Welcome
What I work on, what is on this site, and where to find the code. LLM systems, agents, and inference.
I'm Divith. I work on LLM systems: training and evaluating models that call tools, serving them, and running agents that can fail and recover without losing their place.
Most of what I've built recently falls into three areas. I built a multi-model vLLM serving control plane in Go and Python that routes requests by tenant quota, replica load, and KV-cache state. I built a durable runtime for MCP and computer-use agents with event-sourced state, checkpoints, and crash recovery. And I worked on tool-use post-training and evaluation with PyTorch and QLoRA, measuring schema validity, tool selection, and argument accuracy rather than vibes.
By day I'm a software engineer at CHN Technologies, working on Django and NestJS services, PostgreSQL schemas, and Docker CI/CD. I also run OneDot Communications, where the work ranges from AI products to custom AOSP ROMs.
What's on this site
The experience and projects sections cover the work in more detail, including the mobile projects in Kotlin and Swift. Resume has four versions of the same history, each ordered for a different kind of role. The rest of this blog is occasional writing about things I've built or got wrong.
Everything here is open source. The code for this site is on GitHub, and so is most of what I build.