I build AI systems that run in production, not just in demos.
Automation pipelines, Computer Vision, RAG and LLM integrations, Telegram/Discord bots, async Python services, and full-stack products with Next.js and TypeScript.
Currently delivering client projects with Indext Data Lab as a Top Rated Plus Upwork engineer. Previously built the full AI infrastructure of a live gaming studio and AI tooling for a business estimate platform.
Next.jsTypeScriptRAG SearchPostgreSQLPDF ExportCRM IntegrationPWA / Mobile UX
Full-stack work on a business platform that turns project data, work-day plans, materials, equipment, and AI-generated text
into client-facing cost estimates and A4 PDF presentations. Built an AI RAG search, structured pricing and cost logic,
a flexible timeline/day input system, unified UI components and notifications, and a safe AI estimate editor
that shows old/new diffs, applies only after confirmation, and supports rollback.
Delivered a manager-facing AI-assisted business tool: presentation generation, cost transparency,
versioned PDF delivery, CRM write-paths, and a PostgreSQL SSL fix for stable production. Live demo ↗
Internal automation for an agency to answer Upwork invites faster than competitors. Gmail/IMAP and Upwork API readers detect and
deduplicate invites, Telegram cards and a web dashboard route them for review, Claude/Gemini draft the proposal from the freelancer profile,
and a Chrome Extension browser agent claims approved tasks from a queue and executes them with audit logs.
Turned a manual workflow into a traceable event-driven pipeline: from a new invite to an approved, AI-drafted reply in minutes.
LIVE TRANSLATE: Realtime Translator & Interview Assistant
Open Source
PythonGemini APIEmbeddings RAGSpeech-to-TextSwift Audio CapturemacOS / Windows
Desktop app that listens to English speech in real time, shows the original text with a translation, and splits the conversation into
meaningful segments. In interview mode it drafts short context-aware answers using a custom prompt and a local
embeddings-based RAG knowledge base (CV, portfolio, projects) with similarity control. Captures microphone,
a selected window, or app audio on macOS, and system audio via WASAPI loopback on Windows.
A working realtime speech → translation → grounded answer pipeline. GitHub ↗
02 / Client & Freelance Projects
TCG Card-Replacement Automation
Python + OpenCV automation for a trading-card-game (TCG) card-replacement workflow, built with machine vision and OCR.
Client review: 5.0 ★, "taking an idea I didn't think was possible and making it possible", plus a bonus.
OpenCVOCRMachine VisionUpwork · 5.0 ★
Async Payments REST API
Asynchronous REST API with users, accounts, payments, and webhook processing. JWT auth, Alembic migrations,
Docker Compose, pytest test suite with CI, and a small web UI for manual checks.
Active Telegram bot for Armenian driving license exam preparation: practice tests, 1v1 duels with friends,
and AI-powered explanations. PostgreSQL stores users, progress, and question flows.
Claude Code skills and Python generators for client research: intake interview, market benchmarking, supplier and
equipment research, and auto-generated Excel / PDF / Markdown reports with version archiving.
Claude Code SkillsPythonExcel / PDFReport Automation
03 / Live Gaming Studio R&D · 2025–2026
As an R&D engineer I built the studio's AI infrastructure from scratch: agent orchestration, real-time Computer Vision,
hardware integrations, and a local RAG system. It ran in production 24/7 and made games 30–40% faster.
Dice Automation Detection
Real-time pip detection, trigger waiting, console feedback, and ROI-based reading for sequential dice automation.
Live Wheel Telemetry
Live wheel state monitoring with direction, speed trend, round validation, and final result telemetry.
Color Result Validation
Computer Vision result detection with confidence output and validation logic for ambiguous or incomplete game states.
Live Table ROI Tracking
Table-region monitoring with card state detection, named ROIs, and runtime event logs for production operations.
AGGREGATOR: AI Agent Orchestration
Asyncio service that manages the lifecycle of CV, hardware I/O, and classifier agents over an event-driven WebSocket API:
adaptive heartbeats, zombie detection with auto-restart, and a fallback to the last known good state. 30,000+ lines of code.
Python AsyncioWebSocketsSelf-HealingObservability
WHEEL SENSOR CONSENSUS: Gray-Code Decoding on Raspberry Pi
Noisy wheel sensors gave false results at high speed. Burst-sampled 10 GPIO pins, decoded Gray code, and locked the result
only on a statistical consensus. Round time −37%, no manual verification.
RPi.GPIOThreadingGray CodeStatistics
RACE VISION: Real-Time Finish-Line Tracking
RTSP camera stream → HSV masks → object centroids → signed distance to an operator-drawn finish line
(Ax + By + C) / √(A² + B²). Rankings broadcast via WebSockets with sub-100 ms latency.
OpenCVLinear AlgebraRTSPWebSockets
DICE AUTOMATION: Sequential State Machine
Reads 2–4 dice that settle at different times: a delay queue activates each die's ROI on schedule, counts pips with CV,
and broadcasts partial results. Handles re-shakes and TCP operator overrides without stopping the session.
PyQt5State MachinesArduino / SerialROI Detection
COLOR RESULT VALIDATION: CV + LLM Consensus
OpenCV classifies the winning color; Gemini acts as a second verification layer for ambiguous frames,
so a result is confirmed by two independent sources.
OpenCVGemini APIDual Consensus
Local RAG Knowledge Base
Offline RAG over internal docs: multi-version JSON chunking, TF-IDF index with cosine similarity, and a local LLM
via LM Studio behind a Tkinter UI. Sub-second retrieval, no cloud dependency.
LM Studioscikit-learnChunkingTkinter
OCR Signal Bridge → grandMA3
Tesseract watches selected video regions and triggers grandMA3 lighting changes when specific phrases appear, with a full text log.
Tesseract OCREvent TriggersgrandMA3
Ops Telegram Bot & Test Framework
Telegram bot for IT support: live stats, logs, and remote reboot/reset from a phone. Plus a pytest framework with
unit, integration, smoke, and async WebSocket tests.
I can help with projects, stack, experience, contacts, or a direct message to Harutyun.
Ask about Harutyun
AI portfolio assistant
Hi, I can answer questions about Harutyun's engineering experience, projects, stack, and contact details. Try asking what he can build or which projects are strongest.