I am a data scientist and machine learning engineer based in Yerevan, Armenia, with hands-on experience building predictive models, Python and SQL pipelines, computer vision systems, and practical AI automations.
I am a data scientist and machine learning engineer based in Yerevan, Armenia, with hands-on experience building predictive models, Python and SQL pipelines, computer vision systems, and practical AI automations. I enjoy turning complex data and business requirements into reliable solutions, from regression and classification models to RAG assistants, Telegram bots, real-time market analyzers, and analytics dashboards.
Brief: a clean, fast storefront base to start shop and landing-page projects from. What I made: product grid, product pages with variants, cart, mobile layouts, hand-built HTML/CSS with no framework bloat, ready to connect to Shopify or a headless backend. Result: loads fast on mobile, easy to restyle for a client in a day; code at github.com/GEV44/aurel-storefront-demo. The base I use for Shopify stores and landing pages.
Brief: a Telegram bot that gives HIGHER / LOWER / NO TRADE signals on real forex pairs and never oversells. What I made: analysis of 11 pairs x 4 expiries (multi-timeframe trend, session, news calendar, volatility, spread, LightGBM), safety gates around rollover, high-impact news and spread spikes, a journal that grades every signal twice, a kill switch that disables any setup falling below break-even, multi-user access with risk disclosure, English and Russian. Result: 2 years of Dukascopy data, 18 months walk-forward, a sealed 3-month test, and the bot tells users plainly that no setup is statistically proven. ML with out-of-sample discipline, not hype.
Brief: watch new and moving coins around the clock, filter scams, and measure honestly whether any idea makes money after costs. What I made: a Python system that finds tokens on Solana, Base and BSC, runs a safety gate (mint/freeze authority, LP lock, top-10 holders, honeypot round trip, transfer tax), writes BUY / WATCH / AVOID cards with real win chances, sends SELL NOW and TAKE PROFIT alerts to Telegram, follows every token for 30 days, and paper-trades the Binance top 10 through an independent risk engine with hard limits. Result: 10 pre-registered hypotheses judged on forward data, about 460 automated tests, a dashboard and Telegram bot running 24/7 as a Windows service. No real-money trading: measured, not promised.
Brief: a web app where anyone can turn a PDF or DOCX resume into a scored, rewritten, exportable one. What I made: full-stack app, live with real users: auth, Postgres with row-level security, Edge Functions, Gemini for parsing and rewriting, scoring against 36 roles, 10 export templates. Result: live at resume-constructor-gev44.vercel.app, code at github.com/GEV44/resume-constructor. The same React + Supabase base I use for client web apps.
Brief: find the unmet user needs a product roadmap ignores, without hallucinated evidence. What I made: a six-stage Python/LLM pipeline that reads app-store reviews and a public roadmap, clusters the pain points, ranks opportunities and attaches a verbatim quote to every claim; tested with pytest and mypy. Result: 26 of 26 evidence quotes verified against the source, a judge-facing demo at gev44.github.io/silent-stakeholder-demo and the code at github.com/GEV44/silent-stakeholder.
Brief: put OpenAI/Claude into a product without surprise bills. What I made: a tool-calling assistant with answer caching, a hard monthly spend cap, request logging, and a small dashboard that shows latency and cache hit rate. Result: predictable cost per month, repeated questions answered from cache at zero API cost, and the team can see exactly what the model is spending. Stack: Python, OpenAI and Claude APIs, FastAPI, React.
Brief: data-collection and automation jobs that must run unattended and tell you when something breaks. What I made: a control room in Python for scheduled scrapers (requests, Playwright) with retry and back-off, a run log, alerts, and delivery to Google Sheets or CSV. Result: an example run delivered 1,247 rows from public sources on schedule without manual work. This is how I run scrapers, n8n/Make flows and integrations for clients so they keep working after handover.
Brief: a customer-support assistant that must never invent answers. What I made: a retrieval (RAG) assistant in Python with LangChain and the OpenAI API; every answer cites the document and page it came from, retrieval scores are visible, and when the documents don't cover a question it says "I don't know" and hands over to a human. Result: on a 60-question evaluation set, 54 correct, 5 correctly refused, 1 hallucination (98.3% groundedness). This is the architecture I use for support bots on websites, Telegram and WhatsApp.
Built and tuned more than five regression and classification models in Python and scikit-learn, improving core predictive performance by approximately 15%. Created preprocessing and feature-engineering pipelines across more than ten datasets, applying cross-validation, hyperparameter tuning, and error analysis to control overfitting and improve model reliability.
Performed statistical exploratory data analysis and SQL analysis, identified actionable trends and anomalies, built more than eight automated KPI dashboards, and streamlined recurring reporting workflows.
Build AI assistants and business automation workflows using Python, n8n, LLM APIs, REST APIs, webhooks, Google Sheets, SQL/SQLite, and Telegram integrations. Deliver RAG and chatbot flows, lead-capture and customer-support automations, structured outputs, conversational memory, testing, debugging, and client handoff. Use Claude Code and Cursor to accelerate implementation while validating final system behavior.
Post a job and invite Gevorg, or contact them directly. No hidden commission.
