Public versions of the tooling and frameworks behind the projects here.
Two kinds of repos here. Toolkit is personal tooling I built for myself — my operating layer, my agent library, unambiguously mine to publish. Framework demos are clean reference implementations of the patterns behind production systems I built at work. The real production code stays with whoever owns it; the public versions here are rebuilt from scratch so you can read the architecture without anyone's IP getting in the way.
If you find anything that looks like it shouldn’t have shipped, email me.
Personal tooling — mine, portable, MIT-licensed.
The framework: hook intelligence, TF-IDF knowledge search, domain-aware agent routing, episodic memory via SQLite FTS5, and an MCP server exposing all of it to Claude Code. Plus workflow templates and a bootstrap installer.
$ git clone git@github.com:swsounds42/cascade.gitClone on any machine, run bootstrap.sh, and restore the full Cascade setup in under a minute. 163 specialist agents, 6 custom skills, hook runtime, portable settings template.
$ git clone git@github.com:swsounds42/claude-dotfiles.gitClean reference implementations of patterns behind production systems. Built from scratch.
An opinionated RevOps attribution framework, public and pluggable.
The methodology + a Python package. Rules as composable functions, a persistent learning loop via SQLite, CSV adapter shipped (Salesforce/HubSpot adapters via the five-method Protocol). 5 reference rules, 7 passing tests, 40-case pattern library documented in docs/methodology.md.
$ git clone git@github.com:swsounds42/cascade-attribution.gitThe pipeline pattern: N signal sources → per-identity filter → ranked sections → Slack delivery. Three source adapters (RSS, Hacker News, GitHub releases), deterministic filter with exclude-rules, YAML identity config, GitHub Actions cron. 10 passing tests, working --demo mode with bundled fixtures.
$ git clone git@github.com:swsounds42/weekly-signal-digest.gitAn MCP server that turns structured customer data into a near-complete deck.
MCP server exposing build_qbr(customer_json) → pptx. 7 slide renderers (title, agenda, big-stat, usage-grid, risk-matrix, roadmap, thank-you), deterministic brand resolver that picks palettes by hashing the customer name, full Acme Industries fixture, 6 passing tests.
$ git clone git@github.com:swsounds42/qbr-deck-mcp.gitDigest-card → seeded chat. The deeplink pattern, live in a tiny Next.js app.
Next.js app with 5 imaginary accounts. Click any card, chat opens pre-seeded with that account's context, signals, and last touch. Streaming text/plain responses (no AI SDK dep for the demo; clean swap point for a real LLM). Status-aware suggested openers, 9 passing tests.
$ git clone git@github.com:swsounds42/account-brief-chat.gitMost ops portfolios stop at the case study. The whole point of this one is that the work is real and the code is legible. Toolkit repos let you read my actual operating layer. Framework demos let you see the patterns that power the production systems — the pieces anyone could learn from and adapt — without dragging anyone’s IP into daylight.