How DeepSeek's Super-agent works
DeepSeek wants us to stop watching every session. I went through its prototype and its experiments to understand how it plans to do that.
Architecture
12 posts
Serving agents with vLLM on your own GPU
A Claude Code session is about 174 requests of 100,000 tokens each. What that does to a GPU, why with agents compute runs out before memory, and the same maths on a 16 GB card.
DeepInfer explained: how DeepSeek's inference engine works
DeepSeek open-sourced DeepInfer, the engine it serves V4.1 Flash with. How it works, what it optimises and how to read its benchmarks.
How and Where the Information You Give AI Travels
The prompt you type is only a fraction of what an AI system sends to the model. What actually travels, where it crosses, and what stays on the other side.
AI Sovereignty: Who Controls Your Company's Intelligence?
Knowing where your intelligence runs, what happens to your data and whether you can still change those decisions. A framework for AI sovereignty.
DeepSeek V4.1 Flash and the 890 bytes that matter for agents
Live on Helmcode, replacing V4 Flash at no extra cost. It redesigns inference for a workload that keeps growing: agents that read far more than they write.
Build your own Granola with open models
Record and transcribe your meetings with open models, turn them into structured data, and build a searchable memory of every conversation you have had.
Does building your own infrastructure to run open models pay off? We did the maths
An 8x H200 node serving GLM-5.2 costs €25,250 a month before it serves a single token. The full numbers, the break-even, and when owning the hardware wins.
Cordis: the paper behind the DeepSeek Harness architecture
The five theorems in the Cordis paper, the plugin framework behind DeepSeek Harness, explained in plain language and with what already runs in production.
GLM 5.2 orchestrates, DeepSeek V4 Flash executes
How to design multi-model systems where one open LLM plans, another executes, and the context they hand each other stays efficient.
Which GPU do you need to serve AI models
FP8, NVFP4, H200, B200: what each NVIDIA generation accelerates in hardware, how much VRAM your model needs, and which GPU to buy. With real production data.
Moving Down from Serverless to Kubernetes
Before we start, a bit of context. The infrastructure is hosted on AWS and the architecture was based on Serverless services: