Miraziz Seyidov
Junior AI / ML EngineerBaku, AzerbaijanOpen to remote
I build and ship LLM applications and applied ML models.
Second-year Computer Science student working in Python and JavaScript. Seeking intern or junior AI Engineer, ML Engineer or Data Scientist roles in Baku or remote.
About Me
Everything about Miraziz
Second-year Computer Science student who builds and ships LLM applications and applied ML models in Python and JavaScript, including a tool-using job-search agent with 151 automated tests, a live AI interview-practice platform, and a search-data ranking model validated on unseen clients.
Seeking intern or junior AI Engineer, ML Engineer or Data Scientist roles in Baku or remote.
- Baku, Azerbaijan · Open to remote
- mirazizseyidov1@gmail.com
Skills & Tech Stack
My tech toolbox
ML / Data
Model evaluation
LLM & AI Engineering
Techniques
Backend & Tools
Experience
Where I've worked
Machine Learning Intern — FlyRank AI
Structured internship programme in applied search intelligence on real, anonymised client data; research deliverable, not production code. Completion certificate, evaluation report and recommendation letter.
- Built a content-prioritisation model that ranks client pages by predicted visibility decline and outputs a review queue with reason codes (e.g. low CTR → update title/meta) and human-review rules (Python, pandas, scikit-learn).
- Data: 30,000 anonymised pages from 32 clients (Google Search Console and GA4 metrics). Test set: 6 held-out clients, 2,325 pages.
- Model comparison on the held-out clients:
Test set: 6 held-out clients, 2,325 pages Method precision@20 precision@50 precision@100 ROC AUC Rule baseline 0.20 0.30 0.35 0.557 Logistic Regression 0.35 0.40 0.44 0.700 Random Forest 0.70 0.68 0.70 0.747 - Found data leakage: a random split inflated precision@50 to 0.90; reported the honest client-grouped result (0.68).
- Queried a ~79M-row search-data warehouse on Hugging Face with DuckDB to build a 92,548-unit feature table from a 9.8M-row monthly partition.
AI Engineer Intern (training programme) — DevJoint
4-week programme; completed and passed verification on four practical projects.
-
LLM-Based Text SummarizerGitHub
LLM API application with prompt-engineering experiments.
-
RAG SummarizerGitHub
Retrieves information from documents and generates answers from their content.
-
Tool-Calling AgentGitHub
Interprets user intent and decides when to call external tools/functions.
-
AI Output EvaluationGitHub
Evaluated AI outputs and explored fine-tuning/adaptation approaches for a small model.
Featured Projects
Things I've built and shipped
AI Job-Search Agent
Tool-using job-search agent with a hand-written LLM loop, 8 job sources and 151 offline tests.
- Hand-written LLM tool-use loop (no agent framework) with 7 tools: parse query, choose source, search jobs, fetch job details, score match, draft cover letter, update application status. Each step logs duration, tokens and estimated cost; 15-step limit per run.
- 8 job sources: Remotive, Arbeitnow, Greenhouse, Lever and JSearch APIs, plus rate-limited, robots.txt-compliant scrapers for boss.az, hellojob.az and jobsearch.az, with cross-source deduplication.
- Two-stage CV-to-job scoring: rule-based pre-filter, then the top 15 jobs scored 0–100 by the LLM against an explicit rubric (skills 40, seniority 20, location 15, salary 15, language 10) with reasons and missing skills; final score computed in code.
- 151 offline pytest tests; backend and frontend deployed as separate Vercel services; LLM provider switchable between OpenAI and Anthropic.
Speechive.com — AI Interview-Practice Platform
Live SaaS used by real users on free monthly credits.
- CV-based question generation (GPT-4o-mini), text interviewer and answer scoring (GPT-4o), OpenAI speech-to-text and TTS (tts-1-hd), live voice interviews (Gemini Live API).
- ~2,500-line Node.js backend on Vercel serverless functions; Supabase Auth (email/password, Google OAuth) with JWT verification on every API request.
- PostgreSQL schema (8 tables) with Row Level Security (22 policies); Lemon Squeezy credit-based plans (Free, Plus, Pro) with signature-verified webhooks; 4 languages (az, en, ru, tr).
Legal Document Analysis System — Milli Majlis of Azerbaijan
System that checks draft laws against existing law.
- Contributed to a system that checks draft laws against existing law: parses DOC/DOCX/PDF/TXT into chapter/article/clause structure (regex rules with GPT-4o fallback), retrieves related articles by keyword and tag (simple RAG, no embeddings), and returns a 100-point score with risk levels and issues with legal basis (GPT-4o, JSON mode).
- Converted laws to Akoma Ntoso XML; one base law fully structured (18 chapters, 98 articles); tested on 5 laws; Markdown/DOCX report export.
No public link.
Meta Prompt Generator
Browser-only tool that turns a plain idea into a structured prompt.
- Browser-only tool that turns a plain idea into a structured prompt (Role / Context / Task / Instructions / Constraints / Output Format) for 7 target model families, or a meta-prompt the model completes itself.
- Detects domain, intent, technology and language; strips API keys and passwords; makes no API calls; 14/14 automated tests pass.
Education
Education & certificates
Bachelor's Degree in Computer Science
Academy of Public Administration under the President of the Republic of Azerbaijan, Baku
Data Science & Machine Learning Training (7 months)
Ingress Academy, Baku
Certificates
- Anthropic Academy — 20 courses: Model Context Protocol, prompt engineering, AI evaluation, tool use, context engineering, Claude Code, Claude with Amazon Bedrock (Jul 2026)
- Advanced Learning Algorithms — DeepLearning.AI / Stanford Online (Feb 2026)
- Supervised Machine Learning: Regression and Classification — DeepLearning.AI / Stanford Online (Jan 2026)
- Generative AI: Introduction and Applications — IBM (Dec 2025)
Languages
- Azerbaijani (native)
- Turkish (C1)
- English (B2)
Contact
Let's build something together
Seeking intern or junior AI Engineer, ML Engineer or Data Scientist roles in Baku or remote.