Turlock, California · Est. 2024

We build things
that probably shouldn’t exist
but absolutely need to.

Independent AI research lab. Local-first infrastructure. No cloud dependencies, no vendor lock-in, no asking permission. Genuine curiosity applied to hard problems in automation, machine intelligence, and autonomous systems.

Turlock · CA · USA
Conceived over Pepsi and cat pictures.
Active Research
01
Genuine People
Personalities
Separating identity from cognition in language model prompting. GPP defines who the model is; PPS defines how it thinks. The distinction produces measurably more coherent, stable AI personas — particularly under adversarial or ambiguous inputs. Think-space grounding prevents persona fracture under pressure.
Active · 4B Models
02
Smart-OCR™
Document Intelligence
Vision-model document processing using cognitive instruction prompting rather than behavioral prompting. Handles mixed-content documents — structured text, diagrams, tables, handwriting — with context-aware extraction strategies derived from PPS research.
Active · Production
03
Neural Network
Visualization
Real-time visualization of training dynamics: gradient descent, neuron specialization, failure modes in small networks. Built to make the invisible visible. Watching a network learn XOR reveals more about deep learning than most textbooks manage in a semester.
Research · Ongoing
04
Multi-Agent
AI Analysis
Ensemble of purpose-built analyst personas — foreign policy, market intelligence, investigative journalism, automated fact-checking — operating on live RSS ingestion with extended chain-of-thought and tool-grounded verification. Four agents. One article. Four completely different cognitive paths.
Active · n8n + pgvector
05
Workflow
Automation
Self-hosted n8n orchestrating real operations: property management telephony triage, document pipelines, alert routing with AI severity classification, SITREP generation. Automation that runs actual businesses. Not demos. Flood, fire, and blood get the VP of Property Management. Everything else waits until morning.
Production · Live
06
StrandNet /
Wavefield Cognition
Experimental neural architecture treating consciousness as wavefield dynamics rather than discrete activation states. Consciousness as interference pattern. Attention as resonance. A theoretical framework in search of the right experiment to break it.
Theoretical · Exploratory
Running Systems
KLAB Radio
24/7 AI-DJ station with distinct GPP personas, genre-aware scheduling, Redis-backed play history, and a General Manager who occasionally questions his life choices. If you can hear the infomercials, go to bed.
• Live
Aurora Fleet
Local GPU inference cluster running LMStudio with fixed embeddings endpoint — semantic search across the full document corpus without touching a cloud API.
• Active
Property AI
After-hours telephony triage for property management: voice to transcript to AI schema extraction to property lookup to priority routing. Real operations. Real emergencies. Real consequences for mumbling your address.
• Production
News Pipeline
RSS ingestion feeding four analyst personas with Wikipedia-grounded fact-checking and extended reasoning chains. The think-space work is where the real cognition happens. The output is almost a summary of what already resolved in there.
• Active
Edgar
Voice-first local AI assistant. Whisper input, local LLM cognition, TTS output, GPP personality layer. Designed to feel like someone home, not a product. The long project. Worth waiting for.
○ In Development
Photo Intelligence
Semantic search across 11,500+ personal images. Natural language retrieval without cloud indexing. Built for a specific problem: helping someone find the right photo in a 40-year archive without describing it perfectly.
○ In Development
Current Field Project
AIGS — Agricultural Intelligence Ground System
Autonomous ground patrol vehicles for Central Valley agriculture. Local-first AI base station. No cloud, no subscription, no vendor. One open platform — five mission variants covering the crops and livestock that define this region’s economy. Proposed in partnership with CSU Stanislaus.
gDrone™  Orchards
vDrone™  Vineyards
cDrone™  Corn
dDrone™  Dairy
BahDrone™  Sheep
Philosophy
Build it.
Break it.
Learn from it.
Build it better.

The justsparx Digital Innovation Lab operates on the principle that the most interesting problems exist at the edges of what’s supposed to be possible with the hardware you already have. Resource constraints don’t limit innovation — they direct it.

We are local-first by philosophy, not just by necessity. Data that stays on your infrastructure is data you control. Automation that runs without a vendor is automation that runs when the vendor doesn’t. AI that runs locally is AI that works when the internet doesn’t. These are not inconveniences to route around. They are design principles.

The research here is real research. GPP/PPS isn’t a prompt engineering tutorial — it’s a framework emerging from hundreds of hours of systematic testing across model families and sizes. The multi-agent analysis pipeline isn’t a demo — it runs on live news. AIGS isn’t a concept deck — the chassis specs exist and the first prototype is being sourced.

We’re not trying to disrupt anything. We’re trying to understand how things work by building them, and occasionally discovering that the thing we built is actually useful to someone else.

Lab Infrastructure
100%
Local inference — no cloud APIs for core operations
4B
Optimal model scale for GPP persona coherence (current finding)
4
Active analyst personas on live RSS ingestion
0
Vendor dependencies for core lab operations
Local-First
If it requires a subscription to run, it’s a liability. Everything critical runs on hardware we own.
Open Stack
n8n, PostgreSQL, pgvector, Ollama, LMStudio. No black boxes in the pipeline.
Real Problems
The property management system handles real emergencies. The news pipeline reads real news. Nothing here is a demo.