Onur Barut, PhD

I build production AI systems that run unattended at scale and can prove their own quality. Today I am the Lead Machine Learning Engineer at Wizard Commerce, where I own the AI catalog-enrichment platform: the categories, typed attributes, summaries and embeddings behind a search index of tens of millions of listings assembled from 100+ retailer feeds. Most of my work sits where large language models meet retrieval and data quality — structured-output extraction, hybrid search, embedding pipelines, orchestration on Airflow and Cloud Run, and evaluation harnesses that decide which model actually ships. Before Wizard I was an AI Software Architect at Intel, working on multimodal RAG, VLM-generated synthetic data and hardware-aware inference optimization, and I earned my PhD in Computer Engineering at UMass Lowell for transformer-based network security models. I also developed a graduate generative AI course for Clark University. I am open to new opportunities — Staff or Lead roles in applied AI, GenAI, and search and recommendations.

Platform, model and data-quality work behind a catalog search index of tens of millions of listings. Company-internal names, customers and systems are deliberately omitted.

• AI catalog enrichment platform

  51 registered job types

  Owner and primary author of two successive generations of the service that turns raw retailer feeds into search-ready categories, typed attributes, summaries and embeddings. Took it from research-grade code with manual file handoffs to a job framework that runs whole-catalog campaigns unattended and is ingested continuously by the data platform.

  • Python 3.12
  • asyncio
  • Pydantic
  • Cloud Run
  • Parquet

• Recategorization at catalog scale

  34.0% → 44.9% on a 256-product audit set

  Owned LLM product categorization into a ~6,100-leaf taxonomy across three engine generations, ending hallucinated category paths by generating strict Literal-typed response schemas over retrieved candidates. Replayed an operations error set into a repeatable evaluation loop: on a separate 138-product operations error set, exact-category fixes rose 20 → 39 and right-department fixes 33 → 78 with zero regressions. A controlled A/B on 1,199 random listings with retrieval held fixed lifted agreement with production labels 90.2% → 91.0% (19 wins, 9 losses).

  • Structured outputs
  • Hybrid retrieval
  • Elasticsearch
  • vLLM
  • OpenAI Batch API

• Retailer feed mapping service

  87–99% per-attribute accuracy (median ~93%), 11 garment-size attributes

  Built and owned a Cloud Run service that onboards retailer feeds onto a ~6,750-category taxonomy and a ~1,450-attribute schema: hybrid Elasticsearch retrieval (BM25 + kNN + RRF) proposes candidates and an LLM selects under a dynamically generated Literal schema, so it cannot invent a label. Also cross-mapped a next-generation attribute schema onto the legacy repository, covering roughly 1,280 of ~1,450 attributes (about 88%).

  • Elasticsearch
  • BM25 + kNN + RRF
  • GraphQL
  • Cloud Run
  • Docker

• Orchestration and release ownership

  2× batches in flight (50 → 100)

  Took over the enrichment DAG family in the data-engineering team's Airflow repository across roughly 100 pull requests: retailer and sample-size filters, and an idempotent Cloud Run operator that persists the operation name and re-attaches to an in-flight job on retry. Diagnosed that DAG limits and pools ignore deferred tasks — the limits looked correct but bounded nothing — and shipped per-backend pools.

  • Apache Airflow
  • Cloud Composer
  • Cloud Run
  • PySpark
  • pytest

• Open-weight models, chosen on evidence

  Self-hosted 35B edged GPT-4.1-nano, 0.714 vs 0.698

  Led the move of catalog enrichment from per-token vendor APIs to self-hosted open-weight models (Qwen3 / 3.5 / VL, GPT-OSS, Gemma on vLLM behind a LiteLLM gateway), landing a 68-file, ~20.7k-line integration, keeping the vendor batch path as a pluggable fallback, and holding production up through a compute-capacity interruption by re-pointing a live job to an external GPU-cloud model on a two-line config change. Models were picked on benchmark evidence — a config-driven LLM-as-judge harness (severity rubrics, strict-JSON verdicts, coverage, agreement, defect-rate and cost metrics, text and vision judging) alongside reference-free heuristic rubrics, which showed an 8B open-weight model matching an 80B one on 754-variant name normalization and quantified an open-weight VLM hallucinating numeric specs ~3× more often (4.4% vs 1.4%).

  • vLLM
  • LiteLLM
  • Qwen3 / Qwen3-VL
  • GPT-OSS
  • LLM-as-judge

• Catalog data-quality program

  Applied size coverage 11% → 79% on one feed cohort

  Turned anecdotal quality complaints into measured programs: read-only Elasticsearch audits quantified ~4.7M documents that carried a valid size prefix but no size fields, and a normalized size attribute out of enum on ~31% of ~11.2M documents. The fix was deterministic rather than another LLM pass — a fill-only, evidence-ledgered, dry-run-by-default pipeline routed through the standard ingestion path. Separately built a conformance validator that checked 26 enrichment configs (25 per-vertical plus a fallback) against the schema and taxonomy and enumerated 1,044 conformance issues across eight classes — 365 of them low-severity name-similarity pairs — against staging snapshots, before any of them reached production.

  • Elasticsearch aggregations
  • Delta Lake
  • Dry-run/ledger design
  • Airflow
  • pandas

• Near-duplicate image detection

  Rerank-latency benchmark

  A research benchmark for near-duplicate listing detection in a recommendation rerank step under a hard latency budget: a classical hash / HSV / ORB + RANSAC cascade against triplet-loss fine-tuned CNN backbones and a multimodal embedding model with Matryoshka truncation, all through one harness that sweeps thresholds for precision/recall and measures warm and cold latency against explicit PASS/FAIL budgets.

  • PyTorch
  • OpenCV
  • Triplet loss
  • Qwen3-VL-Embedding
  • Streamlit

• Conversational shopping features

  Catalog-grounded

  Worked on the shopping assistant's prompt layer — soft-constraint entity extraction, brand and entity extraction merged into one async call, and hard-coded size lists replaced with values fetched from the catalog API — then an LLM-judged pipeline that decides which attributes a category genuinely requires, so the assistant asks only the clarifying questions that category actually needs. Later prototyped a retrieval-grounded, steerable shopping experience over a hybrid BM25 + kNN pipeline.

  • Prompt engineering
  • FastAPI
  • SSE
  • Elasticsearch RRF
  • MongoDB

• Multimodal RAG and video super-resolution

  Open-sourced at Intel

  As AI Software Architect at Intel: trained and open-sourced a multimodal visual RAG model on a LLaMA backbone, generated VLM-based synthetic data to fine-tune a smaller language model for document question answering, and open-sourced a video super-resolution model reported at +2% PSNR with reduced artifacts at minimal additional compute.

  • LLaMA
  • Multimodal RAG
  • PyTorch
  • OpenVINO
  • ONNX

1. Wizard Commerce — Lead Machine Learning Engineer

   Oct 2024 – Present · Remote · MA

   AI shopping company. Owner of the AI catalog-enrichment platform: categories, typed attributes, summaries and embeddings for a search index of tens of millions of listings built from 100+ retailer feeds.

   • Owned the AI catalog-enrichment platform across two service generations and three repositories, taking it from a research-grade service with manual file handoffs to a job framework of 51 registered jobs that runs whole-catalog campaigns unattended and is ingested continuously by the data platform.
   • Led the migration of catalog enrichment from per-token vendor APIs to self-hosted open-weight models (Qwen3 / 3.5 / VL, GPT-OSS, Gemma on vLLM behind a LiteLLM gateway), selecting each model on benchmark evidence — a self-hosted 35B model edged GPT-4.1-nano on a 381-variant summary benchmark (0.714 vs 0.698) — keeping the vendor batch API as a pluggable fallback, and holding production up through a compute-capacity interruption by re-pointing a live job to an external GPU-cloud model on a two-line config change.
   • Selected models per task with evaluation-first benchmarks: 754-variant name normalization (98.9% accuracy for the best vendor-API model, with an 8B open-weight model matching the 80B), 1,077-variant vision-language recategorization (92.8% full-path exact match for the leading vendor model; the open-weight 8B led subcategory accuracy), and a 381-variant summary benchmark.
   • Owned LLM product categorization into a ~6,100-leaf taxonomy across three engine generations, eliminating hallucinated category paths by generating strict Literal-typed response schemas over retrieved candidates.
   • Ran catalog-scale campaigns: ~1.7M listings in a taxonomy migration, a multi-million-listing recategorization campaign, 9M+ records lite-enriched, ~3.5M listings size-enriched across 100+ retailer feeds, and a ~275k-product brand pass across 100+ retailer feeds.
   • Migrated catalog embeddings from a paid vendor API to a self-hosted fine-tuned Qwen3 model, backfilled 10 product categories, and fixed a silent data-loss bug so the full set of per-facet vectors reaches the search index.
   • Took over the enrichment Airflow DAG family in the data-engineering team's repository (roughly 100 pull requests), diagnosing that DAG limits and pools ignore deferred tasks and shipping per-backend pools that doubled batches in flight (max active runs 50 → 100), plus an idempotent Cloud Run operator that re-attaches to an in-flight job on retry.
   • Shipped a deterministic, LLM-free size and color fix pipeline — fill-only, evidence-ledgered, dry-run by default, routed through the standard ingestion path — that lifted applied size coverage 11% → 79% on one feed cohort — 100% LLM-judged accuracy on the variant-evidenced letter/numeric/waist tiers (313/313) and 98% on the text-derived default tier — validated on a staging cohort ahead of production rollout.

2. Clark University — Adjunct Faculty (part-time)

   Fall 2025 · Clark University

   • Developed MSIT 3103: Generative AI Model Development, a graduate course on building with generative AI models, for Fall 2025.
   • Published the course assignment notebooks as a public repository so students — and anyone else — can run the exercises: MSIT3103-GenAI-assignments.

3. Intel Corporation — AI Software Architect

   Jun 2022 – Oct 2024 · Remote / MA

   • Generated synthetic data with a vision-language model to fine-tune a smaller language model for document question answering.
   • Trained and open-sourced a multi-modal visual RAG model on a LLaMA backbone for retail security.
   • Designed and open-sourced deep learning models for network-AI projects and optimized them for Intel hardware against customer compute budgets, reported at up to 50% higher throughput.
   • Developed and open-sourced a video super-resolution model reported at +2% PSNR with reduced artifacts at minimal additional compute cost.
   • Led research on network security and video quality enhancement, publishing and presenting more than five papers internally and externally.

4. University of Massachusetts Lowell — Research Assistant, ACANETS Lab

   Aug 2018 – May 2022 · MA

   • Raised malicious traffic detection above 90% and improved model expansion capability with a transformer-based deep learning architecture.
   • Co-organized the NetML 2020 network traffic analytics competition, and collected and open-sourced the HAR-UML20 human activity recognition dataset.
   • Developed a multi-task LSTM model for activity classification and intensity estimation with over 95% accuracy.

5. Intel Corporation — Graduate Intern for Machine Learning

   May – Aug 2019 · CA

   • Built end-to-end machine learning pipelines for network traffic analysis and improved throughput by 15% by optimizing the flow feature selection algorithm.

6. Aselsan — System Design Engineer

   Feb 2014 – Jul 2018 · Turkey

   • Led the team responsible for flight-test data acquisition tasks and system integration tests.
   • Increased flight-test data acquisition speed 5× by replacing the PCM interface with Ethernet.
   • Contributed to Electronic Warfare Self-Protection (EWSP) integration projects with flight data acquisition.

Ordered by depth — the first items in each group are where I have designed and owned production systems.

★A star marks the three most-cited papers; citation counts shown only for those.

1. RET-Net: A CNN Framework for Real-Time Traffic Classification Using Key-Byte Mechanism

   IEEE Transactions on Network and Service Management, 2026.

2. Malware Detection for Portable Executables Using a Multi-input Transformer-Based Approach

   IEEE ICNC, 2024. DOI — Malware Detection for Portable Executables Using a Multi-input Transformer-Based Approach

3. AWS Compute Video Super-Resolution powered by the Intel® Library for Video Super Resolution

   Mile-High Video Conference (MHV '24, ACM), 2024. DOI — AWS Compute Video Super-Resolution powered by the Intel® Library for Video Super Resolution

4. Universal Network Traffic Analysis for Malicious Traffic Detection using RappNet: A Privacy-Preserving Approach

   IEEE CNS 2023 (poster). DOI — Universal Network Traffic Analysis for Malicious Traffic Detection using RappNet: A Privacy-Preserving Approach

5. A Comprehensive Study on Efficient and Accurate Machine Learning-Based Malicious PE Detection

   IEEE CCNC, 2023. DOI — A Comprehensive Study on Efficient and Accurate Machine Learning-Based Malicious PE Detection

6. Weakly supervised learning for network traffic classification

   IEEE NAS, 2022. DOI — Weakly supervised learning for network traffic classification

7. ★Most cited. R1DIT: Privacy-Preserving Malware Traffic Classification with Attention-Based Neural Networks

   IEEE Trans. Network and Service Management, 2022 (early access). · 69 citations. DOI — ★Most cited. R1DIT: Privacy-Preserving Malware Traffic Classification with Attention-Based Neural Networks

8. Multi-Task Hierarchical Learning Based Network Traffic Analytics

   IEEE ICC, 2021. DOI — Multi-Task Hierarchical Learning Based Network Traffic Analytics

9. Machine learning based malware detection on encrypted traffic: A comprehensive performance study

   NSysS 2020 (ACM), 2020. DOI — Machine learning based malware detection on encrypted traffic: A comprehensive performance study

10. TLS Encrypted Application Classification Using Machine Learning with Flow Feature Engineering

    ICCNS 2020 (ACM), 2020. DOI — TLS Encrypted Application Classification Using Machine Learning with Flow Feature Engineering

11. ★Most cited. Multitask LSTM Model for Human Activity Recognition and Intensity Estimation Using Wearable Sensor Data

    IEEE Internet of Things Journal, 2020. · 95 citations. DOI — ★Most cited. Multitask LSTM Model for Human Activity Recognition and Intensity Estimation Using Wearable Sensor Data

12. ★Most cited. NetML: A Challenge for Network Traffic Analytics

    2020. · 50 citations. arXiv

13. Geospatial object detection using deep networks

    SPIE Proceedings, 2019. DOI — Geospatial object detection using deep networks

Full list and citation counts on Google Scholar.

Datasets, challenge baselines, paper code and course material.

• NetML Competition 2020

  Network traffic analytics challenge I co-organized: open dataset, baselines and evaluation for malware and traffic classification.

  26 stars · co-organizer

• HAR-UML20

  Human activity recognition dataset I collected and open-sourced with the ACANETS Lab.

  2 stars · dataset

• Encrypted Malware Detection

  Code for the comprehensive performance study of machine learning malware detection on encrypted traffic.

  6 stars · paper code

• MultiTask-LSTM-HAR

  Code for the IEEE Internet of Things Journal multi-task LSTM activity recognition paper.

  3 stars · paper code

• Multi-Task Hierarchical Learning

  Code for the multi-task hierarchical network traffic analytics paper.

  2 stars · paper code

• Encrypted App Classification

  Code for the TLS encrypted application classification paper with flow feature engineering.

  paper code

• MSIT3103 GenAI Assignments

  Assignment notebooks for the graduate generative AI course I developed at Clark University.

  course material

I was part-time faculty at Clark University, where I developed MSIT 3103: Generative AI Model Development, a graduate course for Fall 2025. The assignment notebooks are public, so the exercises can be run by anyone: MSIT3103-GenAI-assignments. Earlier, as a graduate student at UMass Lowell, I was a Teaching Assistant for the Logic Design course in Fall 2018, coordinating lab work and running tutoring sessions.

I am open to new opportunities and always happy to talk about production LLM systems, search and retrieval, or evaluation — reach me at any of these.

• onurbarut@windowslive.com
• LinkedIn
• GitHub
• Google Scholar