MARCELOBounded GGUF parsing
Architecture, context, tensor count, parameter hints and quantization metadata are read locally without loading the model for inference.
02 · Flagship case study · Source available · 1.0 · Local AI systems
VektorDeck
A local-first workstation manager for the parts of local AI that become difficult to trust once a setup grows: model files, multimodal projectors, launch flags, runtime ownership, RAM/VRAM pressure and benchmark history.
PRODUCT / STATUS1.0.0
Installation statusREADY + NOTES
Blocking issues0
ValidationCI + LOCAL + SMOKE
Primary runtimellama.cpp
SOURCE AVAILABLE / LOCAL PRODUCT
VEKTORDECK / 1.0NEW CAPTURE PENDING
QWEN 27B · IQ3_XS11.15 GB
SMART LAUNCH8K · EXPERIMENTAL
RELEASE READINESS0 BLOCKERS
LOCAL WORKSTATIONFULL PRODUCT CAPTURE WILL REPLACE THIS PANEL
The problem
Local AI should feel like software, not command-line archaeology.
My local AI setup worked, but it was increasingly difficult to reason about. Models lived across drives. Commands depended on exact flags. Vision required the right projector. Ports could already be occupied. A benchmark could look fast even when the machine was already under load.
VektorDeck turns that into one persistent local control surface. It discovers assets, inspects GGUF metadata, validates saved profiles, launches runtimes safely, watches the machine, records evidence and explains why a recommendation exists.
The project deliberately avoids becoming another chat frontend. Its job is to make the workstation itself understandable and trustworthy.
Architecture
One local control deck. Clear boundaries underneath.
The browser never launches arbitrary commands or scans drives directly. System-specific behavior stays behind a loopback FastAPI service and explicit runtime adapters.
UIReact + TypeScriptdashboard, profiles, evidence views
→
APIFastAPIloopback-only control boundary
→
COREPython servicesintelligence, safety, telemetry, runtime control
→
STATESQLite + local logsprofiles, models, evidence, protocol provenance
BOUNDARY
VektorDeck manages llama.cpp, AUTOMATIC1111 and Hermes locally. It does not require a cloud account, and model/runtime control stays on the workstation.
Model intelligence + Smart Launch
Understand the model first. Recommend a launch second.
GGUF inspection and launch recommendations are deliberately linked but kept conceptually separate. Metadata describes what is installed; Smart Launch explains what the current machine should try.
02 / INTELLIGENCELOCAL MODEL INSPECTOR
MODELQwen3.8 27B
QUANTIZATIONIQ3_XS · 11.15 GB
CONTEXT ADVERTISED262,144
PROJECTORVISION PAIRING AVAILABLE
SMART LAUNCH8,192 CTX · EXPERIMENTAL
MODEL INTELLIGENCEREASONS VISIBLE
HARDWARE31.6 / 16.0 GiB
RUNTIMEllama.cpp
HOSTLOCAL WORKSTATION
RULENO SILENT PROFILE CHANGES
EVIDENCE BOUNDARYRECOMMENDATION ≠ PROOF
Projector pairing
Vision projectors are suggested from folder/name evidence, shown with reasons, and never silently applied.
Profile readiness
Saved configurations are checked against indexed assets, local-only host rules and GGUF-advertised context limits before launch.
Transparent heuristics
The current 32 GB-class RAM / 16 GB VRAM workstation maps to an 8K starting point for this 11.15 GB multimodal model, with every reduction rule visible.
Real workstation evidence
Performance results are allowed to say “not proven yet.”
The application preserves old results but separates legacy history from controlled evidence. If the machine is already busy, Protocol v2 refuses the run instead of creating a fake clean benchmark.
release1.0.0
language modelQwen3.8 27B
model file11.15 GB · IQ3_XS
advertised context262,144
local RAM / VRAM31.6 / 16.0 GiB
Smart Launch start point8,192 ctx
release blockers0
release validationCI GREEN · LOCAL BUILD PASS · LIVE SMOKE PASS
READINESSREADY + NOTESPERF EVIDENCEPENDINGFAKE CLEAN RUNS0
03 / EVIDENCEPERFORMANCE TRENDS
BASELINENOT YET QUIET
RUN STATEPRESSURED / RETAINED
PROMOTIONREFUSED
WINNERNONE — INSUFFICIENT PROOF
DESIGN RULENOT PROVEN YET IS A VALID RESULT
Safety rules
Failure is explicit rather than convenient.
01Unknown listener on a port?
Do not replace it. Identify ownership where possible and fail closed when the process is not VektorDeck-managed.
REFUSE02Profile points outside indexed assets?
Do not launch. Model and projector paths must resolve to assets the local inventory knows about.
BLOCK03Benchmark starts on a busy machine?
Refuse the controlled run and surface likely contaminating processes instead of manufacturing clean evidence.
DEFER04Only one good benchmark?
Keep the recommendation experimental. Promotion requires repeated Protocol-v2 HEALTHY/TIGHT evidence for the same profile.
NO PROMOTIONFailure → redesign
The failures became some of the strongest parts of the product.
Benchmark numbers looked comparable until background CPU load changed the result.
Inference-time telemetry could show pressure, but it could not prove the machine was quiet before the run began.
REDESIGN
Protocol v2 records a pre-inference baseline, source, contaminators and inference-time evidence. Busy machines are allowed to say “not now.”
The updater downloaded its own fix but kept executing old PowerShell code already in memory.
The file on disk was correct while the current update process was still running stale logic.
REDESIGN
The updater hashes itself before and after pull and automatically restarts once when its own script changes.
A surviving runtime after backend restart could not simply be trusted by PID.
PIDs are reusable, and blindly adopting a process would create a dangerous ownership assumption.
REDESIGN
Runtime recovery verifies the saved profile, PID, executable path, port, model path and live llama-compatible API before reclaiming ownership.
Engineering surface
What VektorDeck demonstrates.
Python / FastAPI
Local APIs, runtime adapters, process ownership, telemetry, GGUF parsing, validation and orchestration.
React / TypeScript
Stateful control surfaces for runtimes, profiles, telemetry, evidence, Model Intelligence and release readiness.
SQLite
Persistent model inventory, launch profiles, benchmark history, runtime leases and Protocol-v2 provenance.
Systems engineering
Ports, process trees, Windows paths, PowerShell lifecycle issues, loopback networking and fail-closed ownership checks.
Evidence design
Explicit HEALTHY/TIGHT/PRESSURED/UNVERIFIED states, contamination checks and transparent recommendation reasons.
Release engineering
GitHub Actions, local regression/build gates, repair scripts, diagnostics and live post-relaunch smoke tests.
RÉSUMÉ
Built a local-first Windows AI workstation manager using Python/FastAPI, React/TypeScript and SQLite, with safe runtime ownership, GGUF metadata inspection, hardware telemetry, evidence-aware benchmarking, transparent launch recommendations, CI and live smoke validation.
VektorDeck 1.0
A finished tool built from a problem on my own machine.
VektorDeck 1.0 deliberately stops at a complete Windows local-workstation product. Installer packaging, ComfyUI, remote control and plugins are future extensions rather than unfinished core requirements.