devkernels

Projects · Selected Work

Things I've shipped, and things still shipping.

Agent systems, content infrastructure, real-time data pipelines, and the occasional editorial side-quest. Production software over side projects, but I'll shamelessly feature one tasteful static site.

Featured

ACTIVE2025 – PresentArchitect & Operator

OpenClaw — 6-Agent Orchestration Layer

A self-hosted multi-agent system that actually runs 24/7.

Six specialized agents (main, architect, ops, calendar, learning, portfolio) running on OpenClaw with a hybrid Redis + pgvector memory substrate. Reachable through five Telegram bots, each bound to a specific agent role. Voice pipeline with XTTS-v2 for outbound calls and message playback. Self-healing systemd stack on a single Hetzner VPS — zero managed services, full observability.

Uptime 99.6% over the last 90 days across agent gateway + workers
Sub-second p95 latency on memory lookups across Redis + pgvector hybrid
Zero-touch automation for daily content, newsletter, and learning pipelines
OpenClawPythonRedispgvectorPostgreSQLTelegram Bot APIXTTS-v2systemdHetzner

More Work

ACTIVE2025 – Present

devkernels Content Infrastructure

A full content pipeline from idea capture to 5-platform publish.

End-to-end content system: Notion DB as source of truth, Claude-driven humanizer + self-review, platform-native adapters for LinkedIn, Medium, Substack, Instagram, and newsletter. Remotion-rendered carousels and reels with brand-locked theme. Runs on cron with per-stage checkpointing.

  • 8 research sources, auto-ranked into content angles every 3 days
  • 4 platform adapters generating platform-native output from a single canonical article
  • Brand consistency enforced programmatically — one theme, zero copy drift
Next.jsNotion APIClaude SonnetRemotionPythonSendGridReact Email
ACTIVE2026 – Present

claw — Scripts-to-Product Consolidation

Unifying a year of automation scripts into one coherent product.

A consolidation initiative to take dozens of one-off scripts — memory management, knowledge-base synthesis, content pipelines — and pull them under one product surface with shared primitives, a common config layer, and a unified CLI. PRD locked on the claw board; architecture decisions committed to memory.

  • Design doc + product brief finalized (P0 on the claw Notion board)
  • Shared memory layer reused across 4 sub-systems
  • Single CLI replacing 14 ad-hoc scripts
PythonTyperpgvectorChromaDBPydanticuv
WORK2025 – Present

Real-Time CDC Pipelines

Sub-second change propagation over Oracle GoldenGate + WebSocket.

Designed real-time change-data-capture streaming pipelines built on Oracle GoldenGate with a WebSocket event layer for downstream consumers. Paired with a DAG-based orchestration engine — topological sort, concurrent execution, stateful failure recovery — for multi-stage data transformations on OCI.

  • Sub-second propagation from source Oracle DB to downstream consumers
  • DAG orchestration engine handles retry, backoff, and failure recovery per node
  • Scaled across multi-tenant ingest on Oracle Cloud Infrastructure
JavaOracle GoldenGateWebSocketApache SparkOCITerraform
ACTIVE2025 – Present

Karpathy-Style Knowledge Base

A personal wiki that compounds, not a read-it-later graveyard.

Domain-organized markdown KB under /ai/{domain}/, auto-populated by the content consumer pipeline. Articles route to the right folder by tag overlap. Weekly Opus-driven synthesis per domain produces a "State of Knowledge" digest every Friday — a rolling research journal, not a dead archive.

  • 9 domains, hundreds of articles auto-routed by tag overlap
  • Weekly per-domain synthesis surfacing emerging patterns
  • Backs content ideas, build bets, and strategic reads
PythonMarkdownClaude OpusChromaDBcron
SHIPPED2025

Local LLM Inference Stack

Ollama + quantized inference for privacy-first agent workloads.

Optimized local inference pipeline — Ollama for model serving, KV cache quantization, custom embedding generation for agent memory. Used as the privacy-preserving tier for sensitive agent traffic when cloud LLMs aren't an option.

  • Throughput ≈ 40 tok/s on commodity hardware via KV cache quantization
  • Embedding pipeline feeds both pgvector and ChromaDB for hybrid retrieval
  • Drop-in fallback for cloud LLMs — same agent surface, local inference
OllamaPythonpgvectorChromaDBFastAPI
ACTIVE2026

devkernels Portfolio (this site)

Editorial tech portfolio with MDX, command palette, and TOC.

This site: Next.js 16 + MDX, Playfair Display + DM Sans + JetBrains Mono, rust-on-white editorial palette, custom ⌘K command palette, TOC sidebar with active-section highlight, RSS, sitemap, structured data, newsletter capture, and a /uses page.

  • Zero runtime JS on content pages — static export to GitHub Pages
  • ⌘K palette indexes all posts + tags + nav
  • TOC sidebar extracted from MDX on the server, no client parsing cost
Next.jsMDXTailwindLucideGitHub Pages

There's a longer list. The dull ones aren't on here. If something you're building needs an agent layer, a CDC pipeline, or someone who'll ship instead of bikeshed —