Alisher Sharipov

AI that runs on your own hardware

Alisher Sharipov, AI and backend engineer (Python, Go) in Astana. I run a 35B coding model on an ordinary consumer PC, and I built an ERP payroll module that matches 1C to the cent.

Download CV (PDF) GitHub Резюме на русском
  • Python
  • Go
  • FastAPI
  • SQL
  • llama.cpp
  • RAG
  • MCP
  • Docker
  • 1C / OData

Open to remote work.

shura local, offline

› Why is the payroll total one cent off from 1C?

payroll.py rounds every line before summing,
so the rounding errors add up.

- total = sum(round(x, 2) for x in lines)
+ total = round(sum(lines), 2)

Tests: 48 of 48 passed.
Tiel-Coder-35B · RX 7800 XT 16 GB 78 tok/s
Context

What happens when you press Enter

  1. ShuraCodeThe agent sends your question and the files it needs, to your own machine.
  2. GatewayWakes the model if it sleeps and restores the saved session in 0.14 s.
  3. llama-serverReads a 100k-token prompt at about 1,000 tokens per second.
  4. GPU + RAMHot experts on the GPU, cold ones in RAM. The answer streams at about 78 tok/s.

0 answer tokens, zero bytes to the cloud

Projects

Shura

Local AI coding agent

github.com/hikkian/shura
How Shura works: the coding agent talks to the gateway, the gateway runs llama-server, the model is split between GPU and RAM

Problem

Strong coding models do not fit in 16 GB of VRAM, and when they fit, the desktop starts to stutter.

What I built

An OpenAI-compatible gateway in Python that loads the model on demand, survives crashes and saves long sessions. My own llama.cpp fork in C++ with an expert cache and a VRAM governor. A harness of 111 agentic coding tasks.

Result

78 tok/s on short context with MTP, up to 200k context, a 100k-token prompt read at about 1,000 tok/s. A 32k-token session restores in 0.14 s instead of 42 s. 580+ tests, CI on Linux, macOS and Windows.

Python · C++ · llama.cpp · ROCm / HIP · MCP · Docker

Barid

Task board for several AI agents

github.com/hikkian/barid
Barid panel: three agents with their current tasks, a proposed task waiting for approval and the order in which tasks can run

Problem

When several agents (Codex, Claude Code, a local model) work in parallel, they pick the same task and fight over the GPU.

What I built

A queue with dependencies and leases on shared resources, human review gates, an unattended worker with timeboxes, and a web panel that works out what can run together.

Result

I run my own multi-agent work on it every day. Zero dependencies, MIT licence.

Python · HTTP API · JavaScript · MCP

Open Go Academy

A school for Go

github.com/hikkian/open-go-academy
Open Go Academy lesson: theory on the left, a Go editor with Run and Check buttons on the right
Project workspace with tasks and an editor
Skills map with progress by topic

Problem

Learning Go from articles is slow: there is no place to run code right away and get it checked.

What I built

A Go backend compiles and runs student code in isolated Docker containers and grades it with tests. React frontend, an optional AI tutor on a local model or in the cloud.

Result

128 chapters, five projects and exams, a skills map with progress.

Go · React · TypeScript · Docker · PostgreSQL

Experience

ABS LAB

Astana · since February 2026

AI / backend engineer

  • HR and payroll module of an ERPfor a road-construction group, live since September. Python, Frappe, SQL.
  • Connector to 1C:ZUPsyncs seven 1C databases over OData, payroll matches to the cent.
  • RAG assistantin the company website chat, in production. Qdrant, Gemini API.
  • Leads into CRMfrom forms and the AI chat go to Bitrix24 and email automatically.
  • Multilingual sitewith Microsoft 365 sign-in. React, Strapi, PostgreSQL.

Skills

Backend
Python, Go, FastAPI, REST, SQL
AI / LLM
RAG, agents, MCP, llama.cpp, local models
Integrations
1C, Bitrix24, OData
Infrastructure
Docker, Nginx, CI/CD, Linux
Frontend
React, TypeScript

Education

Astana IT University
BSc Computer Science, 2023–2026
Eurasian Olympiad in Informatics
bronze
National Olympiad in Informatics
bronze

Get in touch

Open to remote work: AI, LLM systems and backend in Python or Go.

alishar2005@gmail.com