anthonymuniz.devOpen to remote roles

Anthony Muniz / AI & Backend Engineer

I make LLM agents work with real customers, and I measure it.

Most AI agents look great in a demo and fall apart with real customers. I'm the engineer who finds out why and fixes it, with numbers to prove the fix worked.

Fig. 1, credit analysis pipeline

Five credit bureaus in, one report out. A credit analysis took about an hour, and this pipeline gets it done in as little as 5 minutes. Moving the heavy work to RabbitMQ and Celery workers took capacity from about 10 to up to 100 operations a day.

5+ credit bureausB1B2B3B4B5RabbitMQ queueworkerworkerworkerreportabout 1 hourdown to 5 min

Selected work

  1. 01

    AI support agent for a B2B SaaS

    It answers customers on WhatsApp, web chat and email. I shipped about 30 releases in 4 weeks, and replies got 3x faster.

    • 3x faster replies
    • ~30 releases in 4 weeks
    • 13 of 23 test questions improved

    2026. Python, LLMs, RAG, n8n, HubSpot

  2. 02

    Credit analysis pipeline

    I rebuilt a credit fund's manual credit analysis as Python services with queues and workers.

    • about 1 hour down to as little as 5 minutes
    • about 10 to up to 100 operations a day
    • 5+ credit bureaus

    2024 to 2026. FastAPI, RabbitMQ, Celery, Redis

  3. 03

    LogStream

    My first Go project. A sidecar agent streams logs over gRPC to a live dashboard, and healthmon blocks new requests to an external API once its error rate crosses a threshold.

    • Go and gRPC
    • fsnotify and a ring buffer
    • later retired for Grafana

    Side project. Go, gRPC, WebSocket, React

  4. 04

    RAG side projects

    A PDF notes app and a resume generator on ChromaDB, plus Kai, my personal copilot, with embeddings and cosine similarity in numpy. The bug I remember: every restart wiped the vector index.

    • 1,000-character chunks, 200 overlap
    • duplicates dropped under 0.3 cosine distance
    • 100+ commits on the resume generator

    2026. FastAPI, ChromaDB, Gemini embeddings, Celery

  5. 05

    Tuberculosis surveillance research

    Scrapers for Brazil's public health database feed a dashboard and a forecasting study. No single model won everywhere.

    • MAE 0.0205, the study's best result
    • LSTM, ARIMA, DES and XGBoost compared
    • 67% of cases in men

    2023 to 2025. Python, Selenium, Pandas, Parquet

Have a system that should work better? Write to me.

anthonympodev@gmail.com

LinkedInGitHub

I'm open to remote backend and AI engineering roles with US and European teams. I also take small freelance projects through my company, Insight AI & Automation.