Sanitized public case study
StarkGrid: governed ingestion and operational evidence for artifact-heavy work.
StarkGrid is a private modular MVP/demonstration platform built by Quentin McClellan. It explores deterministic processing of digital and physical artifacts through clear module responsibilities, source-backed evidence, policy-gated orchestration, and read-only review surfaces.
Problem
Operational evidence is often scattered before automation even begins.
Files, scans, spreadsheets, workflow markers, and infrastructure observations can become difficult to trust when they are handled as one-off tasks. StarkGrid was built as a portfolio-safe proof artifact for a different approach: preserve identity, separate responsibilities, collect receipts, and make the current proof boundary explicit.
Architecture
Module boundaries are the product discipline.
StarkGrid is built as a multi-module Java/Maven pipeline running on Kafka. Each module owns a clear responsibility. Active modules in the current working system:
Multi-node execution
Two-node lab with remote GPU-capable worker evidence.
StarkGrid runs distributed across two lab nodes. Jarvis is the primary control-plane: Kafka/Redpanda, all Spring Boot modules, Athena, and athena-observer. Ultron is a GTX 1080 GPU-capable remote worker: Whisper audio transcription sidecar, OCR acceleration pathway (not enabled for current demos), and embedding pathway (not enabled).
Remote worker evidence is captured through athena-observer: the distributed worker status snapshot shows Ultron's GPU presence, worker identity, and evidence kind without exposing private machine details.
Validated proof snapshot
A final Rocketbook proof run produced source-backed distributed worker evidence.
The sanitized proof summary supports a controlled MVP/demo claim: StarkGrid can surface source-backed operational evidence and distributed worker visibility in a governed lab demonstration. It is not a production customer deployment claim.
Caveats: gpuOcrEnabled=false and embeddingsEnabled=false are deliberate guardrails.
The receipt shows remote worker visibility and GPU-capable node evidence, not production GPU OCR or production embedding acceleration.
StarkGrid Audio Lane Validation
Audio notes → searchable StarkGrid evidence
Audio notes became searchable evidence.
StarkGrid validated an MVP audio lane where a local audio note was staged under /starkgrid/audio-inbox, manually
dispatched from an approved AudioTranscriptionCandidateV1 preview, transcribed by the Ultron GTX 1080 Whisper sidecar,
emitted as transcript TXT, JSON, SRT, and manifest artifacts, ingested by Jarvis through the normal StarkGrid pipeline, and
surfaced in Search as an audio-transcript result.
The same validation captured a practical recovery story: an NFS restart caused stale /starkgrid bind mounts in running
worker containers. Search, Demeter, Argus, and Themis showed stale/local views; Demeter DLQ captured ROOT_NOT_FOUND
for DIG-E1F41CFF; Argus DLQ captured NoSuchFileException / ATLAS_IO_FAILURE for
DIG-46A831EB; and the final clean run succeeded after worker containers were force-recreated.
Caveat: Whisper transcripts are AI-generated derived evidence and must be reviewed against source audio for high-stakes use. This was manually dispatched in MVP form; it is not a production GPU OCR, production GPU embedding, or cloud transcription claim.
Caveats and honesty boundaries
This is a sanitized public case study for a private repo.
- The StarkGrid source repository remains private; this page is a public sanitized case study.
- No production customer claim is made.
- No production GPU OCR acceleration claim is made.
- GPU OCR was intentionally not enabled for this proof path.
- Embeddings were intentionally not enabled for this proof path.
- Ultron GTX 1080 evidence demonstrates remote GPU-capable worker visibility, not production GPU OCR acceleration.
- The audio lane was manually dispatched in MVP form; Whisper transcripts are AI-generated derived evidence that require source-audio review for high-stakes use.
- Hephaestus vector backend contract is under active hardening; vector search is not claimed as production-ready.
- This export excludes private runtime evidence files, internal operational commands, secrets, tokens, SSH information, and machine access details.
What this demonstrates
Skills shown without overstating the proof.
Architecture judgment
Clear module ownership, deterministic processing boundaries, and evidence-aware scope control.
Platform discipline
Linux, Docker, Kafka, workflow automation, source-backed summaries, and repeatable rehearsal practices.
Operator empathy
Automation framed as a way to reduce ambiguity, preserve receipts, and keep reviewable decisions visible.