
Kinetiq
A crypto gaming platform with its own provably fair game engine and self-custodial stablecoin payments.
AI product engineer from Yerevan. I design, build and run products end to end: the screens, the backend, the models, the contracts and the server they live on.

I have been building for the web since 2020: first at FBF while still at university, then at Solicy from 2022. Since 2024 I lead a platform team there and work as a core engineer on Demos Network, a cross-chain network built inside Solicy.
I also run SFreedom, my company that builds for clients abroad and ships its own products. Every product on this page was designed, built and deployed by me, on infrastructure I set up and look after: one server, images built in CI, nightly backups and alerts on my phone.
SFreedom, my product studioCase studies and live system maps for each productsfreedom.netScreens are drawn and agreed before code, in a light and a dark theme, with real data instead of placeholders.
TypeScript and Python services, Solidity contracts, React and React Native apps, and the AI pieces in between.
Docker images from GitHub Actions, deploys to my own server, backups every night and alerts when something breaks.
Each one is live or in testing. Open any of them: they are real products with real data, not mockups.

A crypto gaming platform with its own provably fair game engine and self-custodial stablecoin payments.

Prediction markets on an on-chain order book, with smart accounts and gas paid for the trader.

A stylist and your own avatar that dress you in layers for the weather and your day.

Courier dispatch for Yerevan: orders, couriers and cash in one place, and a live map for the customer.

Armenian law in plain language. Answers stand on the official ARLIS texts, in Armenian, Russian and English.

A comfort smart home for Yerevan apartments: light, curtains and climate, priced for your flat in a constructor.

A self-custody wallet that checks every transfer for lookalike addresses before you sign.

A dark-fantasy pixel roguelite: five floors of a sealed crypt, on desktop, Android and in the browser.

A visual novel about a lighthouse keeper, with twelve endings, played in the browser.

Money, food and gym in one chat. Everything stays on the phone: no server and no account.

One help center for all my products: help pages, tickets and policies, with its own admin.
Not only API calls. I fine-tune small models, build the data to train them on, serve them locally and measure whether they got better.
LoRA fine-tuning of Qwen3.5 0.8B with PyTorch and PEFT on a local GPU, teaching it to pull structured preferences out of free text. Every run is scored on held-out data before and after training, and failed runs are kept, not hidden.
A synthetic dataset generator with dataset cards: 50,000 requests across 703 regions in 8 languages, plus run tracking and a live dashboard for every experiment.
GGUF models served with llama.cpp on CUDA and with Ollama. JSON Schemas are compiled into GBNF grammars, so a small model can only answer in valid JSON. One provider layer switches between local models and Claude, Gemini or OpenAI.
CLIP for zero-shot garment tagging, rembg for cut-outs and multilingual e5 for embeddings, all running on ONNX Runtime without a paid API.
RAG over pgvector with full-text search and rank fusion, checked against golden sets: 100 scenarios for the stylist and cited answers over Armenian law.
For clients: a diffusion-based virtual try-on on Stable Diffusion inpainting and image-to-3D characters with Hunyuan3D, both on a local GPU.
Six years of shipping for clients and teams, from small MVPs to platforms with real money moving through them.
Lif3 WalletA cross-platform DeFi wallet across 15+ networks, since closed
SwissCheeseA cross-chain DeFi trading platformWrite-ups of 21 systems I designed: the architecture, the trade-offs and the numbers. Client names are left out on purpose.
Provably fair iGaming platformGaming, real money
Decentralized prediction marketWeb3, trading
Cross-chain interoperability nodeBlockchain infrastructure
Legal research and compliance assistantAI, regulated domainThe tools I reach for first. I pick per problem, not per habit.
React, Next.js, React Native with Expo, TypeScript, PixiJS, Tailwind
Node.js with NestJS and Fastify, Python with FastAPI, WebSockets, BullMQ and RabbitMQ
PostgreSQL, Redis, pgvector, Redpanda, The Graph
Solidity with Foundry, viem, EVM networks, Solana, stablecoin payments
LoRA fine-tuning with PyTorch and PEFT, llama.cpp and Ollama, ONNX Runtime, LangGraph agents, RAG with pgvector, Claude, Gemini and OpenAI APIs
Docker, GitHub Actions, Coolify, Cloudflare, Vercel, backups and alerting
Every project starts with a short call about what you need.
Choosing the stack, designing the system, technical due diligence, or a second opinion before you build.
Web, mobile, backend and smart contracts, from the first screen to production and the server it runs on.
Turning a rushed or AI-generated MVP into software that holds up: structure, tests, security and deploys.
Tell me what it is and where it is stuck. A few lines are enough to start.