Available for freelance projects

Python backends and AI, built for production.

I'm Fazliddin. I design the data model, write the service layer, ship it, and keep it alive — lately with LLM agents, document extraction and computer vision for real businesses.

~4 yearsPython backend & DevOps
AI in productionlogistics · agriculture · manufacturing
Namangan, UZGMT+5 · 4 languages

Work with me Read the write-ups

agent/sanitizer.py Telegram

model draft

Calling ~~get_truck(id=12)~~ now. Truck 12 is in Almaty.

sent to the user

Truck 12 is in Almaty.

WARNING sanitizer_modified_reply

PermitLimitReached · 2/2 DINOv2 · MIN_SCORE 0.55 ✓ pre-filled, not saved vmstat · st 0–2%

The stack I ship with

Python Django / DRF FastAPI PostgreSQL Redis Celery Docker GitHub Actions NGINX Gemini · Vertex AI PyTorch DINOv2 Telegram Bot API

Selected work

Applied AI in boring, important places.

Not demos — systems where a wrong answer costs someone money. Client names are always anonymized.

AI back office for a trucking company

~150 trucks on CIS, Turkey and China routes. A Telegram agent reads contracts, licences and passports with an LLM — and never quietly saves a wrong number.

DjangoGemini · Vertex AITelegram
Read the write-up
Guardrails for an LLM agent

The prompt asks, the code enforces: a small Python layer that keeps tool names, tracebacks and four-paragraph answers away from users.

PythonGeminipytest
Read the write-up
Visual product search

Snap a photo, get the matching catalogue item back in Telegram. DINOv2 embeddings and one matrix multiply — on a MacBook, no vector database.

PyTorchDINOv2Telegram
Read the write-up
On-prem vision for a factory

Face ID attendance, scale events and future camera AI for a ~200-person plant — one server, sized GPU-ready rather than GPU-first.

Computer visionHardwarePostgreSQL

Services

How I can help.

Fixed-scope projects or hourly work. Tell me what you need and when — I reply within one business day.

1–4 weeks

Web application

Full-stack web app with Django or Next.js. Database, auth, admin panel, deployment included.

From 1,500,000 so'm

3–7 days

Telegram bot

Custom Telegram bot for your business — orders, notifications, CRM integration, payments.

From 500,000 so'm

1–2 weeks

Automation & API

Automate repetitive tasks, build REST APIs, connect third-party services, set up CI/CD.

From 800,000 so'm

Flexible

Hourly work

Bug fixes, small features, code review, server setup — anything you need, billed by the hour.

80,000 so'm / hour

Writing

Write-ups from production.

#computer-vision 3 min read

Visual Product Search on a Laptop with DINOv2 (No Vector DB Required)

Snap a photo, get the matching catalogue item. ~700 images, one MacBook, one matrix multiply.

#llm 3 min read

The Prompt Asks, the Code Enforces: Guardrails for an LLM Agent in Production

Verbose replies, leaked tool names, and raw exceptions in front of users — and the small Python layer that fixed them.

#django 4 min read

Never Let the Model Save: LLM Document Extraction in a Trucking Back Office

Two rules that made AI document extraction safe enough for a business where a wrong number becomes a wrong tax payment.

Principles

Lessons I keep relearning.

“The prompt asks. The code enforces. Don't ship an agent with only one of them.”

The Prompt Asks, the Code Enforces Guardrails for an LLM agent in production

“LLMs are great readers and terrible accountants. Let them read. Show a human what they read.”

Never Let the Model Save LLM document extraction in a trucking back office

“A pretrained model, a normalized matrix, and a threshold will get you surprisingly far.”

Visual Product Search on a Laptop with DINOv2 No vector database required

permits/services.py

@transaction.atomic
def issue_permit(*, driver_id: int, permit_type: str, issued_by) -> Permit:
    # Lock the driver row so two staff members can't race past the limit.
    driver = Driver.objects.select_for_update().get(pk=driver_id)

    open_permits = Permit.objects.filter(driver=driver, returned_at__isnull=True).count()
    if open_permits >= driver.permit_limit:        raise PermitLimitReached(
            f"{driver} already holds {open_permits}/{driver.permit_limit} permits. "
            "Return one before issuing another."
        )

    return Permit.objects.create(
        driver=driver,
        permit_type=permit_type,
        issued_by=issued_by,
        issued_at=timezone.now(),
    )

Let's talk

Have a system that has to be right? Let's build it.

Tell me what you're building and where it hurts. Rules that cost money go in one tested place — never in a prompt.

Start a project About me