KNOWLEDGE GRAPH
The map
How the work connects — projects, the tech behind them, and the ideas they share. Every project is wired to the technologies it uses, the capability domains it demonstrates, the agent frameworks involved, and the writing that explains it, so the threads running through the whole portfolio show up as the nodes that more than one thing leans on.
96 nodes · 278 connections — 25 projects · 11 capabilities · 2 frameworks · 54 technologies · 4 posts. Every node grounded in the real work.
Full index (text)›
The same map as text — every project with the technologies it uses, the capabilities it demonstrates, and any writing about it. This layer is fully crawlable and works with JavaScript disabled, with a coarse pointer, or with reduced motion on.
Projects
Greenfield 2026 GEO monitoring SaaS — a tested platform foundation (RLS, append-only credits ledger, Temporal refund saga, replay-safe Stripe) at 24/24 tests.
demonstratesAI-Visibility / GEOBackend & DistributedLLM-Ops & ObservabilityEvaluation & ReliabilityusesNext.jsDockerPostgreSQLBullMQ / RedisDrizzleGrafanaLokiNestJSOpenTelemetryPinoPlaywrightTemporalFastifyRow-Level SecurityStripeProduction AI-visibility / GEO analytics platform he operates as LLM-ops / reliability engineer — diagnosed the severity-1 silent-alerting blind spot.
demonstratesAI-Visibility / GEOBackend & DistributedLLM-Ops & ObservabilityusesNext.jsPostgreSQLBullMQ / RedisDrizzleGrafanaLokiNestJSOpenTelemetryPinoTemporalPM2writing: The five states of AI-answer visibility · The bug that made our alerts lie
Production nonprofit AI research-discovery PWA — resolve a paper across 7 scholarly APIs, chat in 70+ languages, read as HTML, with a first-class GEO program.
demonstratesAI-Visibility / GEOFull-Stack (Next.js/React)RAG & RetrievalDomain VerticalsusesGeminiNext.jsPostgreSQLRAGReactJSON-LDllms.txtVercelFastAPIRailwaySupabaseGEO Platform Suite (Kala)
Three-repo R&D-to-product GEO effort — a Next.js 15 + Elasticsearch backend and a sole-authored Python data-quality pipeline running prompts across 7 providers.
demonstratesAI-Visibility / GEOBackend & DistributedFull-Stack (Next.js/React)usesPythonGeminiOpenAINext.jsDockerReactPerplexityClaudeElasticsearchGenAI Learning Lab
Self-directed project-based GenAI program — multimodal RAG, KG-RAG, LangGraph agents — culminating in a working single-user multi-agent Research Assistant prototype.
demonstratesAgents & MCPRAG & RetrievalMultimodal & Document AIusesPythonGeminiLangChainRAGLangGraphNeo4jChromaGroq / LlamaGenAI GitHub Portfolio
Five-repo public GenAI portfolio — a LangGraph 5-node Research Agent that ran end-to-end, multimodal RAG, and graph-augmented RAG over SEC 10-Ks in Neo4j.
demonstratesAgents & MCPRAG & RetrievalusesPythonGeminiOpenAILangChainRAGLangGraphNeo4jGroq / LlamaOER Matching (OERMatch)
FastAPI + pgvector hybrid (semantic + lexical) RAG spine with evidence-required Gemini reranking — notable for a correct negative-control no_match.
demonstratesRAG & RetrievalEvaluation & ReliabilityusesPythonGeminiPostgreSQLRAGFastAPINeonpgvectorwriting: Teaching RAG to say "I don’t know"
PH 2024 — Data & LLM Pipelines
Multi-version, multi-LLM ETL unifying WhatsApp + Zoom-VTT human-conversation data into one schema, with reversible summarization and an 8-model provider benchmark.
demonstratesLLM-Ops & ObservabilityEvaluation & ReliabilityusesPythonGeminiOpenAIAWSLangChainPerplexityClaudeGroq / LlamaCV Maker (FastCV)
Shipped ATS-first resume-builder SaaS — no-login local drafts, Google OAuth gating, PDF/DOCX/TXT export, a transparent 9-criterion ATS scoring engine.
demonstratesFull-Stack (Next.js/React)Domain VerticalsusesGeminiNext.jsPostgreSQLReactVercelNeonPrismaOpenClaw Personal Agent OS
Self-hosted always-on personal agent layer he integrated/operated — offline WhatsApp voice assistant, 8 cron workers, and a 6,277-AI-Overview SERP scrape run.
demonstratesAgents & MCPAI-Visibility / GEODevOps & Agentic-DevOpsusesPythonDockerLinuxfaster-whisperPH 2024 — Early AI/ML R&D
Six 2024 exploratory systems — TaskWeaver agent-framework evaluation, an LLM semantic-response cache, a Neo4j medical KG with a GPT-to-Cypher assistant, vision/OCR OHLC.
demonstratesAgents & MCPRAG & RetrievalMultimodal & Document AIusesPythonGeminiOpenAINeo4jChromaOpenCVAI-search brand-visibility checker — a CLI + MCP server shipped with its own eval harness; extractor graded against human labels behind a CI gate.
demonstratesAgents & MCPAI-Visibility / GEOEvaluation & ReliabilityusesPythonFastMCPMCPGitHub ActionsSelf-hosted browser-based multi-user remote desktop on AWS EC2 — Guacamole + XRDP/VNC + a hand-built WireGuard split-tunnel firewall, via an AI-agent-driven migration.
demonstratesDevOps & Agentic-DevOpsusesDockerAWSLinuxApache GuacamolenginxWireGuard / nftablesHackathons & Self-Study
Breadth track — a serverless Lambda SEO scraper, a multimodal Gemini subtitle/cultural-lexicon pipeline, applied QLoRA/FSDP fine-tuning, an LLM-as-judge refine loop.
demonstratesMultimodal & Document AIEvaluation & ReliabilityFine-Tuning & Applied MLusesPythonGeminiOpenAIAWSAWS LambdaMulti-Client MCP Lab
A FastMCP/SSE tool server wired to four client surfaces (Cursor, raw SDK, LangChain, LangGraph) plus a custom MIME file-transfer tool round-tripping real binaries.
demonstratesAgents & MCPusesPythonOpenAILangChainFastMCPLangGraphMCPwriting: One MCP server, four clients
Project Hamburg Website
SSR-native nonprofit site re-architected to pass Ad Grants review (3/9 → 9/9), with a bespoke GEO layer and a ~1,124-line in-browser design-telemetry scanner.
demonstratesAI-Visibility / GEOFull-Stack (Next.js/React)usesNext.jsReactJSON-LDllms.txtVercelPlaywrightGEO Knowledge Base
Consolidated GEO domain asset — the 8-signal Entity Authority model, 5 Visibility States, mention/citation hierarchy, and an 11-parser normalization layer.
demonstratesAI-Visibility / GEODomain VerticalsusesPythonJSON-LDllms.txtAWS LambdaOff-Hour Automation Suite
Four independent off-hours sub-projects — a shipped Python Zoom scheduler, an n8n Drive pipeline, a LandingAI PDF agent, and a multi-provider product-intel agent.
demonstratesMultimodal & Document AIDomain VerticalsusesPythonGeminiOpenAIPerplexityYOLOv82025 OCR/API R&D
Production-grade multi-provider academic-paper intelligence pipeline (4 scholarly APIs, field-level provenance) plus OCR/PDF-structure experiments — 312 enriched records.
demonstratesRAG & RetrievalMultimodal & Document AIDomain VerticalsusesPythonGeminiAWS TextractPDF-to-HTML Worker
Secret-authenticated FastAPI microservice converting PDFs to layout-preserving HTML in a pinned Docker image on Railway — SSRF defense, content-addressed caching.
demonstratesBackend & DistributedMultimodal & Document AIusesPythonDockerFastAPIRailwayAR OCR Camera
Real-time AR OCR web app — a custom OpenCV motion-stabilization + ensemble text-detection pipeline gating when to fire an expensive GPT-4o Vision call.
demonstratesMultimodal & Document AIusesPythonOpenAIOpenCVFlaskZeroHedge Newsletter App
Statically-exported Next.js financial calendar of sell-side research + macro events — strict TS domain model, a defensive normalization layer, a green GitHub Actions CI.
demonstratesFull-Stack (Next.js/React)Domain VerticalsusesNext.jsTypeScriptGitHub ActionsOnlySPX
Single-file zero-build S&P 500 dashboard — from-scratch SMA/EMA/MACD/RSI math and a trading-day-aware synchronized pan/zoom engine across three Chart.js panels.
demonstratesFull-Stack (Next.js/React)Domain VerticalsusesTypeScriptChart.jsPaper2Agent (study)
Studied & reverse-engineered an academic multi-agent system that auto-converts a paper codebase into a FastMCP server — an architectural teardown, not a build.
demonstratesAgents & MCPusesPythonFastMCPMCPResuscitation Health App
Offline-first clinical resuscitation decision-support PWA — he built the interaction layer (drag-overlay ghost-clone, FSM-driven wheel UI). A training prototype.
demonstratesFull-Stack (Next.js/React)Domain VerticalsusesTypeScript
Capability domains
The eleven domains the work spans, each with the projects that demonstrate it — the spine that ties the constellation together.
Agents & MCP
GenAI Learning Lab, GenAI GitHub Portfolio, OpenClaw Personal Agent OS, PH 2024 — Early AI/ML R&D, geocheck, Multi-Client MCP Lab, Paper2Agent (study)
AI-Visibility / GEO
CiteStreak, BuzzView, MapleScholar, GEO Platform Suite (Kala), OpenClaw Personal Agent OS, geocheck, Project Hamburg Website, GEO Knowledge Base
Backend & Distributed
CiteStreak, BuzzView, GEO Platform Suite (Kala), PDF-to-HTML Worker
Full-Stack (Next.js/React)
MapleScholar, GEO Platform Suite (Kala), CV Maker (FastCV), Project Hamburg Website, ZeroHedge Newsletter App, OnlySPX, Resuscitation Health App
RAG & Retrieval
MapleScholar, GenAI Learning Lab, GenAI GitHub Portfolio, OER Matching (OERMatch), PH 2024 — Early AI/ML R&D, 2025 OCR/API R&D
Multimodal & Document AI
GenAI Learning Lab, PH 2024 — Early AI/ML R&D, Hackathons & Self-Study, Off-Hour Automation Suite, 2025 OCR/API R&D, PDF-to-HTML Worker, AR OCR Camera
DevOps & Agentic-DevOps
OpenClaw Personal Agent OS, Guacamoly remote-desktop
Domain Verticals
MapleScholar, CV Maker (FastCV), GEO Knowledge Base, Off-Hour Automation Suite, 2025 OCR/API R&D, ZeroHedge Newsletter App, OnlySPX, Resuscitation Health App
LLM-Ops & Observability
CiteStreak, BuzzView, PH 2024 — Data & LLM Pipelines
Evaluation & Reliability
CiteStreak, OER Matching (OERMatch), PH 2024 — Data & LLM Pipelines, geocheck, Hackathons & Self-Study
Fine-Tuning & Applied ML
Hackathons & Self-Study
Agent frameworks
The agentic-assistant frameworks behind the agent work — the specialty that connects the personal-agent and agentic-DevOps projects to the Agents & MCP domain.
OpenClaw
Open-source multi-agent orchestration runtime — skill-based, MCP-integrated, local-first. He integrates, extends, and operates it on his homelab.
powers: OpenClaw Personal Agent OS, Guacamoly remote-desktop
Hermes Agent
Nous Research persistent personal assistant — 40+ tools, subagents, a three-layer memory, and WhatsApp/Telegram/Slack gateways. An agentic-assistant reference point.
powers: OpenClaw Personal Agent OS
Writing
The five states of AI-answer visibility
A five-state model for AI-answer visibility and why the mention–citation gap is cheap to measure and expensive to ignore.
The bug that made our alerts lie
A severity-1 observability blind spot: a log-shipping prefix broke level extraction, so "no errors" meant "we stopped seeing errors."
about: BuzzView
Teaching RAG to say "I don’t know"
Two gates — a model-free score floor and an evidence-required reranker — plus a negative control, so retrieval can honestly return "nothing fits."
about: OER Matching (OERMatch)
One MCP server, four clients
Pointing Cursor, a raw SDK client, LangChain, and LangGraph at one MCP server to test "write once, call from any agent."
about: Multi-Client MCP Lab