Davronbek Malikov
📍 Seoul, South Korea
AI engineer with a PhD in AI Engineering, working across the full stack: classical machine learning, deep learning, computer vision and LLM systems. I take models from data to production: data pipelines, feature engineering, training and tuning, explainability, evaluation, deployment and monitoring. On the LLM side I build post-training pipelines (SFT, DPO, RLHF), hybrid RAG and multi-agent systems; on the ML side, from published tabular ML research to a GPT-style Transformer written from scratch in PyTorch. Currently leading the AI team at TripleH, shipping agents for legal research, financial analysis, meeting automation, and per-employee AI assistants. 5 first-author publications in SCIE journals and international conferences.
- Lead the AI engineering team building autonomous agent systems and enterprise AI products.
- Architect and ship production-grade AI agents for enterprise clients: multi-agent orchestration, legal research, meeting intelligence, and per-employee assistants.
- Drive end-to-end development from research and prototyping to deployment on cloud and on-prem infrastructure.
- Research on machine learning, deep learning and LLMs applied to soccer analytics.
- Models for performance evaluation, salary prediction, and tactical analysis.
- Preparing submissions to SCIE and AI-focused journals.
- Design and deliver advanced courses in machine learning, deep learning, computer vision and LLMs.
- Mentor students and professionals on building end-to-end AI systems.
- Focus on practical applications of AI in industry and academia.
- ML for soccer analytics, covering salary prediction with traditional ML and AutoML, a dual expected/actual-goals model, and injury-type prediction.
- Published 4 first-author papers in SCIE and international conference proceedings.
- Applied advanced mathematical concepts to distributed systems.
- Published on lattice-group structures in Cartesian product spaces.
- Taught advanced mathematics and analytical reasoning for university entrance exams, the SAT, and other international mathematics exams.
- Developed customized teaching materials to strengthen students' problem-solving and logical thinking skills.
Thesis: Advancements in Soccer Analytics: Players' Goals and Salary Prediction Models with Machine Learning
- A full learning platform (course → module → lesson → subskill, progress tracking, onboarding, admin CMS) with one shared AI agent engine behind three assistants: Student Tutor, Teacher Assistant and Admin Assistant.
- Provider adapters for Claude (official Anthropic SDK), OpenAI and local Ollama models, with SSE streaming, usage and cost limits, and safe error handling.
- RAG knowledge base: uploaded PDF, DOCX, TXT and Markdown are chunked, embedded and stored in pgvector; the permission check runs inside the retrieval query, and answers cite their sources.
- Security-first: prompt-injection defence, server-only secrets, signed private storage; 206 automated tests across authorization, HTTP security, the AI engine and RAG, all passing.
- An evidence-first football intelligence platform for coaches, scouts and analysts: answers come with sources and an explicit statement of what the evidence can and cannot support.
- Phase 1 built: a hand-written document indexing pipeline covering loading (PDF per page, DOCX, TXT, Markdown), metadata, cleaning, token-based chunking, Sentence-Transformers embeddings and idempotent storage in Qdrant.
- Stable UUID5 IDs so re-ingestion never duplicates; metadata such as match, team and player IDs is never guessed; unit and integration tests for every step.
- Planned: hybrid retrieval with RRF and reranking, a SQL analytics engine, an investigation agent, a FastAPI and Next.js app, and video intelligence with YOLO tracking.
- A scout describes the player they need in plain words; a Claude agent orchestrated with LangGraph turns the brief into typed tool calls over a DuckDB warehouse of 50,149 players; the model never writes SQL.
- Joined Transfermarkt market data with 297 StatsBomb matches (18 per-90 metrics and percentiles for 594 players), linking 87.5% of players across sources.
- Human-in-the-loop shortlist approval via LangGraph interrupt() with persisted checkpoints; cited scouting reports built on Qdrant hybrid search over scout notes and Wikipedia.
- 120-question golden set in LangSmith with automatic checks for filter accuracy, precision/recall@10, numeric grounding and invented players; 102 automated tests.
- Took a pretrained Qwen2.5 base model through the full post-training pipeline (SFT, DPO, RLHF, deployment) and measured all 14 models on the same 68-question held-out test, working from a written model spec.
- SFT took Qwen2.5-1.5B from 15/68 to 44/68 (LoRA vs QLoRA compared under identical settings). DPO implemented from scratch: v1 learned "shorter is better" (40/68); v2 fixed it with length-balanced pairs and an SFT loss term (44/68, final model).
- Classic RLHF with a reward model and PPO, then RLOO, across four judges (own 0.5B and 1.5B reward models, Skywork-Reward-V2, Gemini as a live judge), with reward-hacking and length-bias checks. RLHF reached 39–41/68: in this setting it was only as good as its judge, and DPO was simpler and better.
- Deployed as TeachHelp: a Claude-style chat in English and Uzbek on Vercel, served from a Kaggle GPU over a Cloudflare tunnel (~2–3 s per answer), plus a fully local version with all models.
- A multi-agent job hub for Uzbekistan, Korea and worldwide (including Telegram job channels), in Uzbek, Russian, Korean and English: job matching, skill-gap roadmaps and company-specific mock interviews.
- A LangGraph orchestrator plans each conversation and routes tasks to 11 agents over Kafka; agents call each other over A2A (Agent Cards + JSON-RPC). A Judge agent checks every output and sends weak work back for a retry.
- Job pipeline from Remotive, Saramin, hh.uz and Telegram: normalisation (so'm / ₩ / $ salary parsing, dedup), a multilingual scam filter, and Postgres + pgvector search with an explainable match score.
- Runs with Claude, Gemini or Ollama, with deterministic fallbacks when no key is set; Jaeger traces across agents, unit and end-to-end tests, and evals for scam detection and answer scoring.
- Implemented a decoder-only Transformer from first principles in PyTorch (multi-head causal self-attention, MLP blocks and weight tying), verified with shape, random-init loss, overfit-one-batch and causality tests.
- Built the full pretraining pipeline: cleaning, filtering, exact and MinHash near-duplicate detection, a trained BPE tokenizer and binary token datasets.
- Training loop with linear warmup + cosine decay, gradient accumulation and clipping, bf16/fp16 mixed precision and resumable checkpoints; perplexity evaluation and temperature/top-k/top-p sampling.
- Industry-level ML pipeline: Baseline → Improved (RF, XGBoost, LightGBM) → Ensemble (Stacking + Voting).
- 21 engineered features without data leakage. Optuna Bayesian tuning (60 trials). SHAP global + local explanations.
- Deployed on Render · kept alive 24/7 via UptimeRobot. Linked to SCIE publication: 17% accuracy improvement.
- A blind stat line (25 goals, 15 assists, no name attached) matched to prime Messi's 2019/20 season at 98% similarity.
- Algorithm bake-off (KMeans, Agglomerative, DBSCAN, Gaussian Mixture) over 6,314 qualified player-seasons, scored on Silhouette, Davies-Bouldin & Calinski-Harabasz.
- Rule-based archetype naming from cluster centroids: Elite Complete Forward, Clinical Poacher, Deep-Lying Playmaker & more.
- Cosine-similarity engine + interactive Streamlit demo: classify any real or hypothetical player and find their statistical twins.
- Halal Uzbek food platform for South Korea with multi-country OTP auth (+82 KR / +998 UZ).
- RAG system over menu database: the AI chatbot answers dish queries with context-aware recommendations.
- Self-improvement loop: user feedback feeds back into the model continuously.
- Real-time expense tracking with category-based spending analysis.
- RAG-powered AI chatbot over transaction history for natural-language finance queries.
- Self-improvement: learns from user corrections to refine categorisation and advice quality.
- AI-powered personal secretary with multi-step agent pipeline: intent → routing → action.
- Graph RAG over personal knowledge base. Persistent conversational memory across all sessions.
- Self-improvement: tracks completed vs. missed tasks, adjusts proactive suggestions autonomously.
Open to new opportunities
LLM · RAG · Agentic AI · ML & Deep Learning · Computer Vision