From binary sensor events to textual representations
A short introduction to my PhD research on using language-model representations to improve human activity recognition across environments.
RESEARCH & ENGINEERING
Research notes, engineering lessons, and ideas I’m still thinking through.
A short introduction to my PhD research on using language-model representations to improve human activity recognition across environments.
A guided map of Wassla: how feeds, language models, storage, search, a web reader, and observability fit into one reproducible system.
How Wassla schedules ingestion, parses RSS records, limits feed responses, removes known URLs, and handles failures at the feed boundary.
A close look at the outer fan-out and per-article LangGraph, its prompts and provider boundary, and what the quality score does—and does not—guarantee.
How the relational schema keeps source articles and language-specific versions together, and how PostgreSQL plus pgvector supports approximate semantic retrieval.
A step-by-step explanation of Wassla search: language-matched query embeddings, cosine retrieval, a score threshold, and an answer prompted with numbered sources.
How the Nuxt reader carries language and filters in the URL, renders translated records, and proxies only safe read methods to the pipeline.
How Wassla traces graph execution, evaluates a small golden set, tests service contracts offline, and where those checks stop short of production evidence.
A system map of the self-hosted chat stack: how the browser, model server, database, cache, metrics, and autoscalers work together under a tight compute budget.
How Open WebUI reaches vLLM, and why model size, context length, KV cache, CPU threads, and per-pod downloads determine the behavior of this deployment.
What PostgreSQL, pgvector, Redis, and ephemeral pod storage each do—and why scaling application pods is different from preserving their data.
Why the project scales vLLM from running and waiting requests with KEDA, while a separate CPU-based HPA manages Open WebUI.
How Prometheus collects model and Kubernetes metrics, how Grafana turns them into a dashboard, and what those measurements can and cannot explain.
A walkthrough of the deployment wrappers, generated secrets, Kubernetes Dashboard helper, and the boundaries that keep this learning stack from being production-ready.
What the phased CPU-cluster load test measured, why the model scaled while the UI did not, and which conclusions should not be generalized.