Skip to content

kobe young · software · systems · inference

Building the behind-the-scenes systems that move data and run AI models, the parts people never see but use every day. Concretely: real-time data pipelines, RAG/LLM serving, and distributed backends on AWS and Azure. Former Navy avionics technician and residential electrician, now working my way down the stack toward systems and inference.

// self-hosting the whole stack on one box · owner of multiple cats

featured · ts-llm-gateway.vercel.app

ts-llm-gateway

TypeScript · Vercel AI SDK · AWS Bedrock + OpenAI · Hono · Upstash Redis

A production-style, multi-provider LLM gateway in strict TypeScript: one unified API plus an OpenAI-compatible surface (a drop-in baseURL swap) that routes to AWS Bedrock and OpenAI behind a single injected provider interface, wrapped in real production policies: per-key/IP rate limiting, retry with backoff and cross-provider failover, a circuit breaker that quarantines a downed provider and probes it back, per-call timeouts, an LRU response cache, and SSE streaming that aborts the upstream call when the client disconnects.

The point: it's the same class of system I built at work, a real-time streaming pipeline, re-expressed as an LLM proxy. Kinesis backpressure became rate limiting; retry/backoff became failover; quarantining a bad dependency became the circuit breaker; shard-consumer cancellation became streaming abort; composite routing keys became provider routing; SignalR fan-out became SSE.

Unified + OpenAI-compatible API

one native endpoint plus a drop-in /v1/chat/completions

Rate limiting

per-key & per-IP token buckets, constant-time key checks

Failover & retry

exponential backoff + jitter across providers

Circuit breaker

quarantines a downed provider, half-open probes it back

Streaming with abort

SSE tokens; client disconnect cancels the upstream call

Cache & state

LRU response cache; in-memory or Upstash Redis, degrades gracefully

~6×
cache-hit speedup (0.22s vs 1.27s)
87
mock-injected tests, keyless in CI
2
providers behind one interface

// ~6× cache-hit speedup and cross-provider failover both verified on the live deployment

also featured · civic-lens.info

Civic Lens

Go · Python · SQLite · FastAPI · React · LLM pipeline

An audit-driven system that measures sampled political discourse across news, Reddit, and X: a polite, crash-safe Go crawler, an LLM analysis pipeline where every output carries a confidence score and a verbatim evidence span, narrative clustering, and a CI regression gate for prompt changes. Open source, deployed, and documented down to the invariants.

~/live

kobeyoung.net live metrics dashboard

dashboard offline
kobellm · offline
model
-
self-hosted
throughput
-
tok/s · last
first token
-
ms · last
responses
-
since deploy
today
-
daily budget
host · hetzner box
cpu
-
utilization
memory
-
used
load
-
1-min avg
machine up
-
host uptime
site · since last deploy
api requests
-
served
abuse deflected
-
honeypot + limits
messages
-
contact form
content
8
7,976 words written
deployed
443b40f
11d ago
api up
-
process uptime
Kobe Young · Software & Systems Engineer