PREFILL
Prompt ingestion
The gateway measures time-to-first-token and input token count. This is the prompt-processing phase, not hidden reasoning.
Build explainer · 2026
A personal reasoning system, made observable.
Aurorium Mind is a QLoRA-adapted Qwen model trained on a 100k synthetic SFT set, wrapped in a streaming Rust/FastAPI service, and surfaced through a Vercel workspace. This page records what was built, what was measured, what failed, and what the evidence can—and cannot—prove.
Project snapshot complete · server can be safely stopped
What shipped
The smallest complete system: one GPU, one adapter, one protected API, and a frontend that exposes useful runtime facts without exposing private traces.
01 · Data100k procedural synthetic examples plus a disjoint 2k eval set.
02 · TrainOne-epoch QLoRA/SFT on a single A10G.
03 · ServevLLM loads Qwen3.5-4B and the `aurorium` LoRA.
04 · GatewayRust and FastAPI stream responses and metadata-only telemetry.
05 · ExperienceNext.js/Vercel chat UI with compact metrics and traces.
Training, in numbers
These are the recorded run metrics—not a claim that synthetic data represents real private conversations.
100ksynthetic train examples
2kdisjoint eval examples
0.05661final eval loss
97.71%eval token accuracy
1,563final global step
13h58mapproximate runtime
64batch size · accumulation 1
512training max sequence tokens
Adapter integrity: `0fba57e246ed38cd6a7ea5b040553079ba9f691e8afb52a8834182efbe0cb464`. The public Hugging Face repository contains adapter files and a manifest only—no raw rows, checkpoints, or base weights.
Inference and telemetry
Every prompt can show throughput and timing while keeping message bodies out of persistent telemetry.
DECODE
Token generation
The UI reports output tokens, tokens per second, total response time, and a natural stop reason when available.
GPU
Hardware signal
A host loop samples A10G utilization, VRAM, temperature, and power. The latest observed idle sample was 0% GPU, 20,019/23,028 MiB, 38°C, 61.55 W.
What went wrong—and how it was fixed
These incidents are part of the project record. They are why the final deployment evidence is explicit.
CAPACITY
AWS us-east-2a could not supply the stopped GPU
The original `g5.2xlarge` restart hit InsufficientInstanceCapacity. A replacement was launched in us-east-2b, restored from S3, and validated before the old instance was terminated.
SNAPSHOT
The migration AMI was too slow to be the recovery path
The EBS snapshot remained slow, so the safer path used a fresh Deep Learning AMI and restored the adapter/code from encrypted S3 instead of waiting indefinitely.
CERTIFICATE
Fresh Caddy storage had no TLS certificate
Let’s Encrypt could not complete its challenge while port 80 was closed. Port 80 was opened only for issuance, the certificate was obtained, and the rule was removed. The security group is HTTPS-only again.
WARMUP
Readiness initially timed out during CUDA compilation
vLLM needed several minutes for model loading, torch compilation, CUDA graph capture, and multimodal warmup. The final readiness probe verified both `qwen35-base` and the `aurorium` adapter.
DEPLOY
Git push did not immediately replace the Vercel production build
The production deployment was promoted explicitly and verified as Ready. A streaming probe then returned a clean `finish_reason: stop`.
Runtime boundary
One A10G is a personal reasoning workspace, not a redundant high-concurrency platform.
| Layer | Role | Public? |
|---|---|---|
| Vercel / Next.js | Chat UI, streaming display, metrics, compact trace | yes |
| Caddy | HTTPS termination and reverse proxy | 443 only |
| FastAPI | Client auth, request shaping, telemetry endpoints | yes |
| Rust gateway | Internal request bridge to vLLM | no |
| vLLM | Base model plus LoRA adapter, internal model route | no |
Cost and limits
Billing is controlled by AWS runtime state and credits; this page deliberately avoids inventing a final invoice.
GPU HOURS
What consumes money
The running `g5.2xlarge` is the main variable cost. The encrypted 300 GiB gp3 volume, EIP, S3 storage, data transfer, and Vercel usage are separate line items.
CREDITS
What credits do
AWS applies eligible credits against eligible usage. The billing console is authoritative; the project cannot infer remaining credit balance from runtime telemetry.
TRADE-OFF
Why stop now
Training, adapter export, evidence, GitHub, Hugging Face, and the explainer are complete. Terminating the GPU stops ongoing instance charges while preserving durable artifacts.
Evidence and honest gaps
Use the links below to inspect the artifacts directly.
GITHUB
Code and stats
MODEL
Adapter publication
Hugging Face adapter
Published files only; no raw data or base weights.
ARCHIVE
Product surfaces
Open the chat workspace
GPU backend terminated
The backend was intentionally shut down after the final snapshot; GitHub, S3, Hugging Face, and this explainer remain available.
Not claimed: this is not a benchmark of frontier-model quality, not proof of high concurrency, and not proof that synthetic rows equal a private conversation archive. The evidence supports the narrower claim: a verified personal-model prototype was trained, served, measured, documented, and published within the stated boundary.