Machine Learning Engineer

A camera. A CNN small enough to run on a wall-mounted board. One question: what is this person about to do? That was my thesis. No dataset existed, so I made one. A good model is only half of the product. The other half is software that keeps running, so I learned that half as a Java developer. In 2026 I put the two together: my thesis model is live on Hugging Face at 98.84% test accuracy, its preprocessing is a package on PyPI. AI is what I care about, and now I can also ship it.

Download CV
Portrait of Sahar Darafsh

2017 - 2022

A good model is only half of the product

  1. Sep 2017 - Jun 2020

    MSc, Artificial Intelligence and Robotics · Amirkabir University of Technology, Tehran

    Thesis: "Designing an Embedded System for Activity Recognition Using a Surveillance Camera"

  2. 2020

    Real-Time Activity & Intention Recognition

    Video-to-decision pipeline that compresses motion clips into one image for a lightweight CNN, designed for an embedded FPGA.

  3. 2021

    First-author arXiv paper on a dataset I built (98.78%).

  4. Jun 2020 - Mar 2022

    AI/ML Researcher and Project Reviewer · AUMS, Abadan, Iran (Remote)

    Reviewed deep-learning projects in medical image analysis.

    Evaluated each project’s model architecture, training and metrics.

  5. Jan 2021 - Mar 2022

    Machine Learning Engineer · danaXa, Canberra, Australia (Remote)

    Built a resource-efficient computer vision pipeline with YOLO on a Jetson Nano.

    Delivered low-latency, on-device video analytics for occupancy monitoring.

  6. Jun 2021 - Mar 2022

    Deep Learning Mentor and Teaching Assistant · Maktabkhooneh, Tehran, Iran (Remote)

    Taught and mentored 100+ students in deep learning, CNNs and reinforcement learning.

    Co-developed the curriculum with industry examples.

2022 - 2026

The other half is software that keeps running

  1. 2022 - 2026

    so I spent 4.5 years as a Java engineer in Munich learning that side properly: TDD, CI/CD, code quality gates, Docker.

    • Mar 2022 - Mar 2025Software Developer · TESTIFI GmbH, Munich, Germany (Remote): Turned stakeholder requirements into core features of QualityOps, a Java tool for evaluating software quality, using TDD.
    • Mar 2025 - Oct 2026Full Stack Developer · Convien GmbH, Munich, Germany: Used Claude Code daily on production code and built two agents to write and review unit tests, raising coverage from 0% to 95–98% with a final team review.
    The software side of this story

2026

In 2026 I put the two together

  1. Nov 2026

    Travel Memory MCP Server: RAG over Qdrant

    Python MCP server that lets Claude search my travel notes and cite the source of each answer.

    RAG pipeline over Qdrant: one chunk per place, metadata filters, re-runnable without duplicates.

    Hit rate@5 of 94% on 100 evaluation questions.

    Clean architecture enforced by import-linter; pytest, Ruff and GitHub Actions CI.

  2. Oct 2026

    LLM Listing Assistant

    Full stack LLM app: a self-hosted LLM (Qwen 2.5 on Ollama, OpenAI-compatible API) extracts sellers’ chat messages into structured JSON edits.

    Validates every output with regex VIN extraction, an allowed-field schema and human approval of each edit before a listing changes.

    NestJS/TypeScript, PostgreSQL, Next.js/React; Docker, GitHub Actions, Kubernetes.

  3. Oct 2026

    Daily Journal Sentiment Analysis (private GitLab repos)

    Classifies English and Persian journal entries as good, neutral or bad days, with a reason.

    Evaluated a zero-shot model, a sentiment model and a local LLM on my own labels for accuracy, latency and GPU use.

    Served over gRPC, tracked in MLflow, released automatically.

  4. Sep 2026

    adaptive-motion-preprocessing: open-source Python package

    Turns video frames into motion images for neural networks with pluggable preprocessing stages.

    Released on PyPI, with CI tests on Linux, Windows and macOS.

  5. rebuilt 2026

    Real-Time Activity & Intention Recognition: MSc thesis to open-source system

    Rebuilt in 2026 as a tested Python CLI; the Hugging Face model reaches 98.84% accuracy (F1 0.99) on held-out data.

Background

Skills

ML / Deep Learning
Keras · TensorFlow · PyTorch · CNN training · fine-tuning · evaluation (accuracy, F1, confusion matrix) · Hugging Face Hub
Computer Vision
OpenCV · video/motion-based feature encoding · image classification · activity recognition · background subtraction · embedded CV design (Xilinx ZCU102, thesis)
NLP & LLMs
Hugging Face Transformers · XLM-RoBERTa · zero-shot classification (NLI) · LLM APIs (OpenAI-compatible API, Ollama, Qwen 2.5) · JSON output from LLMs · prompt engineering · sentiment analysis · multilingual text (English, Persian)
MCP & RAG
Model Context Protocol (MCP) · RAG · embeddings (all-MiniLM-L6-v2) · vector search (Chroma) · chunking · source citations
Python & Tooling
Python · pandas · Git · Claude Code · pytest · ruff · CLI design · clean architecture · pydantic-settings · import-linter
Backend, MLOps & DevOps
Java · TypeScript · NestJS · REST APIs · gRPC · Protobuf (Buf) · MLflow · Spring Framework · Next.js · React · Docker · Kubernetes · CUDA · AWS · CI/CD (Jenkins, GitHub Actions, GitLab CI) · SonarQube · SQL (MySQL, PostgreSQL, SQLite, pgvector) · TDD/JUnit

Education

  • Sep 2017 - Jun 2020

    MSc, Artificial Intelligence and Robotics · Amirkabir University of Technology, Tehran

    Thesis: "Designing an Embedded System for Activity Recognition Using a Surveillance Camera" (see Research & Projects for the resulting repo, paper, and model)

  • Sep 2011 - Sep 2016

    BSc, Computer Engineering · Isfahan University of Technology, Isfahan

Certificates

  • AI for Medicine Specialization, DeepLearning.AI

  • Deep Learning Specialization, DeepLearning.AI

  • Reinforcement Learning Specialization, University of Alberta

  • Generative Adversarial Networks (GAN) Specialization, DeepLearning.AI

Languages

English (Fluent) · German (B1) · Persian (Native)

Let's build something.

The fastest way to reach me is email. I'm always happy to talk about interesting problems, collaborations, or roles.