Detay İnovasyon
SustainabilityA carbon-footprint tracking platform and digital presence for an environmental consultancy, processing 1,000+ emission records every month.
React · Node.js
AI Product Engineer İzmir · open to İstanbul
I build systems that perceive, decide, and ship.
I design LLM pipelines, decision engines, full-stack products and mobile apps, then take them to production. Municipal AI, productivity systems, industrial safety, sustainability tech: real domains, measured results.
About
I'm Tarık — an AI product engineer who actually ships. I don't stop at a notebook or a slide; I carry an idea through the model, the backend, the interface and the deploy until it's live and someone is using it. Fresh out of Software Engineering, I've already pushed a municipal AI call-center, a productivity OS, an explainable decision engine, a process-safety platform and my SAM3 thesis into production. Based in İzmir, open to İstanbul and remote — looking for the team to build the next one with.
- İzmir, TR
- Software Engineering grad
- 7 shipped · 6 live
- Open to İstanbul & remote
Selected work
Four products, each shipped into a real domain with measured impact.
-
01Decision IntelligenceOctostra
A decision-intelligence engine that tells you what to focus on now, and shows its reasoning. Not another task manager.
- Explainable, no LLM black box
- Top 3–5 with reasoning
-
02AI Productivity · Mobile + WebLifeOS
A personal operating system for tasks, time blocking, nutrition and workouts, with AI daily planning synced across web and mobile.
- 2 platforms, one sync layer
- 7 AI features
- Subscriptions live
-
03Multi-Agent StrategyDebateAI
A market-entry strategy engine where 100+ specialized agents debate options, red-team assumptions, score surviving plans and assemble a board-ready playbook.
- Structured strategy intake
- Visible agent workflow
-
04Process Safety · Hybrid RAGPreventA
A process-safety platform that turns HAZOP / LOPA methodology into structured, auditable risk software, with a hybrid-RAG layer grounded in plant documentation.
- Structured HAZOP / LOPA worksheets
- Grounded suggestions, with citations
More projects
Autonomous multi-agent consulting, municipal AI, sustainability, embodied AI, and open source.
MuniGo
Municipal AIAn AI municipal call-center and CRM, co-founded and built end to end: it routed citizen complaints, tracked cases, and reported KPIs for two real municipalities across 300+ complaint categories.
Reworking the request path and model calls cut spoken-response latency from ~7s to ~1.5s and TTS cost by ~80%.
Django · PostgreSQL · GPT API · AWS Polly
Ovri
Embodied AI · In developmentA small physical companion robot paired with a local-only reasoning core that builds memory and personality over time, designed to integrate with LifeOS and Octostra. Currently in product design and early development.
Local LLM · Memory engine · Robotics
Open source
GitHubOctostra, Carbonfootprint, TicketingSystem and more, public on GitHub.
Browse repositoriesThe path
Where I studied, and the roles that each left a measurable mark.
-
2026
BSc Software Engineering · Bahçeşehir University
Graduated 2026. Capstone: DetectIQ, a SAM3 + LoRA fine-tune for structural damage segmentation, taken to production.
3.2 / 4.0 GPA -
2025
Software Engineering Intern · Hepsijet
Backend work on microservices: Java Spring Boot, Docker, Jenkins, Datadog.
~25% API response improvement -
2024–25
Co-Founder & Lead Software Developer · MuniGo
Built an AI municipal CRM end to end: Django, PostgreSQL, GPT API, AWS Polly.
−80% TTS cost · 7s → 1.5s latency · 2 municipalities -
2024
Junior Full-Stack Developer · AtıkNakit
Web and mobile delivery with Node.js, PostgreSQL and Flutter.
-
2022–23
Software Developer · Detay İnovasyon
Carbon-footprint tracking platform and web presence with React and Node.js.
1,000+ monthly emission records
RESEARCH · SHIPPED TO PRODUCTION
DetectIQ: a SAM3 fine-tune for structural damage, from thesis to a live product.
My capstone fine-tunes Meta's SAM3 segmentation foundation model with LoRA to segment bridge and infrastructure damage — cracks, corrosion, spalling and 16 other classes — at the pixel level, on the DACL10K dataset. I trained just 0.25% of its 842M parameters, then shipped the model behind a FastAPI + React inspection platform: upload a photo, get back a colored damage mask with per-class coverage.
0.56mean validation IoU across 19 classes
0.81peak class IoU · 8 of 19 above 0.60
0.25%of 842M params fine-tuned (LoRA)
~10KDACL10K images · 975 validation
- SAM3 base, LoRA adapters on the attention projections (q_proj / v_proj), ~8MB adapter
- 10 epochs on an A100 with a combined BCE + Dice loss; best validation loss 0.27
- Served through a swappable detector seam — mock or the real model via a one-line config flag
Tech stack
What I reach for, grouped by where it lives.
Product & Web
- Next.js
- React
- TypeScript
- React Native · Expo
- Tailwind
- Zustand
Backend & Data
- FastAPI
- Django
- Java Spring Boot
- Node.js
- PostgreSQL
- Supabase
AI & ML
- LLM pipelines
- Hybrid RAG
- LoRA / PEFT fine-tuning
- SAM3 / segmentation
- GPT API
- Decision engines
Infra & Ops
- Docker
- Jenkins
- Datadog
- AWS
- Vercel
- RevenueCat · iyzico
open to AI / product engineering roles & collaborations