🤖 AI Systems Engineer · Hybrid AI architectures
🇹🇷 KVKK / GDPR-aware enterprise AI · Antalya, Türkiye → EMEA
🔧 Edge inference · local vision analysis ↔ multi-agent orchestration

Currently: Multi-domain orchestrator (stealth R&D) + AI NVR camera analytics
Looking for: Forward Deployed / Applied AI Engineer · EMEA remote · Türkiye on-site

About · Tech Stack · Projects · Beyond AI · Certifications · Writing · Stats · Connect

Computer Engineer focused on AI-powered systems — from edge inference on cameras to multi-agent orchestration over enterprise data. I design and build complete ecosystems: backend APIs, mobile apps, AI integrations, and the architecture that ties them together.

My background blends embedded systems & robotics with modern web/mobile and ML. That combination shapes how I think: hardware constraints first, software abstractions second, never the other way around.

Currently exploring how hybrid AI — local detection + cloud LLM semantics, or fully on-premise vision models — outperforms pure-cloud approaches on cost, latency, reliability, and data privacy for real-world deployments.

Focus areas: Hybrid AI architecture · Multi-agent orchestration · Vision LLMs · Prompt engineering & caching · Local inference · Event-driven systems · KVKK / GDPR-aware design

Multi-agent orchestration layer that gives users a single natural-language surface (chat, dashboards, proactive notifications) over many internal systems (ERP, fleet platform, QMS, asset registry, document stores). Designed for mid-size enterprises where data exists but isn't usable.

Architecture highlights:

• Four-layer separation: Channels · Orchestration · Domain agents · Federated data sources (no central warehouse, query-time API access)
• Capability-based model routing — domain reasoning on Claude Sonnet, horizontal/data tasks on Haiku, with fallback chains
• Multi-channel context continuity (a chat thread started in Teams continues in the web dashboard)
• KVKK / GDPR-aware auditability, prompt-injection defense, tool whitelisting per domain

Engineering rigor — every agent ships with a 12-axis checklist: prompt engineering & versioning · Anthropic prompt caching · structured outputs (Zod) · typed error taxonomy · streaming with progressive tool-call UI · OpenTelemetry observability · per-user / per-agent / per-tool rate limits (Upstash Redis) · PII detect+mask pipeline (TC kimlik, phone — deterministic + Haiku) · golden-query evals (Vitest + Promptfoo) · semver agent versioning · per-call cost log + Vercel AI Gateway budget caps · LLM hallucination guard via citation cross-check.

Stack: Next.js · TypeScript · Anthropic SDK · Vercel AI SDK + AI Gateway · Supabase (Auth + Postgres) · Upstash Redis · Sentry · Promptfoo · Vitest

Framed not as "another chatbot" but as a data strategy project — orchestration becomes the forcing function that disciplines underlying data sources and turns latent data into used data.

Implementation repository private during development. Engineering deep-dive available on request.

📖 Architecture vision (public, CC BY-SA 4.0): enterprise-ai-orchestra · live interactive demo

Production-targeted reference architecture for adding AI on top of an existing CCTV/NVR environment without disturbing the original recording system. Combines local detection with local vision analysis for event enrichment — designed to keep image data on-site by default while preserving flexibility for external LLM/API providers when needed.

Current status (M7 in flight · M8 design):

• ✅ Core pipeline running in dev-stack — 5 services healthy, end-to-end verified with test streams
• ✅ Zone state machine, first-entry + restricted-zone alarm, snapshot capture
• ✅ Local vision analysis — truck cab + trailer color/type, per-call latency logged
• ✅ NVR external-alarm bridge code with retry queue; field validation pending
• ✅ Door events — ms-precision in/out log + DMSS mobile push
• ✅ Camera-offline detection → alarm + Grafana panel · operational runbook
• 🚧 M7: operational maturity — backup/disk alarms, restart auto-tests, 6h/24h soak path
• 🔬 M8: forensic behavior intelligence — design/spec stage, not shipped code
• 140 unit tests · ruff · mypy strict · Alembic migrations · GitHub Actions CI
• 11 architecture documents — tech decisions to bottleneck analysis

Why interesting: real cost modeling, explicit trade-off tables, NVR-load-aware design, privacy-first architecture, and a clear production-readiness path from unit tests to soak tests and field pilot.

Not just the model — I build and ship across the full stack: TypeScript · Next.js · Node backends, React Native mobile apps, and Postgres / Supabase data layers. The AI systems above sit on a real product-engineering foundation, not a notebook.

Building toward CKA → AZ-104 → CKS → AZ-500 through 2026-2027. Credly badges will appear here as earned.

• CKA — Certified Kubernetes Administrator · Linux Foundation · target M6-7
• AZ-104 — Microsoft Azure Administrator · target M9
• CKS — Certified Kubernetes Security Specialist · target M16
• AZ-500 — Microsoft Azure Security Engineer · target M18

Technical writing pipeline. Posts will auto-populate here as published.

Planned topics:

• Local-First Vision LLMs: Real Cost Math from Production-Targeted Camera Validation
• Hybrid AI Decision Matrix: When Edge, When Cloud, When Both
• KVKK/GDPR-aware Multi-Agent Architecture: A Practical Reference
• Boundary Agent Doctrine: A Framework for AI Tool Calling at Production Scale

• Lead engineer on the AI Orchestration Platform above — moving from architecture deck to multi-agent runtime: prompt versioning, eval pipelines (Promptfoo + Vitest golden queries), Anthropic prompt caching strategy, KVKK PII pipeline, per-agent cost attribution.
• AI NVR M7 in flight — operational maturity: camera-offline detection, backup/disk alarms, runbook, soak-test path, and field-pilot readiness.
• Exploring capability-based model routing to keep multi-agent systems reliable, auditable, and cost-aware under tight monthly LLM budgets.

📧 Email — ibrahimsumbulll@gmail.com 💼 LinkedIn — linkedin.com/in/ibrahim-sümbül 💻 GitHub — github.com/ibrahimSumbul

"Hardware constraints first, software abstractions second."
_{From embedded systems to multi-agent AI — building across stacks, languages, and boundaries.}