BLOG
Engineering notes.
Technical notes that come out of the systems we build: local models and the serving layer, RAG architecture and access control, data residency, enterprise integration. Field notes, not marketing copy.
What we publish
We do not run this blog to fill a content calendar. We write about the problems we hit while building a system and could not find an answer to online.
Every published post comes out of a real deployment, and where there are numbers we describe how they were measured. No client names, no identifiable case details, in any post.
ARCHITECTURE
The local LLM serving layer: vLLM, SGLang, llama.cpp and Ollama
Once the model is chosen, the first architectural decision is which serving layer will run it. This note separates four layers not by feature lists but by their architectural approach to scheduling, memory, reuse and packaging. COMPARISONTurkish e-documents: direct integration or a special integrator?
Two ways to connect Türkiye’s electronic invoice, archive and waybill documents to your own systems, compared through where responsibility actually sits: who holds the financial seal and the timestamp, who archives, and who carries 7/24 operation. ARCHITECTUREData residency in Turkish banking and healthcare: two regimes, two questions
In healthcare the provision establishing residency never mentions geography; it points at the transfer regime. In banking the provision states geography outright — yet even there the real question is not the country but whether a component falls within the scope of primary systems. ARCHITECTURENES in the browser: not 60 frames, 60.0988
NES hardware produces 60.0988 frames per second; a browser draws at the display's own pace. The decisions we made between those two clocks while building Kaset: keeping the core swappable, moving rendering off the main thread, and treating the composite signal as the medium the games were designed for. GUIDEWhat to check for Turkish when picking a local LLM
A criteria list and measurement method for choosing an open-weight model to run on-premises with Turkish workloads — tokenizer fertility, quantization arithmetic, effective context length, instruction following and your own evaluation set. No leaderboard, no verdicts. GUIDEWhat an MCP server is, and what it is not
MCP standardizes the protocol layer between a language model and the systems it needs to reach. This note walks through the architecture, the server primitives, and where authorization actually happens — plus four things MCP is not. ARCHITECTUREAccess control in RAG: who gets to see what
When corporate documents are handed to a RAG system, the existing permission model usually does not travel with them. An implementation note on which layer enforces access, how identity reaches the query, and when an entitlement change actually reaches the index.Don’t wait for the post. Ask us directly.
No deck needed. 20 minutes. The rest is up to you.