Case Study
SS-1 — Signal Shell
Market intelligence engine for product teams
LangGraph · FastAPI · Claude API · Firecrawl · React 19 · GitPython · SQLite · Tauri v2
The problem
Product teams are reactive by default. A competitor ships on Tuesday. Someone screenshots it on LinkedIn by Wednesday. A feature request lands in the sprint by Thursday — with no strategic assessment, no user impact analysis, no consideration of whether responding is even the right move.
This is not an information problem. There is always plenty of competitive information available. The gap is between raw signal and structured brief — the twenty minutes of thinking that connects a competitor’s move to your specific users, priorities, and roadmap. That brief is almost never written because nobody has time to write it systematically.
SS-1 writes it. Continuously. Before anyone sees the screenshot.
The concept
A Project Manifest — a git-backed markdown file — defines your product vision, competitive landscape, and target archetypes. SS-1 monitors the external world against it continuously. When something significant changes, it produces a structured intelligence brief: what changed, why it matters relative to your product and users, what the manifest should say now, and what the strategic options are.
Every proposed manifest update requires explicit human approval. The brief is the output. The approval is the moment the user evaluates the agent’s reasoning and owns the decision.
Architectural decisions
Manifest as a markdown file, not a database
A markdown file under git gives you a human-readable document, a meaningful diff, and a commit history that reads as a narrative of how your competitive thinking evolved. A database gives you none of those. When the manifest changes, the git diff tells you exactly what the agent concluded and what the human chose to accept. That history is the product.
HITL approval on every manifest write
An agent that updates your product strategy without asking is noise, not intelligence. The manifest represents your product thinking. If the agent draws the wrong implication, an unreviewed write corrupts your strategic record subtly and invisibly. The approval is not a safety guardrail — it is the product.
Five-node LangGraph pipeline over a single agent
Separating Monitor, Analyst, Significance Gate, Propose, and Apply into distinct nodes makes each step testable in isolation and observable at every checkpoint. The significance gate is the critical node: signals below 40/100 are discarded silently. Users see curation, not noise.
Haiku for routing, Sonnet only for briefs
Brief quality is the core product value. Sonnet has the reasoning depth to connect a competitor’s changelog to a specific archetype’s pain point. Haiku handles classification and scoring — faster, cheaper, and the error modes are recoverable.
The connection to TX-1
SS-1 is the intelligence layer above TX-1’s execution layer. TX-1 acts on failures inside your systems. SS-1 monitors for changes outside them. The interaction model is identical because the problem is identical: something changes, an agent reasons about it, a human approves a response, a record is created.
Together they are a proof of concept for a single thesis: enterprise AI tooling in the next decade will not be better dashboards. It will be systems that understand your context, monitor continuously, and bring decisions — not notifications — to the people who need to make them.
What this demonstrates
A production-ready agentic pipeline with typed state, significance filtering, and HITL approval. Every architectural decision documented with reasoning, alternatives considered, and trade-offs acknowledged. End-to-end ownership from problem identification through deployment. And proof that the TX-1 design language and engineering conventions are a reusable system — not a one-off project.
Status and direction
SS-1 demo is complete and live at ant-ss1-signal-shell.vercel.app. Active development continues on the manifest schema, the significance threshold calibration tooling, and the integration layer that connects SS-1 intelligence output to TX-1 execution actions.