Aziz Riskulov

Full Stack AI Engineer

Most AI demos never survive
contact with real users

I build the ones that do. Voice agents that answer the phone and write to a CRM, assistants that answer from a company’s own documents, and the evaluation and deployment underneath them that keeps working on the days nobody is watching.

See the work
CVRésuméGitHub
In production
WhatsApp assistant answering real messages every day
4993 → 3217 ms
first-turn latency, after measuring an architecture and deleting it
5 eval families
route, ground, reach, sales, escalate — run before every change

How I work

The difference is never the model. It is what you do around it

  1. 01

    Measure it, then decide

    I built the specialists-behind-a-router architecture everyone recommends for voice agents, measured it against a single agent, and deleted it. Same eval score, 2.3x slower on the first turn, plus two failure modes the split introduced.

  2. 02

    Evaluate it before a customer does

    Every agent I ship has eval cases that name the tool it must call, the tools it must not, and what the answer has to contain. They run from the command line and fail loudly, which is the only reason the comparison above was possible.

  3. 03

    Design for the wrong answer

    Reads happen immediately. Writes wait. An assistant that can send email will eventually misread something, so nothing that changes state runs unattended, and the confirmation lives in the data model rather than in a prompt.

  4. 04

    Ship it and keep it running

    Containers, TLS, CI that deploys on push, health-gated rollout and one-command rollback. I hand over the runbook, not just the repository. A system nobody can operate is not finished.

About

Who I am

I’m Aziz, a full stack AI engineer with six years of shipping software other people depend on, across pharma, financial platforms and enterprise systems.

I work across the whole thing. Backend at close to senior level, and frontend from high-traffic applications down to the responsive interface someone actually uses. Most of what I have built sits where a product meets the messy systems around it: a CRM that has to stay correct, an ERP that owns the numbers, a queue that must not process the same request twice.

I have been building with language models since 2023, when I shipped an OpenAI plugin with fine-tuning and chatbots to a paying WordPress audience. Now it means agents: voice agents that answer the phone and write to the CRM, assistants that answer from a company’s own documents, the evaluation harnesses that decide which design survives, and the deployment underneath all of it. The interesting problems are the same ones they always were, and the model is rarely the hard part.

I like fixing things, building things, and delivering on time.

What I work with

Languages

  • Python
  • TypeScript
  • JavaScript
  • PHP
  • Java
  • C++
  • SQL

Backend

  • FastAPI
  • Laravel
  • Node
  • Express
  • Nest.js
  • REST
  • GraphQL
  • Yii

Frontend

  • React
  • Redux
  • Next.js
  • Vue
  • Sass
  • GSAP
  • Three.js

Data

  • PostgreSQL
  • pgvector
  • MySQL
  • DynamoDB
  • Redis

Cloud and delivery

  • AWS Lambda
  • SQS
  • S3
  • Azure
  • Docker
  • GitHub Actions
  • Caddy
  • Linux

AI systems

  • LiveKit
  • LangGraph
  • Claude
  • Gemini
  • OpenAI
  • Deepgram
  • RAG
  • fastembed
  • Evaluation suites
  • Guardrails and approval gates
  • Agentic coding

Integrations

  • HubSpot
  • Slack
  • WhatsApp Cloud API
  • Google Workspace
  • Stripe-style payments

Testing

  • Pytest
  • Jest
  • XUnit
  • Integration
  • End to end
  • Load testing

Experience

Where I have worked

  1. 2026 to now

    Full Stack AI Engineer

    Graphnoe, own practice, remote

    • Built and deployed a WhatsApp assistant that answers from a company's own documents and acts in Google Workspace: LangGraph orchestration, Claude and Gemini behind one interface, retrieval over pgvector with embeddings computed locally by fastembed, voice notes transcribed with Deepgram.
    • Built voice agents on LiveKit that identify the caller, answer stock and order questions from live ERP data, book appointments, and write contacts, deals and call summaries back into HubSpot.
    • Wrote the eval harness that makes architecture decisions measurable: cases asserting the tool an agent must call and the ones it must not. It is how a specialists-behind-a-router design lost to a single agent at equal accuracy, 4993 ms to 3217 ms on first turn, and was deleted.
    • Designed the guardrails around write access: reads run immediately, writes wait for human confirmation, and the gate is a conditional database update rather than a prompt instruction, so concurrent confirmations still produce exactly one action.
    • Ran the delivery side: Docker, Caddy with automatic TLS, GitHub Actions building to GHCR with health-gated rollout and one-command rollback, on a hardened VPS.
  2. 2024 to 2026

    Full Stack Developer

    Uravo, remote

    • Built event-driven integrations across CRM and messengers with AWS, Python, TypeScript, Node.js and Nest.js, so records stayed correct in real time instead of being copied by hand.
    • Built the candidate-facing UI of a recruiting platform handling thousands of records with React and Redux.
    • Wired PostgreSQL, REST APIs and webhooks into a two-way sync and stabilised a payment integration, keeping subscription billing accurate across connected systems.
    • Containerised services and automated deployment with Docker and GitHub Actions.
    • Ran sprint planning, standups and code reviews while leading two junior engineers.
  3. 2023 to 2024

    Full Stack Developer

    Stylemixthemes, remote

    • Built AI Bud WP, an OpenAI-powered WordPress plugin shipped to a paying plugin audience: model fine-tuning, long-form post generation and niche chatbots, in 2023.
    • Worked on the Cost Calculator plugin: payment handling, a webhook mechanism and financial formulas.
    • Added quizzes, gamification and real-time analytics to the MasterStudy LMS plugin.
  4. 2022 to 2023

    Full Stack Developer

    WP Webhooks, remote

    • Built more than 25 WordPress plugin integrations, including UltimateMember, WooCommerce and Yoast SEO.
    • Integrated third-party services over REST APIs, including OpenAI, WhatsApp and Zoom.
    • Worked across CI/CD pipelines, code review, and load, unit, integration and end-to-end testing.
  5. 2020 to 2022

    Full Stack Developer

    Zmarket, Tashkent

    • Implemented a CRM system for accountants, managers and vendors with Vue, Laravel and MySQL.
    • Built a payment and penalty collection system using Laravel jobs, queues and Redis.
    • Integrated PayMe, Upay, Humo and Uzcard, plus fiscal tax reporting through 1C over custom REST endpoints.
    • Automated marketing experiments including A/B, auto-targeting and multivariate testing.
Education
BSc Information Technology and Engineering, Inha University in Tashkent
Languages
English, Russian, Uzbek

Selected work

Two systems, in production

Watch demo

AI voice agents for CRM & ERP

Phone agents that identify the caller, answer order and stock questions, book appointments, and write everything back into HubSpot.

LiveKitHubSpotPostgrespgvectorPythonGitHub
Watch demo

WhatsApp assistant over your own documents

Answers from a company's documents and reaches into Gmail, Calendar and Drive, proposing every write before it makes it.

ClaudeLangGraphFastAPIpgvectorDeepgramGitHub
Watch demo

Interview scheduling across HubSpot, AWS and Slack

A form submission becomes a Slack message a coach can claim, without two coaches claiming the same one.

AWS LambdaSQSDynamoDBHubSpotSlack

Chat and file sharing in a recruiting platform

Recruiter and candidate messaging with attachments, on a platform handling thousands of records.

ReactReduxLaravelAWS S3MySQL

Contact

Have something that should answer for itself?

Tell me what it needs to do and which systems it has to reach, and I’ll tell you whether it’s worth building.

vinceere@gmail.comRead the blog