flagship
platform engineering
K3s "Jukebox" — On-Demand Environment Provisioning Platform
An internal Platform-as-a-Service. Engineers pick an O Series version, database flavor,
Java version, and module config from a dropdown. Minutes later, a fully-configured environment
is live — k3s orchestrates, Longhorn persists, Node-RED automates the composition pipeline.
- › Days of manual setup → minutes of self-service.
- › Parameter-driven environment factory, not a fragile script.
- › Empowered non-infra users to spin up real environments independently.
# jukebox.spin
o_series_version: "9.x"
database: "postgres-16-ha"
java_version: "17"
mdu_profile: "sales-tax-full"
# t+0:00 → submit
# t+0:08 → ready
k3sLonghornNode-RED
DockerPostgreSQLPython
flagship · personal
full-stack platform · 18 months in production
Buddy — Personal Operations Platform
A self-hosted family operations platform I designed and built from scratch — and use every
single day. FastAPI single-page dashboard backed by a Neo4j knowledge graph, role-scoped
access for five household members, six external system integrations, AI-driven email
parsing via a local Ollama model, and a security stack I rewrote end-to-end.
- › 16 sections (inventory, health, automation, calendar, travel, ancestry, finance, recalls, admin) in ~10 KLOC.
- › Live PnL aggregator fans out to 3 trading bots — decimal-safe math, 10s in-process cache, per-system stale fallback.
- › Auth rewrite: argon2id with on-login self-heal of legacy SHA-256, double-submit-cookie CSRF, TOTP, IP-ban, 49-test pytest suite.
- › Real production fix: regex w/ negative lookahead that stopped two MEDIA paths concatenating when the LLM dropped a newline.
# in active use
household_users: 5
belongings_tracked: 98
receipts_parsed: 80
contacts_graphed: 47
market_candles: 27_000_000+
integrations: 6
in_production: "~18 months"
FastAPIPydanticNeo4j
SQLiteOllamaargon2id
pyotphttpxpytest
Home AssistantFrigateBlueBubbles
TelegramGmail IDLE
vanilla ES6nginxsystemd
Vertex Emerging Tech · production AI
demoed at Exchange 2025
heading to product
enterprise AI
Oracle → O Series AI Reconciliation Agent
An AI agent that reconciles tax data between Oracle source systems and Vertex O Series — surfacing
mismatches, classifying their cause, and proposing resolutions in language a tax operator can act on.
Demoed at Vertex Exchange 2025 in front of customers, and being incorporated
into the product roadmap.
- › Replaces a manual, error-prone tie-out that tax teams previously did line-by-line.
- › Reasons over the data, doesn't just diff it — explains why two values disagree.
- › Customer-facing demo validated the pattern; now productizing it.
# reconciliation_agent.run
source: "oracle.tax_lines"
target: "vertex.o_series"
mode: "explain + propose"
demo'd_at: "Exchange 2025"
status: "productizing"
LLM agentVertex O Series
OracleRAG
Python
patent pending
// applied AI · Vertex Emerging Tech
Natural-Language Tax Readout
A PoC that takes the raw output of an O Series transaction — the kind of dense, jurisdiction-heavy
breakdown only a tax expert can read — enriches it with surrounding context, and explains
the result in plain language for the human running the query. Closes the gap between
"the engine returned a number" and "here's why."
- › Transaction output → contextual enrichment → narrated explanation.
- › Patent application in process.
LLMEnrichment pipeline
Vertex O SeriesRAG
patent pending
// applied AI · Vertex Emerging Tech
SME-Driven Best-Practice Agent
Built an agent that ingests recorded SME interviews, distills them into structured Q&A pairs,
and uses those pairs to enrich standard RAG retrieval — so the assistant doesn't just
quote the docs, it answers the way the expert would. Captures tribal knowledge before it walks
out the door.
- › Interview audio → Q&A pair extraction → RAG-augmented expert answers.
- › Patent application in process.
LLM agentQ&A extraction
RAG enrichmentWhisper
// AI orchestration
Multi-Agent Home Orchestration
Parent Agent routes intent to scoped child agents (Home, Weather, Calendar, Finance),
each with its own tools and reasoning. PostgreSQL state, MCP transport, local LLMs —
a state-aware decision engine, not a chatbot. Brought the underlying n8n + MCP pattern
into Vertex as a secure orchestration platform for enterprise PoCs.
n8nMCPOllama
Home AssistantPostgreSQLFlowise
// edge AI · personal time
Real-Time Visual Tax Checkout
Roll a tennis ball past the camera. Jetson Nano detects the object, classifies it,
and calls Vertex O Series live for the correct tax. Vision → classification →
enterprise tax engine, end to end.
Jetson NanoObject detection
O Series APIPython
// AI decisioning
Autonomous Crypto Trading System
Multi-signal AI evaluates 1-min and 5-min OHLC, CryptoPanic sentiment, and aggregate
trend. Executes via OCO orders with dynamic stop-loss/take-profit. Tracks rolling
performance over 10-min and 1-hour windows. Real-time decisioning under uncertainty.
FlowiseNode-RED
BinanceCoinbase
CryptoPanicPostgreSQL
// data foundation
Enterprise Tax Knowledge Corpus for LLMs
The data layer behind AI assistants for tax. Aggregated product docs, legislative
content, and KB into a tagged, vector-searchable corpus — schema covers document type,
tax entity, and jurisdiction so retrieval is precise enough for enterprise use.
Chroma / MilvusPostgreSQL
FlowiseSemantic tagging
// infrastructure
Distributed AI Compute Fabric
Multi-node k3s with NVIDIA GPU scheduling, Longhorn persistence, and HA Postgres
running Ollama, vector DBs, and custom inference services at home — production-shaped
substrate for AI workloads.
k3sNVIDIA GPU
LonghornOllama
// emerging
MCP + API Integration Layer
MCP-compatible services that let LLMs invoke real tools — Home Assistant via secure
tokens, internal APIs through FastAPI shims. Tool servers as first-class infrastructure.
FastAPIMCP
Home AssistantDocker
// experiment
Personal Knowledge Graph
Family, travel, home inventory, and smart-home state unified as a Neo4j graph —
the substrate for a conversational interface. A personal intelligence hub instead
of seven disconnected dashboards.
Neo4jCustom APIs
Graph modeling