Skip to content

AI Engineer · 25 years in tech

I work across the AI stack, from GPU kernels to agents.

I worked inside OpenAI, on a project the company acquired. My work spans GPU kernel optimization with Triton, production RAG handling 70M+ documents, and full agentic workflows. I take LLM systems from research to production.

Download CV Get in touch GitHub LinkedIn
Daniel Romero
Remote · Fortaleza, Brazil

Worked at

The stack

Where I work in the stack.

Five layers sit between a question and the hardware that answers it. Here is what I do at each one.

  1. Agents

    Orchestration and tool use

    End-to-end agentic workflows: planning, tool calls, multi-step execution against real systems.

    LangChain · LlamaIndex · DSPy

  2. Retrieval

    Hybrid search at 70M+ documents

    Dense, sparse and late-interaction embeddings combined, with re-ranking well past what keyword search returns.

    Qdrant · HuggingFace · OpenAI

  3. Serving

    Ingestion and query APIs

    Ingestion, embedding generation and query endpoints, packaged and scheduled on infrastructure I've run for two decades.

    FastAPI · Docker · Kubernetes

  4. Model runtime

    Precision and quantization

    Mixed-precision training in FP16 and BF16, plus INT8 quantization. Throughput up, memory footprint down.

    PyTorch · CUDA

  5. Kernels & silicon

    Custom GPU kernels

    High-performance Triton kernels shipped as PyTorch extensions, written after profiling the GPU rather than guessing at it.

    Triton · Nsight · PyTorch Profiler

Experience

Where I've worked.

  1. Stealth AI startup

    Acquired by OpenAI

    AI Engineer
    Remote, USA
    07/2025 – 02/2026

    • Built large-scale LLM workloads with pipelines integrated directly into OpenAI production systems, and kept working on the project inside OpenAI after the acquisition.
    • Wrote high-performance Triton kernels shipped as PyTorch extensions, cutting GPU inference latency at the kernel level.
    • Profiled GPU bottlenecks with Nsight and the PyTorch Profiler, then applied mixed-precision training (FP16/BF16) and INT8 quantization to raise throughput and shrink memory footprint.

    Python · PyTorch · Triton · CUDA · LangChain · LlamaIndex · Qdrant · FastAPI

  2. Doximity

    Senior Software Engineer / AI
    Remote, USA
    07/2024 – 07/2025

    • Designed and shipped a production RAG platform for semantic search across 70M+ medical documents, cutting retrieval time for healthcare professionals.
    • Combined dense, sparse and late-interaction embeddings with re-ranking, measurably beating the keyword-search baseline on retrieval precision.
    • Built the FastAPI backend behind it, covering embedding generation, document ingestion and vector search against Qdrant.

    Python · FastAPI · Qdrant · HuggingFace · OpenAI · LangChain · Docker · Kubernetes

  3. Qdrant

    ML Engineer / AI
    Remote, Germany
    01/2024 – 07/2024

    • Wrote the tutorials and documentation that got developers onto the Qdrant ecosystem faster, covering semantic search, vector search and RAG.
    • Contributed framework integrations for DSPy, LangChain and LlamaIndex, smoothing adoption in production retrieval pipelines.
    • Recorded a tutorial series for the Qdrant YouTube channel, still up as a public playlist.

    Python · LangChain · LlamaIndex · DSPy · Qdrant · OpenAI · HuggingFace

  4. DNSFilter

    Platform Engineer / ML
    Remote, USA
    06/2022 – 12/2023

    • Cut deployment time 87%, from 5 minutes to 40 seconds, by moving Rails services off bare metal onto Kubernetes (EKS) with Terraform and Ansible.
    • Closed the blind spots in the monitoring stack with continuous alerting and observability, improving incident response across production.
    • Contributed to the deployment pipeline the data science team used to ship ML models to production.

    Terraform · Ansible · Docker · Kubernetes (EKS) · Loki · Grafana · Datadog · eBPF · PyTorch

  5. Henku

    Data Scientist / ML
    Remote, Estonia
    04/2021 – 06/2022

    • Cut model training and deployment 90%, from 30 minutes to 3, by replacing ad-hoc notebooks with a continuous training pipeline serving 10K+ active users in Spain.
    • Took a food-nutrition computer vision model from scratch to production: a CNN in PyTorch, in front of those same 10K+ users.
    • Automated the model lifecycle with Airflow, Kubeflow and Weights & Biases, making experiments reproducible.

    Python · PyTorch · TensorFlow · Scikit-Learn · Airflow · Kubeflow · Weights & Biases

  6. Elastic

    Senior Site Reliability Engineer
    Remote, USA
    09/2018 – 04/2021

    • Held 99.95–99.99% uptime across tens of thousands of Elasticsearch clusters on AWS, Azure and GCP, leading on-call in a multi-region, multi-cloud environment.
    • Contributed to a Golang CLI for cluster maintenance that took ~8 hours of manual work per week off the SRE team.
    • Deployed new regional infrastructure to widen Elastic Cloud's global footprint and cut latency for international customers.

    Golang · Python · Ansible · Terraform · Kubernetes · AWS · Azure · GCP · Elasticsearch

Before that

2015 – 2018
DevOps Engineer
AWS, GCP, Docker, Kubernetes, Ansible, Terraform, Prometheus, Grafana
2007 – 2015
Software Developer
PHP, Java, Ruby on Rails, MySQL, PostgreSQL, LXC, Docker
2000 – 2007
Sysadmin
Linux, iptables, proxies, networks, Bash, Perl

Physical AI

Lately I've been building and training my own humanoid robot.

It's the same work I do with language models, pointed at hardware. The difference is that the feedback loop is physics, so nothing can be mocked out.

Writing about it

Work in progress · dev.to/infoslack

Contact

Get in touch.

I'm happy to talk about retrieval at scale, inference performance, or agent systems in production. Email is the fastest way to reach me.

infoslack@gmail.com LinkedIn GitHub X YouTube PT-BR Download CV