Production-grade software
for the AI era

Not every problem needs a generative model. For text classification, intent detection, entity extraction and similarity I use smaller, faster BERT-based models — a BERT classifier for intent, an embedding model for retrieval and an LLM for generation, working together to balance cost, latency and accuracy.

In fine-tuning I focus less on training the model and more on identifying the right training problem: RAG for fresh knowledge, fine-tuning for behavior and task discipline, tool calling for real-time actions — treated as complementary layers that solve different problems, not alternatives.

I serve open-source models at high concurrency with engines like vLLM, using continuous batching, PagedAttention and KV cache management to make GPUs efficient. I focus not just on the model running, but on time-to-first-token, tokens/sec, GPU memory, concurrent request capacity and cost per request.

Instead of using the largest model for everything, I build systems that route tasks to different models by complexity: simple classification on small models, complex reasoning on large ones. This routing lowers cost while improving latency and scalability.

Chunking documents into a vector database is never enough on its own. Based on a document's structure, headings, tables and semantic coherence I apply semantic chunking, recursive splitting, parent-child relationships and type-specific parsing strategies.

I evaluate embedding models not by specs alone but by retrieval success on real data. By measuring how highly the correct document ranks, I jointly optimize the model, chunking structure and index settings.

A fast embedding model retrieves a broad set first; a cross-encoder or BERT-based reranker re-orders them in a second stage. This significantly improves answer accuracy in enterprise systems full of near-duplicate documents.

Rather than giving a model unlimited authority, I explicitly define the allowed tools and operational boundaries. For critical actions — money transfers, record deletion, database updates — I build guardrails: user verification, authorization checks, parameter validation and explicit approval.

I use BERT and its variants for tasks needing low latency and high consistency, extracting person, date, amount, product and action type from a message into structured business data — e.g. "invoice five pens to Hüseyin" is parsed into customer, product, quantity and action type, then safely passed to the accounting service.

Instead of turning model output directly into an application action, I insert schema validation and business-rule checks in between. The data types, required values and allowed bounds of generated fields are verified before any action runs.

Rather than scattering prompts uncontrolled through application code, I treat them as a versioned, testable and measurable part of the system.

I assess quality not with a handful of sample questions but with test sets that represent real usage, comparing the impact of model, prompt, embedding, reranker and retrieval changes through measurable metrics.

In a production AI system, it must be traceable which model, with which prompt version, using which documents, called which tools. That traceability is critical both for debugging and for understanding how each answer was produced.

In enterprise and multi-tenant systems, to ensure one user can never reach another company's data I enforce access control not just at the prompt level but across the API, database, retrieval and authorization layers.

My goal is not just an LLM integration but designing the AI layer as a scalable, sustainable system that works alongside the rest of the application. In one project, BERT intent classification, Qdrant retrieval, cross-encoder reranking, an open-source model on vLLM and Python REST APIs can run together in the same architecture — with no lock-in to a single model or provider.