Explicit boundaries
Keep decision, execution, memory, risk, and resource cost separable enough to inspect and replace.
Building SelectiveLLM
I build AI systems that decide what computation is actually needed.
I'm Altan Çetin Çelik — an Electrical-Electronics Engineering student working across adaptive LLM inference, agent architecture, backend systems, and edge AI.
InputRequest / state
→ControllerRoute · budget · abstain
→RuntimeMinimum sufficient compute
Selected work
Systems, not isolated demos.
I like projects where architecture, runtime behavior, measurement, and failure modes are visible — not hidden behind a polished output.
01 / Research
SelectiveLLM
Adaptive model-capacity orchestration for language models. Select, load, cache, and budget adapters, experts, models, and future model blocks instead of treating all available capacity as permanently resident.
Current focus: learned + calibrated capacity routing
02 / Agents
seed-showcase
A dependency-free distillation of a local-first agent architecture with model routing, hybrid memory, approval gates, risk-aware capabilities, and recoverable self-editing.
03 / Edge
Industrial systems
Public engineering work around industrial telemetry, edge-side data handling, monitoring dashboards, backend services, and automation-oriented system design.
04 / Product
luna-companion
A stateful interactive web prototype exploring memory-driven UX, character interaction, responsive UI, and lightweight product architecture.
Research direction
Make the runtime choose.
The recurring question in my work is whether an AI system can spend resources selectively — choosing the right model capacity, memory, context, tool, or escalation path for the request in front of it.
“What is the minimum computation required to produce a useful, reliable answer under real memory and latency constraints?”
01Adaptive inference
02Conditional computation
03Model / expert / adapter routing
04Agent orchestration
05RAG & memory systems
06Backend infrastructure
07Edge AI
08Industrial telemetry
Engineering principles
Measure the system you claim to have built.
Real baselines
Compare against simple controls before treating a more complex mechanism as an improvement.
Failure is evidence
Preserve null and contradictory results instead of optimizing the story around a demo.
Human control
High-impact actions should expose approval, fallback, and abstention paths rather than assume autonomy is always desirable.
Altan Çetin ÇelikAI / ML systems
AboutElectrical engineering background.
Electrical engineering background.
Software-first execution.
My work spans ML systems, agents, backend infrastructure, and hardware-adjacent edge systems. I'm especially interested in the layer between a model's capability and the runtime decisions that make that capability practical.
LanguagesPython · TypeScript · C · C++ · SQL
AI / MLPyTorch · Transformers · PEFT · RAG · LLM routing
SystemsFastAPI · Pydantic · Docker · Linux · WebSockets
EdgeRaspberry Pi · STM32 · ESP32 · MQTT · Modbus · OPC UA