I build production infrastructure — and the AI that runs on it. Looking for DevOps and AI internships where reliability and cost per task are the metrics that matter.
Right now, on my rack95.2%cluster uptime20server TPSlive
Summary
Third-year CS student who runs a single-node Kubernetes cluster the way a team runs production: every deployment declared in Git, every service scraped and alerted on, and every cost measured rather than estimated. Most of my work sits where infrastructure meets applied AI — most recently reproducing an NVIDIA research result about replacing large models with small ones, using my own agent telemetry as the dataset.
Reproduced NVIDIA Research's “Small Language Models are the Future of Agentic AI” end to end, running their six-step LLM→SLM conversion procedure twice — on a classifier and on an agentic subtask — against my own logged agent traffic rather than a public benchmark.
Fine-tuned (QLoRA) a 3.2 GB model that beats a purpose-built 9.5 GB tool-tuned model 0.82 to 0.62 on a 61-task held-out set, at every interface and sampling configuration tested. Converting a second model family reached the identical score from a base 21 points apart, showing the result came from the pipeline rather than one checkpoint.
Ran a training-set size curve from 460 to 10,000 examples and found a plateau: 21× the data bought 3.3 points and regressed beyond 2,500. Ruled out compute starvation with a pre-registered epoch-matched arm that scored 6.5 points worse at 5.4× the compute — while validation loss read 0.000 on that worst model.
Gated every promotion on a rule fixed before each run, which rejected eight of my own proposals and forced five published numbers to be corrected by my own instruments; wrote up 25 failure modes, most of them defects in the measuring instruments rather than the models.
Built the router the pipeline feeds: hindsight replay over the telemetry ledger measured a 66% ceiling on locally-servable spend and a defensible 57% against a real 35B, with routing quality at 0.810 versus a 0.286 random-assignment floor — and published with its caveats, including an estimate that moved 45%→24% once a dominant session was sampled more than once.
Run a single-node k3s cluster as a production environment: GitOps delivery through ArgoCD app-of-apps, so no service is deployed by hand.
Instrumented every workload with Prometheus, Grafana and Loki, plus Istio mTLS between services and cert-manager for TLS issuance.
Secrets stay out of Git via External Secrets backed by Azure Key Vault; 7 alert rules across 5 services page a Discord incident bot before users notice.
Total infrastructure cost: $0/month, on hardware under my desk — with uptime published live rather than asserted.
Delivered full-stack observability — metrics, logs, distributed traces and AI incident analysis — for $5.12/month, about 60% under the serverless equivalent.
Added Tempo tracing and a local Phi-3 model that summarises incidents, keeping analysis on-premises instead of shipping logs to a vendor.
Built a native SwiftUI player from scratch with zero third-party dependencies, including a real 10-band parametric EQ on AVAudioEngine.
Backed it with a Python service that resolves and caches streams, and covered the app with 364 automated tests across 39 files, currently passing with 0 failures.
Wrote Azure Functions that probe homelab services every 15 minutes and process incident events off a Service Bus topic, with Pydantic-validated payloads.
Cut alert fatigue with a 5-minute dedup window that collapses repeat failures into a single Discord notification instead of a storm.