibrary Project

A full-stack web-based library management system built with ASP.NET Core and Entity Framework Core. Backend highlights: dynamic waiting list, email notifications when books become available, and seamless PayPal API payment integration. Robust SQLite data storage and modular MVC structure for maintainability.

  • User registration and authentication
  • Book browsing, borrowing, and waiting list
  • Email notifications & checkout summary
  • Real PayPal API payment processing
  • Clean MVC architecture, Entity Framework Core
  • SQLite for lightweight, fast storage
View on GitHub