SignalX

SignalX is a local-first startup and finance workbench with three workflows:

• Ideate for shaping rough ideas and reasoning
• Evaluate for structured pressure-testing and reports
• Expert for domain discussion, concept learning, pre-screening, and deck analysis

It is built to feel more like a sharp operator than a generic chatbot:

• it stays two-way instead of dumping generic advice
• it shows evidence and grounded reasoning in Expert
• it supports both open-source local runtime and API-key runtime
• it stores sessions locally so users can resume work
• it can turn an Ideate session into a refined pitch artifact

Quick Start

Clone-and-run with API keys

Best path if you want the easiest local setup and prefer providers like Groq, Cerebras, OpenAI, OpenRouter, Anthropic, or Gemini.

git clone <your-repo-url>
cd signalx
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
npm install
npm --prefix frontend install
cp .env.example .env
npm run mvp:api

Open http://127.0.0.1:7860, choose Use API key, then pick your provider.

Run fully local with open-source models

If you want the app to stay local and avoid paid APIs, use the same setup steps above, then:

ollama pull llama3.2
# Optional sharper local model if your machine can handle it:
ollama pull qwen3:8b
npm run mvp

SignalX will auto-start Ollama if it is installed and not already running.

What You Get

Ideate

• open-ended discussion
• sharper framing of problem, customer, wedge, and pitch
• uploads and session resume
• refined pitch generation from the conversation

Evaluate

• Idea review for adaptive pressure-testing of a startup idea
• Deck review for direct .pdf and .pptx pitch-deck assessment inside Evaluate
• structured reports with verdict, evidence gaps, weak claims, and next steps
• slide/page-aware feedback instead of a generic one-shot chatbot answer

Expert

• concept explanations and comparisons
• evidence-backed answers using the bundled knowledge corpus
• pre-screening and deck analysis
• source, confidence, and analysis panels in the workbench

Runtime Modes

npm run mvp

Single-port local app with open-source models.

• serves frontend and backend together on http://127.0.0.1:7860
• auto-starts Ollama if needed
• good default for solo local use

npm run mvp:api

Single-port local app without Ollama startup.

• serves frontend and backend together on http://127.0.0.1:7860
• optimized for API-key-based usage
• supports groq, cerebras, openai, openrouter, anthropic, and gemini

npm run mvp:lan

LAN sharing for people on the same network.

npm run dev

Frontend and backend split for active development.

• frontend: http://127.0.0.1:5173
• backend: http://127.0.0.1:8000

Providers

Open-source local

• ollama
• recommended local fast model: llama3.2:latest
• optional sharper local model: qwen3:8b

API-key providers

• groq
• cerebras
• openai
• openrouter
• anthropic
• gemini

API keys can be entered in the UI for a session or exported in the shell before launch. For public deployments, set the provider key on the server so visitors do not need to bring their own key.

Example:

export GROQ_API_KEY=...
npm run mvp:api

Fast hosted open-weight defaults:

• Groq: openai/gpt-oss-20b for Fast, openai/gpt-oss-120b for Sharper
• Cerebras: gpt-oss-120b

Frontier OpenAI defaults:

• OpenAI: gpt-5.4-mini for Fast, gpt-5.5 for Sharper

Deck Review

Evaluate now has two modes:

• Idea review
• Deck review

Deck review is meant for serious pitch-deck feedback, not a prompt hack inside chat.

What it does:

• reads uploaded .pdf and .pptx decks into an ordered slide/page artifact
• evaluates the deck against the built-in rubric inside Evaluate
• reports what is working, what is weak, what is unproven, and what is still missing
• cites slide/page references where possible
• marks missing material as not shown, unclear, or unverified instead of guessing

Important runtime behavior:

• if the active model supports vision and the deck has renderable slide/page images, the review can assess slides more directly
• if the active model is text-only, SignalX runs a bounded transcript review from extracted deck text and explicitly avoids fake visual claims
• Qwen2.5-VL is the recommended local Ollama path for stronger deck review
• Gemma 3 is a lighter local fallback when you want multimodal support on a smaller setup

If you update the deck or improve the review pipeline, re-run Deck review to generate a fresh report. Existing saved reports do not automatically rewrite themselves.

Bundled Knowledge

The Expert workbench ships with a bundled JSON corpus under knowledge_base/expert/.

That means:

• clone-and-run does not depend on a private /Users/.../Desktop/data path
• /api/health reports whether the Expert corpus loaded
• local users and hosted deployments read from the same bundled source by default

Project Layout

• frontend/ - React + Vite UI
• backend/ - FastAPI API and services
• knowledge_base/expert/ - bundled Expert corpus
• data/ - local runtime state, sessions, uploads, exports
• memory.py - SQLite persistence
• signalx_app.py - single-port launcher
• docs/ - runbook, architecture, and deployment notes

Deployment

For the current codebase, the cleanest first public deployment is Render.

Why:

• the app is a single long-running React + FastAPI service
• it uses SQLite and uploaded files on disk
• it expects same-origin frontend and API behavior
• the repo already includes render.yaml plus a deployment checklist

Use:

• Execution Guide for local and dev commands
• Deployment Checklist for hosting

Main Commands

npm run mvp
npm run mvp:api
npm run mvp:lan
npm run admin
npm run dev
npm run build
npm run knowledge:vc

Current State

• SignalX is now a three-workflow MVP: Ideate, Evaluate, Expert
• both open-source and API-key runtime paths are supported
• the Expert corpus is bundled in the repo
• local API-key mode works without Ollama
• Evaluate supports both Idea review and structured Deck review
• long analytical answers auto-continue once when providers stop on length
• hosted provider keys can now be configured server-side for public demos
• session history, uploads, and reports persist locally

Docs

• Execution Guide
• Deployment Checklist
• Platform Overview
• Architecture
• Changelog