Izmir, Turkey · San Francisco, CAProduct Engineer at Lovie

Emre Ozyoruk

AI Product Engineer · Agents, MCP & Fintech

Product engineer who takes products from zero to live — alone or with a team. I ship on every layer and every platform: Go and Python backends, web apps, iOS and Android apps in the stores, macOS and Windows desktop apps, and AI agents running in production. And I go below the API: I wrote a GPT from scratch and train and run models on my own hardware.

Shipped toApp StoreGoogle PlaymacOSWindowsWebGitHub Marketplacenpm

Download CV Email me

Live work

Everything here is running today

Case studies

Lovie

lovie.coSan Francisco fintech · current

I ship on every product — banking, cards, company formation, equity, fundraising and AI agents — from the Go backend to the iOS, Android, macOS and Windows apps.

Go · gRPC · Next.js · Swift · Electron · Stripe · MCP

Job Radar

jobradar.live · isradar.com.trBuilt and run alone

Crawls 21 job platforms every day into one index of 2.2M+ open jobs, matches them to each user's CV with vector search and emails the best ones. Live in the US and Turkey, with paid plans.

Python · Django · Celery · PostgreSQL + pgvector · Docker

Qalentis

qalentis.comBuilt from scratch · Tech Talent Space

An AI-native hiring platform: CV screening, generated technical tests and AI video interviews — then AI performance reviews after the hire.

React · Node.js · 18+ microservices · multi-LLM · RAG

hush

github.com/emreozyoruk/hushOpen source · GitHub Marketplace + npm

A bot that labels new GitHub issues — and stays quiet when it isn't sure. Tested on 800 real issues from 14 projects: 78% matched the label the maintainers chose.

TypeScript · GitHub Actions · zero dependencies

Experience

Where I ship

Full background

Product Engineer, AI-Native · Lovie

04/2026 – Present

San Francisco Bay Area (Remote) · every product, every surface: Go backend, web, iOS & Android, macOS & Windows apps, CLI, agents

Company formation — US incorporation automated end to end: state filings, IRS EIN by fax, DUNS via browser agents, e-signature.
Banking — card issuing and payments on Stripe Issuing, Treasury and Plaid; LLM transaction enrichment (categories, counterparties, recurring charges).
Equity & fundraising — cap table and SAFE rounds, plus Raise Agents that run investor outreach and act only after a human approves.
MCP server — 90+ production tools with guardrails, rate limits and audit logs; a gateway that turns any gRPC service into an MCP tool.
FreeTokens — one OpenAI-compatible key over every free LLM provider; shipped on Electron, then rebuilt natively in Swift.
Operations agents — an always-on SEO agent with multi-LLM voting and a budget kill-switch; Slack approval bots.

Full Stack Developer & Team Lead · Tech Talent Space

09/2025 – 07/2026

Izmir (Hybrid)

Led a 4-person engineering team and built Qalentis from scratch: 18+ microservices, multi-LLM (OpenAI, Claude, Gemini), RAG.
Secured and ran a TUBITAK R&D grant (Turkey's national research council), owning the technical documentation and compliance.

Full Stack & AI Developer · Skymod Technology

12/2024 – 09/2025

Izmir (Hybrid) · SkyStudio enterprise AI platform

Enterprise RAG applications on LangChain, LlamaIndex and Milvus; cut vector-search latency; shipped Node.js and FastAPI services on AWS.

Earlier

Freelance Full Stack Developer · FOK Dijital
2026
Business Analyst · Mobilpark
2024
.NET Bootcamp · Rea Teknoloji
2024
Software Support Specialist · ARKEN
2023 – 2024
Frontend Developer Intern · ASELSIS
2023

Below the API

I wrote a GPT from scratch

Reading about Mixture-of-Experts, LoRA and test-time training didn't make them click, so I built a language model instead — character-level tokenizer, hand-written attention with causal masking, a 4-layer Transformer trained on my own chat corpus.

Then the three ideas, one at a time: a 4-expert MoE with a top-2 router (loss 1.48 → 1.38, no expert collapse), LoRA adapters on 48 attention layers, and test-time training — which first made the model 36.8% worse, until a sweep of nine configurations showed where it actually pays off.

I train and serve models locally with PyTorch and Ollama, down to an LLM node on a Raspberry Pi.

838,775

parameters in a 4-layer Transformer with hand-written attention

97 s

to train it on a MacBook; loss 4.93 → 1.48

6.8%

of parameters trained with LoRA to turn it into a single voice

43%

lower loss on text it had never seen, through test-time training

Skills

Tools I ship with

AI: Agents, tool calling, MCP, human-in-the-loop, RAG, evals · Transformers, MoE, LoRA, PyTorch · Anthropic, OpenAI, Gemini · LangChain, pgvector, Ollama
Languages: TypeScript, JavaScript, Go, Python, Swift, C#, SQL
Frontend: React, Next.js, Angular, Tailwind · React Native, Expo · Swift (macOS) · Electron · Chrome extensions
Backend: Go, gRPC / Protobuf, Node.js, Express, Django, FastAPI, .NET (ASP.NET), Celery · Playwright, RPA
Data & infra: PostgreSQL, Redis, MongoDB, MSSQL, Supabase · Docker, Kubernetes, ArgoCD, Terraform, AWS, Cloudflare, Vercel
Fintech: Stripe Issuing & Treasury, Plaid, Persona KYC, ACH & wire
Education: B.Sc. Computer Engineering, Pamukkale University (2024)
Spoken: Turkish (native), English (professional)

Building something? Let's talk.

Open to product engineering and AI roles — remote or on site — and to freelance builds for teams that need a product shipped end to end.

emreozyorukdev@gmail.com Contact