abdulaziz abdujalilov RUTashkent
Oct '26

Abdulaziz Abdujalilov, AI engineer and full‑stack developer in TashkentI build software, then I hope prove it works.

Hi, I'm Abdulaziz (people call me Balu). I build Telegram bots, websites, web apps and AI features for clients anywhere, and this is my working notebook: the projects, the numbers and a few pages I'm not allowed to show you.

since 2022building software for clients
20+projects shipped, public and under NDA
1–2 weeksto a working bot or site
EN · RU · UZcalls and docs in your language

↑ the short version

Need it built?
→ email me
01

01 · what I build, and how long it takes

Typical timelines. After a 20-minute call you get a fixed scope and a date.

Telegram bots and Mini Apps

Bots that sell, book, remind and answer, plus Mini Apps that work like a small website inside Telegram.

1–2 weeks

3–4 weeks with payments and an admin panel

details →

Websites

Fast marketing sites and landing pages that load in a second and send every enquiry to your Telegram or inbox.

1–2 weeks

3–4 weeks for multilingual sites with a CMS

details →

Web apps and MVPs

SaaS MVPs, dashboards, internal tools and calculators: real accounts, real data, deployed and monitored.

4–8 weeks

first working version in 2–3 weeks

details →

AI features and automation

Knowledge bots over your documents, document checks, agents with human approval, and evals that stop bad releases.

2–4 weeks

plus an evaluation set so quality is measured, not guessed

details →

also good at: verification in niche domains (COBOL migration, quant models, optimization), compliance tooling and local AI for teams that cannot use the cloud.

remote · project or contract · UTC+5: full overlap with Europe, US East Coast mornings · EN / RU / UZ · I hand over code, docs, a runbook and tests

02

02 · things that are provably correct

cobol-equiv0 diffs / 9,769

Differential testing harness and a Python runtime for COBOL numeric semantics. GnuCOBOL is the oracle; unsupported cases are refused, never guessed.

  • 9,769 generated cases, 0 diffs vs GnuCOBOL 3.1.2
  • LLM translations: 92% one-shot, 100% with repair rounds
  • mutation tested, 3.1.2 vs 3.2 compared

pricing-crosscheck8 findings!

Cross-validation of derivative pricing engines: QuantLib's analytic, tree, PDE and Monte Carlo engines against independent NumPy references.

  • 27k+ grid points across 4 instrument types
  • 8 documented engine findings
  • protocol fixed before any result

quasar-solvergap 86.7% → 3.7%

Reproducible benchmark of QUBO simulated annealing vs classical heuristics on TSP and tiny CVRP, measured against proven optima.

  • Held-Karp proven optima as reference
  • 95% confidence intervals
  • swap-SA gap 86.7% to 3.7% at 20 cities

Telegram Web A Lab7 PRs

Six new features built inside the open-source Telegram web client, each in its own branch and upstream pull request.

  • Remind Me, Share as Card, Chat Insights
  • Storage Usage, Catch Up, Chat Calendar
  • 7 PRs opened to Ajaxy/telegram-tt
03

03 · more things I built

04

04 · pages I can't show you

Most of my paid work is under NDA. Here is what kind of systems they are, without names, screenshots or internals. Details on a call.

client:

NDA

Interactive labor-law toolset

HR-tech platform · Uzbekistan. A suite of tools embedded into a large HR reference platform: leave and dismissal calculators, interactive checklists in Russian and Uzbek, decision flowcharts, a typical-procedures guide and an HR audit with a violation registry.

Next.js + static + iframe + Vercel + Lovable

client:

NDA

Smart document constructor

HR-tech platform · Uzbekistan. Fill-in-once document builder for HR paperwork: conditional fields, a live A4 preview with two-way highlighting, date logic from the labor code, print, PDF and DOCX export.

TypeScript + DOCX + Vercel

client:

NDA

AI document-check bots

Legal and consulting firm · Uzbekistan. Three Telegram services that read uploaded documents (company charters and contracts, construction documentation, HR documents) and return a structured report of what violates the law and how to fix it. Russian and Uzbek, scans included.

Python + LLM + OCR + Supabase + Railway

client:

NDA

Due-diligence Mini App

Legal and consulting firm · Uzbekistan. Telegram Mini App that works as a landing page for a counterparty-check service ahead of Telegram Ads: tariffs, sample report, lead capture straight into a spreadsheet.

Telegram Mini App + Google Sheets

client:

NDA

Payment-control rulebook

Construction · Uzbekistan. Rule engine design for checking construction acts against estimates before payment: a regulatory map, 100 deterministic machine checks, data model and test corpus. No AI in the core, full audit trail.

rules engine + XLSX import + audit trail

client:

NDA

Lead and webinar bots

Education and events · Uzbekistan. Telegram bots for course and webinar funnels: bilingual flows, registrations to Google Sheets, scheduled reminders and segmented broadcasts.

Python + aiogram + Google Sheets

client:

NDA

Occupation classifier search

Public data · Uzbekistan. Smart search over the national occupational classifier: 10,910 occupations parsed from an official PDF into a fast, typo-tolerant index.

PDF parsing + search + web

client:

NDA

AI video production

Media · Uzbekistan. Short vertical quiz videos with a recurring mascot, produced scene by scene from structured prompts, in Russian and Uzbek.

Veo 3 + JSON prompts + 9:16

client:

NDA

Websites for small businesses

Local businesses · Uzbekistan. Fast marketing sites for two travel agencies, a 3D-printing studio, a wildlife reserve and a private tutor, with forms that land in Telegram.

Next.js + Tailwind + Framer Motion

client:

NDA

Voice English tutor bot

EdTech · Central Asia. Telegram bot for spoken English practice: voice in, voice out, aimed at learners in Central Asia.

grammY + speech-to-text + TTS

client:

NDA

Internal ops tooling

Corporate · Uzbekistan. A task-management Telegram bot for a company team and an on-prem file server layout with role-based access for a law firm.

Telegram + access control

05

05 · code I sent to other people's repos

06

06 · how a project goes

  1. discover

    One call, then a short written brief: who uses it, what done looks like, what can break.

  2. spec

    A protocol before code: data, rules, acceptance criteria and test cases agreed up front.

  3. build

    I direct AI coding agents and review every diff. Fast iterations, small PRs, nothing unreviewed.

  4. verify

    Tests, evals and a reference to compare against. Results are reproducible.

  5. ship

    Deploy on Vercel or Railway, add monitoring, hand over docs and a runbook.

07 · where I've worked

Employers anonymised, on purpose.

AI Specialist · Full-Stack Developer · LLMOps · Digital Media Strategist

  1. 2026 → now

    Full-stack developer · legal and consulting firm

    AI document checks, Telegram bots and Mini Apps, internal tools, LLM integrations and RAG.

  2. now

    AI tools and content · HR-tech reference platform

    Labor-law calculators, interactive checklists and flowcharts, a document constructor, AI video production.

  3. 2026

    Remote contract · client software

    Two-month remote engagement building client-side software.

  4. 2023 → 2026

    Independent YouTube creator

    Ran several channels end to end, from scripts and editing to thumbnails, analytics and monetization.

  5. 2022 → 2024

    Full-stack developer · professional reference publisher

    A multilingual vacation-day calculator for HR (RU/UZ) and a Telegram bot that routes documents and chases deadlines.

  6. 2021 → 2023

    Technical editor · same publisher

    Technical editing for professional reference systems in business, HR and finance.

08 · recently

2026-10
  • opened 7 pull requests to Telegram Web A (Ajaxy/telegram-tt)
  • reported a Leisen-Reimer precision issue upstream to QuantLib (#2902)
  • published cobol-equiv and pricing-crosscheck
  • building three AI document-check bots for a legal firmnda
  • launched a typical HR procedures guidenda
2026-09
  • started a smart HR document constructornda
  • rebuilt a vacation-pay calculator to an expert specnda
  • designed a 100-rule construction payment-control enginenda
2026-08
  • moved 46 Uzbek-language checklists into interactive formnda
  • launched a due-diligence Telegram Mini Appnda
  • interactive HR audit went live for platform usersnda
2026-07
  • published quasar-solver, a QUBO vs classical benchmark
  • competitive analysis for a consulting groupnda
2026-06
  • started a BBA at University of the People
09

09 · questions clients ask

How do we agree on scope?

After a 20-minute call and a short written brief you get a fixed scope and a date, so the plan does not grow on its own.

How long does it take?

A bot or a site usually takes 1–2 weeks, an AI feature 2–4 weeks, a web app or MVP 4–8 weeks. You see a working version early and then every week.

Who owns the code?

You do. At handover you get the repository, documentation, a runbook and tests, deployed to your accounts.

Will you sign an NDA?

Yes. Most of my paid work is under NDA, which is why client names on this site are hidden.

What happens after launch?

I stay available for fixes and changes. Ongoing support or a monthly retainer can be agreed separately.

Which languages do you work in?

English, Russian and Uzbek. My working day in Tashkent (UTC+5) overlaps Europe fully and US East Coast mornings.

Got something that has to work in production?

Email me two lines about the task: what it should do and when you need it. Calls work too, my day overlaps Europe fully and US East Coast mornings.

copy emailgithublinkedin