02 / live surface
Self-BOT — operator surface
A real assistant surface riding the same VM runtime: provider routing, dual-context retrieval, and streamed progress on top of Meridian-backed context.
↗
TypeScriptQdrantRAGSSE
Live · Meridian + Self-BOT on GCP VM · UTC−3Backend systems
Backend systems
for agentic AI.
I'm Luis Adrian, a self-taught backend engineer in Buenos Aires building and operating my own agentic stack: Meridian for context transport and skill dynamic injection, and Self-BOT as an AI assistant.
About
A glimpse of how I work.
Operating the system is part of the design brief. I prioritise reliability, observability, and repeatability in the tools and processes I use to build and maintain software.
~/portraitBuenos Aires · UTC−3
Engineering profile
Backend systems
Self-taught. Linux-native. CLI-first.
Based in
BAS · UTC−3
Now playing
What I'm shipping.
- ▶Running Meridian on a GCP VM.Containerised context transport and skill injection at runtime with MCP and gRPC in the same live stack.
- ▶Keeping Self-BOT on that runtime.Provider routing, retrieval, and streamed responses share the same operating surface.
- ○Hardening deploy edges.Traefik, Docker Compose, and boring runbooks over novelty.
Primary stack
PythonFastAPITypeScriptgRPC · MCPQdrantPostgresLinux
“If it isn't running, it's still a hypothesis.”
// principle 01
Selected work
Systems that made it into runtime.
Open-source work, but not toy work. Meridian is the anchor runtime; the rest shows the surfaces and API discipline around it.
03 / API discipline
Inventory API — foundations
A smaller proving ground for contract discipline: auth, clean REST shapes, and the boring operational edges that make services trustworthy.
↗
PythonFlaskREST
Systems — live
What's running right now.
Meridian and Self-BOT are both running on the same GCP VM. This section is the operating view: runtime shape, uptime window, and the numbers I care about while it is live.
▣meridian.core· :7000
operational90 days ago30dtoday
Uptime 90d
99.94%
p99 frame
17.6ms
Active agents
6/6
⌬self-bot.api· :8443
operational90 days ago30dtoday
Uptime 90d
99.62%
p99 reply
1.41s
Requests 24h
4.2k
◊github / viper9009adr· contributions 52w
628 commitsMay 2025lessmoreMay 2026
Lab notes
Numbers from the system.
Short technical write-ups built from real operating questions: tail latency, retrieval quality, and provider routing.
benchmark2026-04-29~6 min
Cutting Meridian's p99 from 42ms to 17ms.
Six weeks tracking a single tail-latency regression across the merge stage. The culprit was a synchronous protobuf re-serialization sneaking in on fan-out.
Before
42.1ms
After
17.6ms
Reduction
-58%
Affected
100%
retrieval2026-04-14
Dual-context retrieval lifts recall@8 by 23pp.
Combining vector RAG with recursive CTE graph traversal on a 12k-doc corpus. The graph hop fills what cosine similarity misses.
vector only61%
+ graph CTE84%
routing2026-03-22
What Self-BOT actually called last month.
Provider routing log over March 2026. Local Ollama handles half; the heavy reasoning still goes to Claude.
ollama 51%claude 26%gpt-4o 15%other 8%
Interactive surfaces
Demos with their own routes.
The stack is a live system, and I want to show it as such. These demos link to interactive routes where you can explore the runtime and its components in more depth.