📚LLM Wiki
An internal support tool: a self-maintaining knowledge base my AI sessions read before answering and update after building, so context compounds instead of resetting each time.
The AI Lab · self-hosted · local-first
I ship what I spec.
Product management is my craft; shipping is my proof. Everything below is a real, running system: designed, built, and operated end-to-end on my own infrastructure, AI-assisted and human-directed across a modern toolchain. No tutorials, no toy demos. Agent platforms, product concepts, native apps, and pipelines that run every day.
0
Local models served
0
Specialized agents
0%
Self-hosted
Featured projects
A closer look at a few. Click any screen to zoom.
Fleet ops · GenAI concept
Routesight
Fleet managers drown in dashcam events; insight arrives too late to coach anyone.
A driver-safety platform concept where managers chat with their fleet's event data and an autonomous flows engine runs scheduled coaching interventions: drafting coach emails, flagging fatigue patterns, and tracking whether interventions actually reduced events. Built on a strict "LLMs explain, tools compute" rule: every number comes from the database, never the model. Featured as a full case study on the main page.
Under the hoodFastAPI + Postgres row-level-security multi-tenant data plane · real MCP server · visual no-code flow editor · 6 seeded flows, 33% median tailgating-event reduction in simulation
Product concept
Account Truth Layer
Customer-success teams rebuild "what's happening with this account" in slides every single week.
A continuously-maintained, queryable account-state model: every meeting note, ticket, and status change flows into a living dossier, and any stakeholder can ask "what's really going on with account X?" and get a current, citation-backed answer. No deck reconstruction, no chasing reps.
Under the hoodNext.js 16 + FastAPI · event-driven ingest (S3 → SQS → Lambda) on emulated AWS · Claude with native citations over a prompt-cached corpus · fully containerized, <30s cold start
ask this account · every answer cited
QWhere does this account stand before the call?
1 security demo notes
2 Okta SSO ticket
3 commitments ledger
Account risk
Northwind Logistics
High0
updated 4m ago · 2 open requests
Key commitments
cost_center adapter prototype, end-to-endMay 29 · overdue
SSO via Okta go-live
status update to Sarah ChenMay 29 · overdue
SSO via Okta go-live
svc-checkout · $40k weekly cost spikeus-east-1 · flagged by deterministic detector
Awaiting approval
Agent loop · observe to verify
1
Observeextract parameters
auto2
Investigatepull evidence, tools step-by-step
frontier3
Reasonsynthesize hypothesis
frontier4
Proposerecommend action
approve5
Verifyconfirm against ledger
gatedROI gate · pre-executioncleared
64,213×vs 100× customer floor · ~$172k saved / mo
Per-investigation$1.04 / $1.50
Daily cap$9.44 / $50
recommended: Revert PR #4127 · svc-checkoutreversible
Product concept
Bounded-Autonomy FinOps
Cloud-cost anomalies get investigated days late; autonomous agents nobody trusts get switched off.
An interactive product concept for bounded-autonomy FinOps: an agent investigates spend anomalies through an Observe → Investigate → Reason → Propose → Verify loop, but every action passes a permission envelope (Auto / Approve / Locked) and a customer-tunable ROI gate. Failed investigations stay on the ledger by design: trust comes from showing the misses.
Under the hoodReact + Framer Motion animated agent loop · live ROI simulator with tunable spend knobs (ROI floor, per-investigation budget, daily cap) · outcome scorecard incl. COGS%
Platform
CrabCake
Cloud LLM bills explode when every task, trivial or hard, goes to a frontier model.
A self-hosted multi-agent platform: a router classifies each incoming request and sends cheap work to local models and hard work to Claude, behind one streaming gateway with pluggable providers (claude-code / ollama / openai-compat). Runs my actual life through a Telegram interface: CV pipeline, finance tracking, market discovery, cron reports.
Under the hoodNode.js gateway, 3-layer security · 9 specialized agents, 11 models in rotation · ≤$1/day cloud budget enforced by routing policy · streaming via Telegram edit-in-place
0Agents
0Models
≤$1Per day
3-layerSecurity
router → local models / Claude → Telegram
Desktop app
GemmaChat
ChatGPT-class assistance for work that can never leave the machine.
A local-first desktop AI app: chat, vision, artifacts, document generation, and an agent-style tool belt, running entirely on local models. Zero tokens leave the machine; conversations, memory, and files stay on disk.
Under the hoodElectron + React 19 + Tailwind 4 · Gemma 4 via Ollama (256K context, vision) · 18 tools incl. Python sandbox, web search, PDF/Excel · auto-extracted long-term memory · SQLite + FTS5
GemmaChat
offline
Summarize Q3-report.pdf and flag the top risks
★
📄 read Q3-report.pdf · 24 pages
Q3 revenue is up 14% — three risks stand out: supplier concentration, FX exposure, and thinning runway.
Message GemmaChat…↑
live scan · 4 topics
Opportunity feedautotuning
01
Fleet maintenance scheduling
92
PST
02
Driver onboarding friction
87
PST
03
Claims evidence automation
81
PST
04
Route risk benchmarking
78
PST
autotuner rewrote 4 queries · next cycle 02:00
PM tooling
Radar Discovery
Product discovery that depends on manually trawling Reddit and HN doesn't scale past one topic.
A multi-topic discovery pipeline that scrapes community signals nightly, clusters them, and scores opportunities on three axes: Pain (40%), Scale (35%), Timing (25%). The twist: an LLM autotuner reviews each cycle's results and rewrites its own search queries and sources for the next one. Discovery that improves itself while I sleep, at $0/cycle on local models.
Under the hoodReddit / HN / GitHub / web scraping → clustering → 3-axis scoring with cross-source multiplier · LLM autotuner rewrites its own queries weekly · dashboard with topic tabs + program editor · nightly launchd schedule
The long tail
More from the lab, each one real and in use.
⭐Tiny Wins
A native iOS parenting app: log a child's positive moments and watch a reward garden grow. Shown in full on the main page.
🚢PortPilot
Native macOS app that shows which process owns every listening port and which project it belongs to.
🐝Claude Swarm
One click launches a grid of parallel Claude Code agents in tmux, each pointed at its own directory.
🧠Second Brain
The household hub: one dashboard for cron reports, market discovery, image generation, and family spending.
📸Immich AI Scout
A daily pipeline that reads my phone screenshots, recognizes AI tools with Claude vision, and catalogs them, then cleans up after itself.
📄CV Tailoring Pipeline
Job description in, tailored CV out: 11 protected base facts, 12 tailoring rules, and a validator that rejects drift.
📈CryptoBot Pro
A paper-trading bot where bull and bear analyst agents debate every trade before a judge, risk panel, and portfolio manager sign off.
📊AlgoTrader Skill
A 27K-line Claude Code skill encoding a full equities trading methodology, with backtest-to-live parity.
🔍RAG System
A local retrieval pipeline over personal documents: the predecessor whose lessons became the LLM Wiki.
🎨ComfyUI Pipeline
Local FLUX.1 image generation wired into the dashboard: prompt → tuned workflow → gallery, no cloud.
⏰Cron Reports
Every two hours, a local model turns raw cron output into executive summaries with action items.
👨👩👧Family Proxy
A deliberately tiny reverse proxy that exposes exactly two dashboard routes to my family over a Cloudflare tunnel. Everything else is a 403.
🌐This Site
dor-pm.ai itself: served from a Mac mini through a Cloudflare tunnel with zero inbound ports, hardened headers, and custom-domain email.
✏️Excalidraw Toolkit
"Diagram this repo" → a live, editable architecture canvas. A Claude Code plugin with its own MCP server.
💬Telegram Interface
The whole system answers in one chat: CV requests, budgets, market radar, cron status, streamed live.