Portfolio · updated 31.10.2026 · Русский

Vilsur Garifov

AI Implementation Project Manager · AI Project Manager

Founder of AI Talent Genome · Founder & CEO of Q Core Center (2021–2026)

I find where a process loses hours, put the right tool in, and teach people to use it.

Since 2019 it has been one job, only the tool changed: CRM then, AI now. I choose the model by measurement and check quality before a person sees the result. With AI agents I took AI Talent Genome, an AI interview platform, all the way to interviews with real candidates.

Book a 15-minute call See the case

23

AI interviews with real candidates in the first month of the voice interviewer

29.09–24.10.2026

36 → 13–15 min

to analyze a 30-minute recording after switching the model by a blind test; from question to launch in 4 days, the test cost $0.07

15.10–19.10.2026

215

releases; now 188 automated checks before each one

30.07–27.10.2026

2,500+

commits in development with AI agents over 3.5 months; I do the acceptance myself

13.07–27.10.2026

AI Talent Genome · candidate journey
  1. 01hh.kzvacancies and applications are pulled in automatically
  2. 02Resume screeningon the vacancy's axes, one scale for everyone
  3. 03Voice AI interviewera 3D avatar in the browser, no recruiter
  4. 04Recording analysistranscript, speaker roles, posture, voice, language passes
  5. 05Report for HRevery score has a quote and a timestamp
For the company

What I do for a company

I own the path from task to result: research, prototype, launch, user training.

I put AI where it pays off and decide by measurement

First I measure where the hours go, and only then choose the tool. In my product I switched the model for recording analysis after a blind comparison that cost $0.07; analysis now takes 13–15 minutes instead of 36.

Measurement and model switch

I release AI to people only after a check

If a check fails, the report never reaches HR: the system verifies the transcript, roles, quotes and scores itself and sends the candidate for a re-run. Now there are 188 automated checks before every release.

How the check works

I take a decision to production in days

With AI agents a feature goes from decision to working in days: a new landing page in under 8 hours, pricing with a migration and 24 tests in 2 days. The voice interviewer with a 3D avatar went from a document to real candidates in about two months.

Speed table

I take implementation all the way to people: processes, CRM, training

Implementation ends when people actually use the tool. Before AI I implemented amoCRM for 10 clients in four countries; now I am the author of a practical-AI professional development program for a department of Saginov Karaganda Technical University.

KarTU department post · CRM implementations

Main case

AI Talent Genome: AI implementation in hiring

12.2024 — present · Head of Product and AI Implementation (founder)

An AI interview platform: from an application on hh.kz to an HR report with quotes from the conversation. My role: I decide what to build, measure the result and accept every release myself. Since July 2026 the code is written by AI agents under my direction. The product is now in support mode.

Usage

466candidates in 19 vacanciesas of 19.10.2026
97interview recordings analyzedas of 19.10.2026
23AI interviews with real candidates29.09–24.10.2026

17.10.2026Booth at the “Creative Technologies of Karaganda” event: analyzed visitors' speech live, including in Kazakh.

The task

Hiring means a lot of information: hours of interviews, recordings to re-watch, and everything about every candidate has to be kept in your head. The task was twofold: take this routine off HR and at the same time make the selection deeper, so that a manager reads a report in minutes where every score rests on the candidate's own words.

Milestones

  1. 2024–2025Idea and customer discovery
  2. 26.05–15.08.2025Terricon Valley Accelerator, project on the demo day of 15.08.2025
  3. 12.2025MVP with a contractor: the first interview was run by a live recruiter; the project was presented by the Akimat (regional government) of Karaganda Region
  4. 07.2026Developing the product myself with AI agents
  5. 29.09.2026The voice AI interviewer works with real candidates
  6. nowProduct in support mode

What is built

  1. 01

    Entry via hh.kz. Vacancies and applications are pulled in automatically.

  2. 02

    Vacancy axes and resume screening. The model derives the skills for the role and freezes them so that all candidates are scored on one scale; frequent market requirements from hh are added by profession. Applications are scored on the vacancy's axes and ranked.

  3. 03

    Voice AI interviewer. The candidate follows a link, picks a time, checks camera and microphone, gives consent and talks to a 3D avatar in the browser, with no recruiter. Questions are built from the vacancy and the resume; the interviewer asks follow-ups.

  4. 04

    Recording analysis. Transcript, speaker roles, posture and gaze, emotion in the voice; then the language passes: scoring on the axes, methodology, profile, timeline, cross-check against the resume.

  5. 05

    Report for HR. Every score has a quote and a timestamp. Strengths, risks, what to ask next; a link for the manager without login; comparison of the vacancy's candidates.

    Match = axes 0.6 + methodology 0.3 + company culture 0.1

Around it: a Telegram bot with events and server monitoring; an admin panel (users, limits, promo codes, language-model switch); a booth for events.

How it is built: one person and AI agents

Who does the work
Since 13.07.2026 only I and AI agents (Claude Code) work in the repository. Agents write the code; what to build, how to measure and what to release is my decision.
Scale
The contractor's MVP: about 9k lines, 8 services, 19 migrations. Now: about 223k lines including tests, 15 services, 78 migrations.
Pace
2,544 commits of mine over 82 days with commits (13.07–27.10.2026); 77% of the repository's commits are marked as co-authored by an AI agent. 215 successful releases in 44 days.
Rules for agents
A 1,066-line rulebook: limits of authority, the measure of quality, what an agent is not allowed to do. A log of 307 measurements with their conditions, 161 analyzed cases.
Control over agents
Any change to prompts, thresholds, axes or score wording is tagged “ANALYTICS CHANGE”: 242 such commits, and I see every one. Important fixes are re-checked by an independent group of agents: the fix of 25.10.2026 had three reviewers and one refuter per finding, nine in total.
Tests
161 backend test files, 105 frontend specs, 13 E2E specs. Now 188 automated checks before a release; I cut the full run from 50.9 to 25.7 minutes.
Stack
FastAPI, PostgreSQL, Redis, RabbitMQ, Vue 3; Whisper, pyannote, YOLOv8-pose, MediaPipe; local and cloud language models; Unity WebGL for the avatar.
Whose work it is
The MVP with hh login and a basic analyzer was started by a contractor. The original avatar head model and face animation came from a third-party developer, with rights for commercial use. The voice interviewer, the current analysis, the checks and everything done after July 2026 is my work with agents.

Speed: from decision to working

WhatPathTime
Moving the language stages of analysis to a cloud model, together with a blind comparisonquestion 15.10 → code 18.10 → comparison and launch 19.10.20264 days
A booth for an eventbrief without prototypes 11.10 → live 12.10 → event 17.10.2026about 1.5 days
New landing page, 149 of 149 checks passedlayout chosen 05.10 → live 06.10.2026under 8 hours
Per-vacancy pricing: migration, admin button, 24 tests“pricing — OK” 05.10 → live 07.10.20262 days
The “candidate spoke but was not heard” failure: fix and independent review by nine agentsincident 23.10 → live 25.10.2026about 2 days
Voice AI interviewer with a 3D avatardocument 31.07 → first real candidates 29.09.2026about 2 months

Decisions with numbers

I switched the model only after a blind test

  • The measurement showed the bottleneck: about 70% of analysis time went to the local language model.
  • Blind comparison on 19.10.2026: five candidates, one prompt, three models — the local one and two cloud models. Five AI evaluators judged; the answers were shuffled and the author not named. All five ranked the junior cloud model first and the local one last.
  • Grounding of conclusions in the transcript: 8.2 out of 10 for the chosen model, 6.4 for the second cloud model, 2.2 for the local one. The local model inflated the match by 17–33 points and passed off a line from the resume as the candidate's words.
  • The comparison cost $0.07. From question to launch — 4 days.
  • Result: the language stages take 35–87 seconds instead of about 12 minutes; the whole analysis of a 30-minute recording takes 13–15 minutes instead of about 36. The local model stayed as a fallback: a switch in the admin panel.
  • One more argument: in the local model's context window an hour-long recording was cut to 59% of the speech; in the cloud model all 21 long recordings fit entirely.

Speaker-role errors → an interviewer of my own

A check against a manual reference showed that 23.6% of segments were labeled wrong, and the confidence flag did not see it. The errors inflated the match: 57.5 versus 50.1 with correct labels. The decision — an interviewer of my own, where the speaker roles are known by construction.

The model copied the example from the instruction into the report

A made-up “negotiator” from the prompt's example ended up among the candidate's risks in 9 cases out of 12. After the fix — 0 out of 8; examples now show the form, not the values, and a separate check step guards the rule.

Speech recognition made up text

Instead of a 26-second answer it produced a meaningless Latin string. After the fix the transcript's similarity to the reference rose from 0.631 to 0.739, and the invented Latin disappeared.

Hypotheses are tested on a real recording

Noise suppression looked like an obvious improvement, but at the event on 17.10.2026 it made the models invent text. After the test run it was rejected.

How I check that AI did its job

  • An unfinished report is never shown to a person. The system checks the transcript, roles, quotes and scores itself. If a report is empty through the system's fault, the candidate is re-run and the event goes to Telegram.
  • Quotes are verified by code against the transcript. Over 147 runs and 885 quotes, unmarked cut-offs were brought down to zero: there were 188 at the start of a quote and 356 at the end.
  • I compare models blind before replacing one. The AI evaluators do not know whose answer is whose.
  • In the conversation the plan is driven by code, not by the model. A question pointer, at most three follow-ups per question, each reply tied to an axis. Praise, scoring and hints out loud are forbidden by code.
  • What is deployed, I ask the system. At release the image is stamped with the commit it was built from.
  • Kazakhstan's AI law is taken into account. The candidate is told that an AI is working with them; consent for emotion detection is asked; they may object to automated processing. Behavioral signals are not part of the match: they are a hint for a person, and the candidate is compared with their own baseline, not with a standard.

Cases from operation

The candidate was not heard

On 23.10.2026 a candidate spoke for three minutes and the interviewer did not hear him. On 25.10 the fix went live: the voice threshold is set from the noise level, the candidate gets a hint, HR gets an alert. The same night nine agents reviewed the fix independently, and what they found was fixed and released within an hour.

“Success” with no result

The nightly backup of recordings failed during archiving for lack of space, and the task reported exit code 0: PowerShell changes the exit code only for external commands. Lesson: the success marker checks that the file is there and has the right size, not the return code.

Honest status, no quality

The run finished, status “analyzed”, and the system logged a quality caveat: the worker held more memory than allowed, and that never reached the status. Lesson: after a run, ask the system about quality, not just look at the status.

Someone else's system with no hints

Docker Desktop stopped starting after an emergency stop: stuck zero-size socket files that Windows would neither open nor delete. The solution — rename the parent folder with all processes stopped; the whole sequence (stop, disk compaction, socket cleanup, platform liveness check) is in one script run by the scheduler.

Recognition

  • The Akimat of Karaganda Region presented the project in an official release on 18.12.2025. Over 18–19.12.2025 the release produced 10 publications, among them Sputnik Kazakhstan and Profit.kz; in January 2026 Liter.kz ran a separate piece with my comments. All links
  • Terricon Valley: Accelerator participant (26.05–15.08.2025), project on the demo day of 15.08.2025. Demo day story
  • The Department of Economics and Enterprise Management of Saginov KarTU thanked “the founder of AI Talent Genome” for a talk at a round table with business (20.10.2026). Department post
  • Antiplagiat: “leading AI expert Vilsur Garifov (founder of AI Talent Genome)” — about the Winter School of Buketov Karaganda University (16.01.2026). News item
  • Said — done. In December 2025 the in-house AI interviewer appeared in the release as a plan. Since 29.09.2026 it works with real candidates.

What did not work and what it taught

  • Five weeks of fixing symptoms instead of the cause. A small local model failed again and again, and each failure was fixed separately until I asked whether it had to be local at all. Now it is a rule: any question about capacity starts with “can the constraint be removed entirely?” and only then “how to speed it up?”.
  • The system was confidently wrong and could not see it. 23.6% of speaker roles were labeled wrong while the confidence flag stayed calm. Lesson: measure quality against a manual reference, not the system's self-assessment, and sometimes change the product's design rather than fix the model.
  • An obvious improvement turned out harmful. Noise suppression at a live event made the models invent text. Lesson: a hypothesis is tested by a run on a real recording, not by reasoning.
  • I paid for accuracy with locality. The language stages moved to a cloud model; speech recognition, vision and voice stayed on our own server, and the local model is a fallback.
Founder

Q Core Center: a company from scratch

09.2021 — 10.2026 · Founder & CEO

An outsourced English-speaking call center in Karaganda: through it specialists from Karaganda worked remotely for foreign companies, including US ones. I built the company from scratch and ran it end to end.

  • People: employee profile, hiring, intern training, KPIs and motivation.
  • Processes: chose and set up the CRM, wrote the procedures and the knowledge base.
  • Operations: office management, procurement, document flow, legal and accounting, finance, management.

Bridge to AI. AI Talent Genome was tested on 50 real interview recordings from Q Core Center (vacancy: Manager of Customer Care Department).

Why it matters for you. I built the process from scratch myself — from hiring to procedures — so I know where it loses hours.

Interview. “Terricon Stories: Q Core Center” — Terricon Valley's channel, 29.05.2024, 21 minutes, in Russian: how to open an English-speaking call center, launch and promotion, working with the team, training interns, working with clients.

Video on YouTube · Terricon Valley post

Before AI

CRM implementations

The implementation cycle is the same, only the tool changed: CRM then, AI now.

SalesBand

04.2021 — 08.2022 · Senior Business Analyst & CRM Integrator

System integration and business process automation on amoCRM.

10amoCRM implementations and improvements
4countries: Kazakhstan, Kyrgyzstan, Russia, Montenegro

Full cycle

  1. stakeholder interviews
  2. BPMN maps
  3. solution concept
  4. configuration
  5. webhooks and lead transfer
  6. staff training, dashboards, reporting

Clients

  • BuildInvest
  • Mogo Kazakhstan
  • Namba One
  • Change Fitness
  • RTA
  • Chestny Agent
  • ARCADIA VR
  • GOISRAEL
  • RPP Camp
  • PROFREAB

Wooppay

09.2019 — 05.2021 · Business Process Optimization Manager → Senior Corporate Account Manager

Fintech, payment services. Business process optimization: BPMN maps, procedures, knowledge base, finding bottlenecks. Set up amoCRM myself, trained 5 managers to use it, coordinated 2.

Teaching

Talks and teaching

Departments and faculty

Author of a practical-AI professional development program for a department of Saginov KarTU: taking routine off the teacher, lesson plans, interactive assignments, grading and feedback, AI in the processes of the department and the university. A talk at the Winter School of Buketov Karaganda University for university staff; workshops for teachers.

Result: at the EdSkills workshop teachers built lesson plans with AI.

KarTU department post · Antiplagiat

Beginners

The course “AI tools and freelancing: from skills to income” — 15 sessions, from an introduction to AI tools to defending one's own AI projects; the “Vibe coding” intensive.

Result: the final assignment of the KarNU intensive — a prototype AI consultant for university applicants, built by the participants.

Course page · KarNU department post

Teams

Open workshop “Claude + Asana: how to embed AI in task management”.

Result: an open recording, 1 h 44 min, in Russian.

Recording

All talks · newest first · materials in Russian

DateTopicVenueLinks
20.10.2026 Talk “AI in forecasting: new opportunities for effective business decisions” at the round table of the Department of Economics and Enterprise Management with business Saginov KarTU department post
02.07.2026 Open workshop “Claude + Asana: how to embed AI in task management” Terricon Valley recording, 1 h 44 min · organizer's report
30.04.2026 Lecture “Hiring, conflicts and clients” with a role-play practicum for students Saginov KarTU post
03.2026 Intensive “Vibe coding: an AI consultant for university applicants”: two sessions and a week of practice, the final assignment — a prototype Buketov KarNU, marketing department department post
09.02.2026 Professional development program on the practical use of AI (author): working meeting with the department Department of Economics and Enterprise Management, Saginov KarTU department post
29.01.2026 Workshop on educational innovation and digital tools for teachers (EdSkills program) Karaganda Higher Polytechnic College college post
21.01.2026 “AI in education: new opportunities for teachers” Bukenov College, Karaganda; with Terricon Valley college post · organizer's report
01.2026 Talk for university staff at the Winter School “Academic Foresight 2026” Buketov KarNU Antiplagiat · post
28.11.2025 Workshop “How to find a niche with AI and run quick market and competitor research” Buketov KarNU post
26.11.2025 AI Workshop “From skills to income” Astana Hub, Aqmola Hub, Kokshetau event page · post
23.10.2025 Workshop “How to turn a product into a service with AI and launch an MVP without big budgets”: customer discovery and quick prototypes Saginov KarTU post
08–26.09.2025 Speaker of the “AI and automation” block of the course “Digital Creator 2025: design, video, AI and Telegram bots” Terricon Valley, Osakarovka course results · announcement
29.04.2025 Workshop “Startups, products and research with AI” Karaganda Medical University post
04.2025 Workshop “Startups, products and research with AI” Saginov KarTU post
03–28.03.2025 Course “AI tools and freelancing: from skills to income” — 15 sessions, from an introduction to AI tools to defending one's own AI projects Terricon Valley; Saran Higher Humanitarian and Technical College and online course page · results · post
19.11.2024 Lecture “Fundamentals of goal setting” Terricon Valley; Saran Higher Humanitarian and Technical College and online recording, 1 h 10 min
01.02.2024 Goal-setting and soft-skills training for students of the “Two Foreign Languages” program (grant project) Faculty of Foreign Languages, Buketov KarU, KazTEA project page · post
Press

Media about the project

In December 2025 the press described an MVP where the first interview was run by a live recruiter, and an in-house AI interviewer was in the plans. Since 29.09.2026 that interviewer works with real candidates. The materials are in Russian unless noted; titles are translated.

Where I am named

  1. 22.01.2026Liter.kz“AI does the hiring” — a piece with my comments
  2. 16.01.2026Antiplagiat, antiplagiat.ru“An Antiplagiat expert spoke at Buketov Karaganda University”
  3. 19.12.2025Profit.kz“AI will help employers find staff”
  4. 19.12.2025Registr TV, regtv.kz“AI will help employers find staff”
  5. 18.12.2025Akimat of Karaganda Region, gov.kz“AI against staff turnover: how AI Talent Genome helps employers find their people”
  6. 18.12.2025QazMonitor (in Kazakh)“A domestic platform that helps select staff has started working in Kazakhstan”
  7. 18.12.2025Registr TV, regtv.kz“AI against staff turnover: how AI Talent Genome helps employers find their people”
  8. 18.12.2025biznescentr.kz“AI against staff turnover: how AI Talent Genome helps employers find their people”
  9. 18.12.2025vaktau.kz“Kazakhstani platform AI Talent Genome uses AI to improve hiring and retention”

About the project, without my name

  1. 18.12.2025Sputnik Kazakhstan“For the first time in Kazakhstan employers look for new staff with AI”
  2. 18.12.2025Finratings.kz“The resume is not the main thing: Kazakhstan has started selecting employees with AI”
  3. 18.12.2025Ustinka LIVE, ustinka.kz“AI will help employers find staff” (Sputnik reprint)
  4. 18.08.2025Terricon Valley, terricon.kz“Demo day for the finalists of business programs held at the IT hub”

Video interviews

  1. 17.12.2025Akimat of Karaganda Region, official pageInterview about the AI Talent Genome project
  2. 29.05.2024Terricon Valley, YouTube“Terricon Stories: Q Core Center” — how to open an English-speaking call center, 21 min
Contact

Let's talk for 15 minutes

In 15 minutes I will show the product live and walk through one decision with numbers — how I chose the model, or how I keep defective output away from people. You will see whether this approach fits your task.

Message on Telegram Email me Message on WhatsApp
Phone, WhatsApp
+7 700 600 00 36
Email
nx0036@gmail.com
Telegram
@will_sure

What I am looking for

Today one person with AI agents can build a working product — I did it with AI Talent Genome. But scaling a product alone is slow and a team is faster, so AI Talent Genome is now in support mode, and I am looking for a company where AI is at the center of the product or its processes.

Prove it in practice

Instead of a test assignment, at your request I will run a demo interview for a role from your vacancy: you will see the conversation with the AI interviewer and the report that HR receives.

Terms

  • Remote. Karaganda, Kazakhstan, UTC+5.
  • Different forms of cooperation: employment, contract, project — we will pick the one that suits you.
  • Languages: Russian (native), English (B1).