# Tarık Deveci — AI Product Engineer

> Software Engineering graduate and AI Product Engineer based in İzmir, open to İstanbul
> and remote. Builds
> intelligent systems from idea to production: LLM pipelines, decision engines,
> full-stack products, mobile apps, and sustainability/safety tech.

## Summary
Tarık Deveci designs the systems around AI, not just API calls. Work spans municipal
AI (MuniGo), productivity systems (LifeOS), explainable decision intelligence (Octostra),
industrial process safety (PreventA), and DetectIQ, a shipped SAM3 + LoRA damage-segmentation thesis.

## Selected work
- MuniGo — AI municipal call-center & CRM. Role: Co-founder & Lead Developer (2024–2025).
  Stack: Django, PostgreSQL, GPT API, AWS Polly. Impact: TTS cost −80%; latency 7s → 1.5s;
  live across 2 municipalities, 20+ departments, 300+ complaint subcategories.
- LifeOS — AI productivity OS for tasks, time blocking, nutrition, workouts. Web + mobile.
  Stack: Next.js, React Native, Expo, Supabase, Zustand, RevenueCat, iyzico. 10K+ tasks,
  7 AI features, 2 platforms. https://lifeos.tr/
- Octostra — explainable decision-intelligence engine. Deterministic rule engine + transparent
  scoring + aggregation; surfaces top 3–5 tasks with traceable reasoning. Stack: FastAPI,
  PostgreSQL, React, Docker. https://github.com/tarikdeveci/octostra
- PreventA — process-safety platform on HAZOP/LOPA. Stack: FastAPI, PostgreSQL, React.
- DebateAI — AI-powered debate / structured argumentation product.
- Detay İnovasyon — carbon-footprint tracking platform; 1,000+ monthly emission records. React, Node.js.

## Experience
- 2025 — Software Engineering Intern, Hepsijet. Java Spring Boot, Docker, Jenkins, Datadog. ~25% API response improvement.
- 2024–2025 — Co-Founder & Lead Software Developer, MuniGo.
- 2024 — Junior Full-Stack Developer, AtıkNakit. Node.js, PostgreSQL, Flutter.
- 2022–2023 — Software Developer, Detay İnovasyon.

## Education
BSc Software Engineering, Bahçeşehir University (graduated 2026), GPA 3.2/4.0.
Capstone: DetectIQ (SAM3 + LoRA structural damage segmentation).

## Research
DetectIQ — fine-tuning Meta's SAM3 segmentation foundation model with LoRA for bridge/structural
damage segmentation on the DACL10K dataset (capstone thesis, 2026), then shipped to production.
Trained ~0.25% of 842M parameters (LoRA on q_proj/v_proj), 10 epochs on an A100, BCE+Dice loss.
Mean validation IoU 0.56 across 19 classes (peak class 0.81; 8 of 19 above 0.60). Served via a
FastAPI + React inspection platform (upload image → colored damage mask + per-class coverage),
with a swappable mock/real detector seam. Live: https://detectiq.com.tr

## Contact
- Email: tariikdevecii@gmail.com
- GitHub: https://github.com/tarikdeveci
- LinkedIn: https://www.linkedin.com/in/tarik-deveci