TAS18:18:33
···13:18:33
00:00
X0.0000
Y0.0000
Not what a model outputs - how the system decides, executes, and holds under load.

Shahriyor

Abdullayev

Role

AI Systems Engineer

Company

Biznesni Rivojlantirish Banki
March 2026 – Present

Optimising: Residuals • Not: Roles
Shahriyor Abdullayev
Software Engineering - TATUPython • PyTorch • LLMs • RAG
Scroll
About

Turning data into decisions.

Curious by default. Practical by choice.

I'm Shahriyor Abdullayev, a machine learning and AI engineer. I analyze data and build simple, effective technology solutions for real business problems - currently as an ML/AI Engineer at the Business Development Bank, where I work on data analysis, ML models, and bringing AI into business processes.

My path started with Python: automation scripts, small tools, and data workflows. That pulled me into data science - forecasting, classification, and turning messy datasets into something a business can actually act on.

Today I focus on the full lifecycle of an ML product: understanding the data, building and evaluating models, and integrating them into real processes. My goal is reliable, fast AI that delivers genuine value to the people using it.

Data Before Models

A model is only as good as the data behind it. I start by understanding the data, cleaning it, and framing the real business question - before reaching for an algorithm.

Reliable, Fast, Useful

A model that only works in a notebook isn't finished. I care about accuracy, latency, and shipping something a team can actually rely on in production.

Simple Solutions to Real Problems

I look for the simplest approach that solves the actual problem. Clear, maintainable solutions beat clever ones nobody can run six months later.

What I don't do

- I don't chase complexity for its own sake. If a simpler model does the job, that's the model.

- I don't ship black boxes I can't explain. If I can't reason about why it works, it isn't ready.

- I don't ignore the data. Most ML problems are data problems first.

• Highlights •
ML/AI Engineer
2026 – present
Business Development Bank
5+ AI Projects
2022 – present
Forecasting • classification • analytics
95% Model Accuracy
2026
Credit risk scoring model
Credit Risk Scoring
2026
Python • Pandas • Scikit-learn
AI Assistant
2025
Document Q&A • LLM • FastAPI
Customer Analytics
2024
Segmentation • KPIs • Power BI
ML/AI Engineer
2026 – present
Business Development Bank
5+ AI Projects
2022 – present
Forecasting • classification • analytics
95% Model Accuracy
2026
Credit risk scoring model
Credit Risk Scoring
2026
Python • Pandas • Scikit-learn
AI Assistant
2025
Document Q&A • LLM • FastAPI
Customer Analytics
2024
Segmentation • KPIs • Power BI
Experience & Education

The trajectory.

2022
2023
2024
2025
2026
2027
Lead Math & Critical-Thinking TutorFeb 2023 – Jun 2023
Wunderkind Learning Center
Taught maths and logical reasoning to students aged 6–18, building individual study plans and prepping students for olympiads.
Data ScientistSep 2023 – Jun 2025
Freelance
Delivered NLP, computer vision, and analytics projects — sentiment analysis, object detection, and Power BI / Tableau dashboards across domains.
Data ScientistApr 2025 – Oct 2025
Universalbank
Built predictive models and process automation — saving $50k+/yr, cutting costs ~20%, and reducing the database footprint 50%.
AI EngineerMarch 2026 – Present
Biznesni Rivojlantirish Banki
Production AI for banking — fraud detection, credit-risk models, and LLM/RAG assistants deployed with Docker, FastAPI, and CI/CD.
BSc Software Engineering2022 – 2026
Tashkent University of IT (TATU)
Software engineering degree with a focus on machine learning, data science, and applied AI.
Impact

Proof, not promises.

Scale & PerformanceAccuracy & QualitySystems & DeliveryReliability & Ops
Hover a point - every number is delivered, not projected.
Selected Work

Things I've built.

Real-Time Fraud Detection System

Production
Biznesni Rivojlantirish BankiBiznesni Rivojlantirish Banki
ProblemFraudulent transactions had to be caught as they happened - batch review surfaced fraud only after the money had already moved.
SystemBuilt a real-time scoring service that scores every transaction against learned behavioural baselines and flags anomalies instantly.
DesignEngineered features from transaction history and customer profiles; handled severe class imbalance so rare fraud signal isn't drowned out.
OutcomeSuspicious operations now flagged automatically in real time, cutting manual review load and strengthening the bank's security posture.

Credit Scoring & Risk Model

Production
Biznesni Rivojlantirish BankiBiznesni Rivojlantirish Banki
ProblemLending decisions leaned on manual heuristics that were slow, inconsistent, and hard to defend when a borrower defaulted.
SystemTrained ML scoring models that estimate a customer's repayment ability from financial and behavioural history.
DesignCalibrated outputs into interpretable risk bands so credit officers get a probability, not a black-box yes/no.
OutcomeRepayment-ability scoring now supports data-driven, auditable lending decisions across the bank's retail portfolio.

Banking LLM Assistant with RAG

In Development
Biznesni Rivojlantirish BankiBiznesni Rivojlantirish Banki
ProblemCustomers and staff needed answers from dense internal banking documents - a plain LLM would confidently invent policy that doesn't exist.
SystemBuilt a RAG system that retrieves the relevant document passages first, then has the LLM answer grounded in that evidence.
DesignAdded an intent-classification pipeline that routes incoming customer queries to the right flow automatically.
OutcomeStaff and customers get document-grounded answers to internal-policy and product questions, reducing escalation to human agents.

Customer Churn Prediction & Segmentation

Production
Biznesni Rivojlantirish BankiBiznesni Rivojlantirish Banki
ProblemThe bank learned a customer had left only after they were gone - too late to do anything about it.
SystemBuilt churn-probability models that flag at-risk customers early from changes in their activity patterns.
DesignLayered segmentation on top so retention effort is targeted at the right groups, not sprayed across the whole base.
OutcomeRetention teams can now act before a valuable customer leaves, prioritising outreach by predicted churn risk.

Banking Process Automation Suite

Production
UniversalbankUniversalbank
ProblemCore analytical and reporting workflows were manual and repetitive, costing analyst hours and money every month.
SystemBuilt automation tooling that takes over the repetitive data and reporting work end-to-end.
DesignRe-worked the data layer - trimming a bloated database and streamlining queries for speed.
Outcome$50,000+/year saved, costs down ~20%, database size cut 50%, with projects delivered ahead of deadline.

NLP Text Classification & Sentiment Analysis

Shipped
FreelanceFreelance
ProblemClients had large volumes of unstructured text - feedback, messages, documents - with no scalable way to sort it or read sentiment.
SystemBuilt transformer-based NLP pipelines that classify text and detect sentiment automatically.
DesignStandardised the flow on Hugging Face transformers so models could be swapped and fine-tuned per use case.
OutcomeText is now auto-categorised and scored for sentiment, turning raw text into structured, actionable signal.

Computer Vision & Object Detection

Shipped
FreelanceFreelance
ProblemClients needed software to understand images - recognise content and locate objects - not just store them.
SystemBuilt image-recognition and object-detection models in TensorFlow and PyTorch.
DesignWired in OpenCV for real-time frame processing so detection runs live, not just on stored files.
OutcomeDelivered working recognition and detection with real-time analysis for the client's use cases.

Business Intelligence Dashboards

Shipped
FreelanceFreelance
ProblemStakeholders were making calls on raw spreadsheets - slow to read, easy to misinterpret, impossible to explore.
SystemBuilt interactive dashboards in Power BI and Tableau on top of cleaned, query-ready data.
DesignDid the unglamorous work first - cleaning and reshaping large datasets with SQL, Pandas, and Excel.
OutcomeDecision-makers get interactive, trustworthy views of large datasets instead of static, error-prone reports.
Stack

Tools I build with.

Profiled under load. Not just imported.

Languages
Python(data · ML)
SQL(PostgreSQL, MySQL)
JavaScript(JS / TS)
Bash
ML & Data Science
scikit-learn
PyTorch
Pandas / NumPy
Hugging Face
LLM / RAG(LangChain)
Backend & APIs
FastAPI
Next.js
REST APIs
Data & Infra
PostgreSQL
Docker
Git
Power BI
Contact

Hard problems welcome.

Optimising: Residuals • Not: Roles

Heads-down building right now - not looking for roles. But if you've got a hard problem, a wild idea, or just want to talk shop about LLMs, distributed systems, scientific ML, or why this site is unreasonably over-engineered for a portfolio, I'm always up for that.

🎨 Vision & design by Shahriyor Abdullayev • ⚡ Engineered with Claude Code • 🚀 Deployed on Vercel