ONLINEcascade://samwarren.iov0.1.0
all projects
FIRING · PROD · DAILY
ORCHESTRATION LAYER· 01 / 06

Cascade

An AI-native operations layer for revenue teams.

A conversational layer across the entire GTM stack. Classifies intent in plain English, dispatches to the right specialist, pulls live context from every connected system, and learns from every correction. Replaces the operational glue work that used to eat a full day every week.

225
SKILLS
142
AGENTS
8+
MCP SOURCES
~90 min/day
TIME RECLAIMED
▸ architecture·signal chain
promptnatural languageintent routerclassify · dispatch+ reason over memoryskill225 callableagent142 scopedmcp8+ live systemsresultready to act onhybrid memory · tf-idf semantic + sqlite fts5 episodic
signal pathmemory read/writelive production system
PROBLEM

The modern GTM tech stack makes everything easier except the one thing that matters: deciding what to do next.

A revenue team today touches a dozen systems before lunch. CRM. Marketing automation. Revenue intelligence. CDP. Reverse ETL. A data warehouse. A spreadsheet that nobody talks about. Each tool is good at what it does. Each one assumes the human between them is going to read every signal, apply every rule, and execute every handoff in the right order.

That assumption is where revenue operations quietly burns a full day every week. Not on strategy. Not on judgment calls. On the operational glue — the classification, the context-gathering, the figuring out which rule applies to which record in which pipeline at which moment. It's work the stack should be doing for the humans, not the other way around.

APPROACH

One layer across the stack, spoken to in plain English.

Cascade sits between the human and the tools. A prompt comes in. The system classifies intent, pulls relevant context from an indexed knowledge base, picks the right specialist for the work — a skill, a sub-agent, or a live production system — and returns a result that's ready to act on.

The infrastructure that took real work is the memory. A hybrid store: TF-IDF semantic search over 550+ indexed knowledge entries for "what do I know about this," plus a native SQLite FTS5 log for "what did I actually do about it last week." Every prompt queries both. Every tool call writes back. The system builds institutional memory as a side effect of doing the work.

20+ learned attribution rules. 95% accuracy, up from 50%. That isn't a model. That's institutional memory, finally captured.
WHAT'S INSIDE

Six coordinated layers, all running automatically.

Cascade is a runtime that coordinates six loosely-coupled components. Everything below fires on every session without asking.

  • ·Intent router — classifies the prompt and dispatches to the right specialist
  • ·Hybrid memory — TF-IDF for semantic recall, SQLite FTS5 for episodic recall, queried on every prompt
  • ·Specialist agents — 142 scoped agents handling narrow domains with depth
  • ·Skills library — 225 callable skills across operations, design, infrastructure, and productivity
  • ·MCP integrations — 8+ live production systems across the GTM stack (CRM, MA, revenue intelligence, community, calendar, docs, support, and more)
  • ·Hook runtime — session tracking, knowledge indexing, drift detection, pattern learning, all in the background
OUTCOME

A daily operating system, not a demo.

Cascade runs every workday. Real operational work, real corrections, real memory. The stats on the homepage aren't marketing numbers — they're the system's own runtime metrics, read out of its observations database.

The quiet win is what the work becomes. An hour and a half a day that used to go to classification, context-gathering, and the ten systems of record shows up somewhere better: strategy, higher-leverage judgment, the calls and conversations a human should still be in. The execution got automated so the work could get more strategic.

▸ stack
claude codenodepythontf-idfsqlite fts5mcp
▸ clone it
▸ public repo · ready
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.
▸ connects to
NEXT · 02 / 06
Customer Intelligence Hub
Quarterly Business Reviews, generated on demand.