PhD in AI Engineering · LLM · ML · Deep Learning

I build ML, deep learning and LLM systems
from data to production.

LLM EngineerRAG EngineerAgentic AI EngineerML & Deep Learning EngineerComputer Vision

AI engineer across the full stack: classical machine learning, deep learning, computer vision and LLM systems. I take models from data to production: pipelines, training, evaluation, deployment and monitoring, from published ML research to multi-agent LLM platforms.

LLM & RAG Projects

LLM Systems Agents, RAG & Post-Training

ERA Academy: an AI learning platform with one agent engine behind three assistants and permission-aware RAG. TACTIQ: evidence-first football intelligence, phase 1: the indexing pipeline into Qdrant. Touchline: a Claude scouting agent on LangGraph with cited reports, traced in LangSmith. TeachHelp: SFT, DPO and RLHF on Qwen2.5, 15/68 → 44/68. Portiq: 11 agents over Kafka and A2A with a Judge agent.

CARD 09Mythic

ERA Academy

AI EdTech Platform · Multi-Agent Engine · RAG · pgvector

“One AI engine, three assistants (student tutor, teacher and admin), and each one only ever sees what its user is allowed to see.”

PWR90
SCL88
AUT82
INT90
PLH92
  • Full learning platform: course → module → lesson → subskill, progress tracking, goal-based onboarding and an admin CMS
  • One shared AI agent engine powers 3 agents (Student Tutor, Teacher Assistant, Admin Assistant), streamed over SSE
  • Provider adapters for Claude (official Anthropic SDK), OpenAI and local Ollama models such as Qwen2.5, with usage and cost limits
  • RAG knowledge base: PDF, DOCX, TXT, MD → chunking → embeddings (Voyage, OpenAI or Ollama) → pgvector, answers cite sources [S1]…
  • Permission-aware retrieval: the access rule runs inside the SQL query, so no one retrieves another course's or a teacher's private files
  • Prompt-injection defence: documents and selections fenced as data; authorization, not the prompt, is the real control
  • "Ask ERA AI": select any text in a lesson and get an explanation grounded in that lesson
  • 206 automated tests (authorization, HTTP security, AI engine, RAG) all passing, plus headless-browser UI checks
Next.js 15TypeScriptPrismaPostgreSQL + pgvectorSupabaseClaude APIOllamaVercel
CARD 10Legendary

TACTIQ

Football Intelligence · RAG · Qdrant · Investigation Agent

“Ask football questions. Get evidence, not guesses.”

PWR80
SCL78
AUT72
INT86
PLH80
  • Evidence-first analysis for coaches, scouts and analysts: answers with sources and a clear statement of what the evidence can't support
  • Built a document indexing pipeline: load (PDF per page, DOCX, TXT, MD) → metadata → cleaning → token-based chunking → embeddings → Qdrant
  • Stable UUID5 document and chunk IDs: re-ingesting never duplicates, and stale chunks are removed
  • Metadata is never guessed: match, team and player IDs come only from human-written sidecar files
  • Sentence-Transformers embeddings (all-MiniLM-L6-v2, L2-normalised), dimension read from real model output and validated
  • Hand-written pipeline (no LangChain) so every step is visible and testable; unit and integration tests for each step
  • Next: hybrid retrieval with RRF and reranking, a SQL analytics engine, an investigation agent, then V2 video intelligence (YOLO + tracking)
PythonQdrantSentence TransformersHugging FacePyMuPDFpython-docxDockerpytest
IN DEVELOPMENT · PHASE 1
CARD 11Mythic

Touchline

AI Scouting Copilot · Agentic RAG · LangGraph · LangSmith

“Describe the player you need. Get a ranked shortlist and a cited report. The AI never makes a number up.”

PWR90
SCL84
AUT88
INT94
PLH86
  • Plain-language scouting briefs → Claude fills typed tool forms; Python builds the query, so the model never writes SQL
  • LangGraph agent with 7 strict tools (search, profile, similar, compare, notes, shortlist) and an 8-round cap
  • Human-in-the-loop: pauses with a shortlist via interrupt(), resumes from a checkpoint, even after a restart
  • Cited scouting reports: every style claim links to its source passage (Claude citations over Qdrant hybrid search)
  • Warehouse of 50,149 players (Transfermarkt) + 297 StatsBomb matches → 18 per-90 metrics for 594 players
  • 120-question golden set in LangSmith: filter accuracy, precision/recall@10, numeric grounding, invented-player check
  • 102 automated tests · ChatGPT-style Next.js app with voice, file uploads, folders and a model picker
Claude APILangGraphLangSmithQdrantDuckDBFastAPINext.jsPostgres
IN DEVELOPMENT
CARD 12Mythic

TeachHelp

Post-Training · SFT · DPO · RLHF (PPO / RLOO) · Uzbek–English

“From a base model that only continues text to a teaching assistant: SFT, DPO and RLHF, every stage measured on the same 68 questions.”

PWR88
SCL68
AUT64
INT96
PLH86
  • Full post-training journey on Qwen2.5: base → SFT → DPO → RLHF → deployment, 14 models on one 68-question held-out test
  • SFT: Qwen2.5-1.5B 15/68 → 44/68; LoRA vs QLoRA under identical settings (4-bit cost only 2 points)
  • DPO from scratch: v1 learned "shorter is better" (40/68); v2 with length-balanced pairs and an SFT term → 44/68, the final model
  • Classic RLHF (InstructGPT recipe): reward model + PPO, then RLOO, with 4 judges incl. Skywork-Reward-V2 and Gemini as a live judge
  • Reward-hacking checks: a rule score the policy never sees and a length-bias gate. RLHF reached 39–41/68, only as good as its judge
  • Spec-first: MODEL_SPEC.md defines behaviour; DPO v2 refuses all harmful and cheating requests (8/8), "Bilmayman" 5/5
  • TeachHelp: Claude-style chat in English and Uzbek on Vercel, Kaggle GPU backend over a Cloudflare tunnel (~2–3 s per answer)
PythonPyTorchTransformersPEFTLoRA/QLoRADPOPPO / RLOOKaggleFastAPIVercel
LIVE · GPU BACKEND ON DEMAND
CARD 13Legendary

Portiq

Multi-Agent System · LangGraph · Kafka · A2A · pgvector

“Your door to the right job: 11 agents find jobs in Uzbekistan, Korea and worldwide, map your skill gaps and run company-specific mock interviews.”

PWR88
SCL90
AUT92
INT88
PLH84
  • LangGraph orchestrator plans each conversation: language → mode → country → asks only for what's missing → routes the work
  • 11 agents over Kafka topics and A2A (Agent Cards + JSON-RPC): profile, match, company research, interviewer, roadmap, judge…
  • Judge agent checks every output and sends weak work back for a retry before the user sees it
  • Job pipeline: Remotive, Saramin, hh.uz and Telegram channels → normaliser (so'm / ₩ / $ salaries, dedup) → scam filter → Postgres + pgvector
  • Explainable match %: skills you have and miss, salary and visa fit; CV parsing from PDF, DOCX and TXT
  • Mock interviews: 11 rounds (intro … stress … salary … reverse) in Uzbek, Russian, Korean and English, scored on a STAR rubric
  • 30/60/90-day skill-gap roadmap with free official resources; company briefs from Wikipedia + Google News, every line sourced
  • Claude, Gemini or Ollama with deterministic fallbacks; Jaeger tracing across agents, unit tests, end-to-end script and evals
LangGraphKafkaA2AFastAPIPostgres + pgvectorRedisNext.jsDockerJaeger
LIVE PREVIEW · BACKEND LOCAL
ML & Deep Learning Projects

Machine Learning & Deep Learning

LLM from Scratch: a GPT-style Transformer written by hand in PyTorch, with its own tokenizer, data pipeline and training loop. Soccer Player Salary Predictor: a supervised ML pipeline with stacking ensembles, Optuna tuning and SHAP explainability (SCIE 2025). TacticalDNA: unsupervised clustering that matched a blind stat line to prime Messi at 98% similarity. Computer Vision coming next.

CARD 14Legendary

LLM from Scratch

Deep Learning · Transformer · PyTorch · Pretraining

“Every layer written by hand (attention, tokenizer, training loop) and made to pass four sanity checks before a single training step.”

PWR82
SCL70
AUT60
INT92
PLH84
  • GPT-style decoder-only Transformer implemented from first principles in PyTorch: multi-head causal self-attention, MLP blocks, weight tying
  • Verified before training with 4 sanity checks: tensor shapes, random-init loss, overfit one batch, causality
  • Data pipeline: cleaning, filtering, exact and MinHash near-duplicate detection on Hugging Face text data
  • Trained its own BPE tokenizer (8k vocabulary) and packed the corpus into binary train/validation token files
  • Training loop: linear warmup + cosine decay, gradient accumulation (effective batch 128) and clipping, bf16/fp16 mixed precision
  • Resumable checkpoints (weights, optimizer, step) for Kaggle/Colab sessions; perplexity evaluation; temperature, top-k and top-p sampling
  • One YAML config names the whole experiment: architecture and training hyperparameters together
PythonPyTorchTransformers (from scratch)BPE TokenizerHugging Face DatasetsKaggle / Colab GPU
PRETRAINING IN PROGRESS
CARD 15Epic

Soccer Salary Predictor

Supervised ML · Regression · Sports Analytics

“Can xG predict a player's salary? Spoiler: yes, with SHAP to prove it.”

PWR85
SCL78
AUT70
INT90
PLH80
R² S88
  • Zero data leakage: split first, fit preprocessor on train only, no target imputation
  • Baseline (Linear/Ridge/Lasso/DT) → Improved (RF/XGBoost/LightGBM/GB) → Ensemble (Stacking + Voting)
  • 21 engineered features: xG/game, efficiency index, contribution index, with the target never touched
  • Optuna Bayesian tuning (60 trials, TPE sampler) on XGBoost
  • SHAP global beeswarm + local waterfall show which features drive salary
  • Live Gradio demo deployed on Render, kept alive 24/7 via UptimeRobot
  • Linked to SCIE publication: 17% accuracy improvement over baseline
PythonXGBoostLightGBMSHAPOptunaGradioRenderScikit-learn
CARD 16Epic

TacticalDNA

Unsupervised ML · Clustering · Sports Analytics

“No name attached, just 25 goals and 15 assists, and the model matched it to prime Messi at 98% similarity.”

PWR82
SCL80
AUT65
INT92
PLH82
TWIN98
  • 6,314 qualified player-seasons (≥900 min) across Bundesliga, La Liga, Serie A, with zero salary/target leakage
  • Algorithm bake-off: KMeans beats Agglomerative, DBSCAN, and Gaussian Mixture on Silhouette, Davies-Bouldin & Calinski-Harabasz
  • DBSCAN empirically collapses to one cluster, documented as a real density finding, not hidden
  • Rule-based archetype naming from centroids: Elite Complete Forward, Clinical Poacher, Deep-Lying Playmaker & more
  • Cosine-similarity engine finds each player's closest statistical twins within their own archetype
  • Interactive Streamlit demo: classify any real or hypothetical player, see radar chart + PCA map
PythonScikit-learnStreamlitPlotlyRenderUptimeRobot
Coming SoonRepository in progress
CARD 17Legendary

Computer Vision

Object Detection · Image Classification

“Teaching machines to see what humans see.”

PWR85
SCL70
AUT60
INT90
PLH72
  • Object detection with YOLO (v8/v11) and real-time inference
  • Image classification with CNN and transformer architectures
  • Custom dataset annotation and management via Roboflow
  • Transfer learning on domain-specific image datasets
  • Real-time video processing pipeline with OpenCV
PythonYOLOPyTorchRoboflowOpenCV
COMING SOON
Product Projects

Full-Stack AI Products

Three live products, built solo end-to-end: Palov (food ordering with a RAG menu chatbot), Hamyon (finance tracking with AI-generated insights), and Kotib (an AI secretary with Graph RAG and persistent memory).

CARD 18Rare

Palov

Full-Stack · RAG · AI Chatbot

“Halal Uzbek food in Korean cities, with an AI chatbot that actually knows the menu.”

PWR70
SCL65
AUT55
INT75
PLH82
  • Multi-country phone auth (+82 KR / +998 UZ) with OTP verification and Telegram OAuth
  • RAG system over the menu database: the chatbot answers dish queries with real-time recommendations
  • Self-improvement loop: user feedback and order patterns continuously refine suggestions
  • Full-stack food ordering platform connecting the Uzbek community in South Korea to halal cuisine
Next.jsRAGLLM APIVercel
CARD 19Rare

Hamyon

Personal Finance · RAG · AI Insights

“"Pulingiz nazoratda": your money, under an AI's watch, not just a spreadsheet's.”

PWR68
SCL62
AUT58
INT78
PLH80
  • Real-time expense tracking, budget monitoring, category-based spending analysis
  • RAG-powered AI chatbot over transaction history for natural-language finance queries
  • Self-improvement loop: learns from user corrections to refine categorisation & advice
  • AI-generated budget insights, not just static charts
Next.jsSupabaseRAGLLM APIVercel
CARD 20Epic

Kotib

AI Agent · Graph RAG · Conversational Memory

“A secretary that never forgets, and quietly gets better at guessing what you need next.”

PWR74
SCL68
AUT72
INT82
PLH78
  • Multi-step agent pipeline: intent classification → action routing → task/reminder/note execution
  • Graph RAG over the user's personal knowledge base for deep contextual understanding
  • Persistent conversational memory maintaining full context across all sessions
  • Self-improvement: tracks completed vs. missed tasks, adjusts proactive suggestions autonomously
Next.jsGraph RAGLLM APIVercel
Publications

Research Publications

Five first-author papers in SCIE journals, a KSCIE journal and an international conference, on machine learning for soccer analytics (salaries, performance, injuries) and mathematics for distributed systems.

01
Predicting Soccer Player Salaries with Both Traditional and Automated Machine Learning Approaches
SCIE Journal · Applied Sciences · 2025First author17% increased prediction accuracyRead paper ↗
02
Beyond xG: A Dual Prediction Model for Analyzing Player Performance Through Expected and Actual Goals in European Soccer Leagues
SCIE Journal · Applied Sciences · 2024First author12% improvement on player performance predictionRead paper ↗
03
Comparative Evaluation of Machine Learning Models for Predicting Soccer Injury Types
KSCIE Journal · 2024First authorNovel injury prediction model using constructed featuresRead paper ↗
04
The Application of Machine Learning on the Injury Prediction of Soccer Players
ISE@APSEC · CEUR-WS · 2022First authorInjury prediction without GPS technologyPDF ↗