Portfolio

Production Systems I've Built

Every project ships with verified outcomes, real problem-solving, and measurable results. No fluff - only production-grade work.

All8 projects
Featured

Hermes Agent Framework

AI Agent Orchestration System

Production-grade AI agent framework with multi-model routing, MCP integration, and a 600K+ node knowledge graph. Orchestrates intent classification, entity extraction, and tool execution across microservices.

ProblemNeed a scalable architecture where multiple AI models collaborate, context is maintained across sessions, and tools are dynamically dispatched based on intent - all with sub-second latency.
SolutionBuilt a modular agent framework with an Intent Router (BM25 classification), Entity Extractor, MCP server integration, and a unified knowledge graph. Deployed as 4+ microservices with health monitoring cron.
Outcome600K+ node memory index, 94% tool reduction via intent routing, sub-100ms classification, production deployment on Proxmox cluster.
PythonFastAPIMCPGraphRAGChromaDBBM25+3
View Sourcehermes-agent
Featured

CTV Platform Development

Cross-Platform Streaming Applications

Production streaming apps across all major smart TV platforms: Roku, FireTV, Android TV, Samsung Tizen, LG webOS, and Vizio. Includes SDK toolchains, CI/CD pipelines, and YouTube channel feed integration.

ProblemEach CTV platform has its own ecosystem, SDK, packaging format, and certification process. Maintaining feature parity across all platforms requires significant effort and specialized tooling.
SolutionBuilt a unified development infrastructure with GitHub Actions runners for each platform, shared component libraries, automated CI/CD pipelines via BrighterScript (Roku), Java/Kotlin (FireTV), and webOS CLI toolchains.
OutcomeProduction apps on 6 platforms, automated CI/CD for all, YouTube channel integration with 249+ videos indexed. GitHub Actions runner handling multi-platform builds.
RokuBrighterScriptFireTVAndroid TVTizenwebOS+4
View Sourcectv-platforms
Featured

CodeGraphContext (CGC)

Code Knowledge Graph Toolkit

MCP server and CLI toolkit that indexes codebases into knowledge graphs for intelligent code discovery, call graph analysis, data flow tracing, and automated bug detection.

ProblemUnderstanding large codebases requires traversing files, functions, and relationships manually. Existing tools don't provide semantic search across function definitions or cross-service dependency tracking.
SolutionBuilt a knowledge graph indexer that parses code into structured representations (nodes for functions, classes, imports; edges for calls, data flows). Added BM25 ranking, Cypher query support, and MCP integration.
OutcomeProduction toolkit used for cross-service analysis, hot-path detection (loop depth, memory allocations), architecture clustering. Supports full-text search with structural boosting.
PythonLSPCypherMCPChromaDBBM25+2
View Sourcecodegraph-context
Featured

Intel Arc B70 AI Toolboxes

GPU-Accelerated LLM Inference Infrastructure

Production Docker toolboxes for vLLM and llama.cpp on Intel Arc Pro B70 GPUs. Supports NVFP4/INT4 quantization, XPU optimization, and CI/CD pipeline for automated image publishing.

ProblemIntel ARC GPUs require special setup for LLM inference (XPU drivers, SYCL/vulkan). Deploying vLLM with tensor parallelism across 4 GPUs needs careful shared memory, security, and driver configuration.
SolutionCreated isolated Docker toolboxes for llama.cpp (SYCL + Vulkan) and vLLM with proper XPU environment variables. Configured --tensor-parallel-size 4, --shm-size 200g, --security-opt seccomp=unconfined for multi-GPU inference.
OutcomeDeepSeek V4-Flash deployment on Intel Arc Pro B70 with NVFP4 quantization (~68-85GB weights). ~25-45 tokens/sec generation, 5K-15K+ tokens/sec prefill. Automated CI/CD via GitHub Actions.
PythonvLLMllama.cppDockerIntel GPU (XPU)NVFP4+3
View Sourceintel-b70-toolboxes

Mem0ai Integration

Persistent Memory for AI Agents

Multi-layer memory system for AI agents with NER extraction, embedding quality monitoring, ChromaDB storage, and episodic/constructional memory indexing.

ProblemAI agents need durable context across sessions but traditional approaches either lose all state or accumulate unbounded noise. Need structured memory that's queryable by type (persona, episodic, instruction).
SolutionImplemented 4-layer memory system (L0 conversations -> L1 structured memories -> L2 scene blocks -> L3 persona synthesis). Added NER retraining pipeline targeting >90% accuracy for 8 entity types.
OutcomeStructured memory extraction working across all layers. NER models trained on target types. Embedding quality tracking logs to JSONL format.
PythonChromaDBBGE EmbeddingsNERBERTJSONL+1
View Sourcemem0ai-integration

All-Agentic-Architectures

35 Production Agent Architecture Reference

Comprehensive reference library of 35 production-grade agentic AI architectures including retrieval agents, coding assistants, research systems, and multi-agent workflows.

ProblemThe agentic AI space is growing rapidly but developers lack a curated reference of proven architecture patterns with implementation details and trade-off analysis.
SolutionDocumented 35 architectures from real implementations: DeerFlow, AutoGPT-style agents, subagent orchestration, MCP-based tools, RAG pipelines. Each with pattern description, complexity analysis, and when-to-use guidance.
OutcomeProduction reference library used for architecture decisions across multiple projects. Benchmarked and tested patterns in real deployments.
ResearchArchitectureMCPRAGSubagentsDocumentation
View Sourceall-agentic-architectures

AI Knowledge Router

Multi-System RAG Routing

Intelligent knowledge routing system that classifies queries and dispatches them to the optimal MCP backend (GraphRAG, wiki-rag, unified-rag, codebase-memory) with a fallback mechanism.

ProblemMultiple knowledge retrieval systems exist but users/tools need a single entry point that picks the right backend for each query type without manual configuration.
SolutionBuilt classification layer that routes queries to GraphRAG (entity relationships), wiki-rag (documentation), unified-rag (semantic across all collections), and codebase-memory (project-specific code). Includes router confidence scoring.
OutcomeSingle knowledge query endpoint serving 4+ backends with automatic routing. Router handles entity lookup, semantic search, and project-aware queries.
PythonFastAPIMCPGraphRAGChromaDBBM25+1
View Sourceai-knowledge-router

Homelab Infrastructure

Production Home Lab with Kubernetes & Proxmox

Multi-node Proxmox cluster running Kubernetes, Docker containers, VMs, and dedicated services - all exposed via Cloudflare tunnel. CI/CD pipeline using Ansible for zero-downtime deployments.

ProblemHome lab needs to support multiple concurrent workloads (AI inference, CTV development, media streaming, monitoring) with high availability and automated deployments.
SolutionProxmox cluster with 2 nodes + Raspberry Pi quorum. 3 Kubernetes nodes for HA. Docker networks with static IPs. Cloudflare tunnel for external access. Ansible playbook webhook listener for CI/CD from GitHub.
OutcomeProduction-grade infrastructure running on repurposed hardware. Zero-downtime deployments via Ansible webhooks. Multi-service orchestration including Jellyfin, Nextcloud, Uptime Kuma, Portainer, VS Code Server.
ProxmoxKubernetesDockerAnsibleCloudflare TunnelOpenWRT+2
View Sourcehomelab-infra