agfian's space

AI Engineer · Indonesia · UTC+7

I ship computer vision, MLOps pipelines, LLM systems, AI evaluations, backend services, DevOps pipelines, automations and reliable software to production.

Hi, I'm Agfian. For 6+ years I've taken AI from a promising notebook to systems people rely on every day: vision models on edge devices in palm-oil mills, the MLOps and backend that keep them honest, and now LLM workflows that take hours of busywork off a team's plate.

See how I ship Work with me
years shipping AI to production
6+
more experiments every week
10×
off the monthly cloud bill
57%

An animated portrait of Agfian drawn with ASCII characters. Particles scatter and rebuild the picture, then computer-vision style bounding boxes label the face and smile. The bars below it switch between the portrait and the lightning-bolt logo.

Scroll

02 / How I ship

From raw pixels to a system that runs itself.

Most AI projects stall between the demo and production. I've owned every stage in between, so nothing gets lost in a handover. Scroll through the pipeline.

An ASCII picture that redraws itself for each stage of the pipeline described alongside it: a labelled dataset, training curves, an edge device, a server rack, a drift chart and an LLM workflow, each with computer-vision style bounding boxes.

  1. Stage 01 / Data

    Great models start as great datasets.

    I build the labelling and curation loop first: self-hosted CVAT with AI-assisted annotation, then FiftyOne to catch duplicates, gaps and mislabels before they ever reach training.

    weeks per annotation cycle, down from 3
    1
    better data quality after curation
    +25%
    • CVAT
    • FiftyOne
    • MinIO
    • OpenCV

  2. Stage 02 / Train & evaluate

    Experiments you can repeat. Numbers you can trust.

    Reusable training templates, every run tracked and versioned in ClearML, and evaluation that goes past a single mAP: per-class errors, hard cases, and a held-out set that looks like the field.

    experiments a week, up from 1
    ~10
    to train a model that used to take 2 weeks
    1–2 days
    • PyTorch
    • Ultralytics YOLO
    • ClearML
    • ONNX

  3. Stage 03 / Ship to the edge

    Fast enough for the factory floor.

    Models compiled with TensorRT and served with Triton, running on Jetson devices right next to the cameras, where the network is flaky and every millisecond counts.

    to grade a fruit bunch, down from ~2 min
    ~30 s
    grading accuracy in a palm-oil mill
    ~87%
    • TensorRT
    • Triton
    • NVIDIA Jetson
    • Docker

  4. Stage 04 / Serve

    APIs and infrastructure that stay up.

    FastAPI services with PostgreSQL and Redis behind them, packaged in Docker, running on Kubernetes and described in Terraform, so a new environment is a pull request, not a week of clicking.

    from merge to production
    <5 min
    saved each month moving from AWS to GCP/Azure
    57%
    • FastAPI
    • PostgreSQL
    • Redis
    • Kubernetes
    • Terraform
    • CI/CD

  5. Stage 05 / Monitor

    Know a model is drifting before users do.

    RED metrics for the service, distribution checks for the data, and alerts that land in chat with enough context to act on, so a silent accuracy drop becomes a ticket instead of a complaint.

    faster error recovery
    35%
    app startup, down from 10 minutes
    ~5 s
    • Prometheus
    • Grafana
    • Loki
    • Alerting

  6. Stage 06 / Automate with LLMs

    LLM systems with evals and guardrails.

    Document pipelines, agent workflows and automations that do real work, each shipped with an eval set, so swapping a prompt or a model is a measured change rather than a hopeful one.

    to redact a 300-page document that took a week by hand
    ~60 s
    • OpenAI API
    • Claude Code
    • AI evals
    • n8n
    • RabbitMQ

03 / Field notes

I write down what I learn.

All posts

    04 / Contact

    Got a model that needs to leave the notebook?

    I like the hard middle part: turning a promising model into a system that's fast, observable and cheap to run. If you're building with computer vision, MLOps or LLMs, I'd love to hear about it.

    Say hello LinkedIn GitHub