AlphaVoice

A translation layer between quant models and human conversation.

AlphaVoice is a self-hosted MCP server (spec 2025-11-25, Streamable HTTP) that turns an Alexa+ agent into a quant research desk. Speak a market question, get a spoken answer with real numbers behind it.

Data flow

Utterance becomes an Alexa+ agent call, the agent selects tools over MCP schemas, the engine runs, a speakable summary comes back with the detail payload one question away, and the summary is spoken.

Related MCP server: Rekko MCP

Architecture

Four tools

• signal_scan: scans six signal families and combines them with a ridge combiner to rank trade ideas.

• mispricing_check: prices full-book VWAP across Polymarket and Kalshi, sizes positions with fractional Kelly, paper only. It never executes.

• news_microstructure: fits a Hawkes process to news-to-price impact. It reports alpha_10 (impact magnitude), half-life ln(2)/beta, and whether the regime is sub-critical or explosive.

• market_brief: the composed multi-tool path. Runs all three tools on the synthetic fixtures and synthesizes one spoken verdict, the centerpiece of the demo.

Four design decisions

1. MCP, not a custom API. The server follows the MCP spec 2025-11-25 over Streamable HTTP. It is self-hostable on free infra, so the agent talks to it through a standard protocol instead of a bespoke endpoint.

2. Narrow typed tool interfaces. Each tool takes a small set of typed inputs and returns two things: a speakable summary and a detail payload. A voice answer gets a 30-second budget, and the detail is always one question away.

3. Honest numbers or silence. Every output exposes confidence and freshness. When the data cannot support a claim, the system says so out loud instead of inventing one.

4. Designed for ears, not eyes. Speech rules govern every answer: never more precision than the data earns, confidence stated qualitatively, and every claim ends with an offer for more detail.

Quickstart

AlphaVoice wraps two existing quant engines, so clone them as siblings first:

mkdir -p ~/workspace
git clone https://github.com/RosarioM123/signal-research-lab.git ~/workspace/signal-build
git clone https://github.com/RosarioM123/odds-prediction-mispricing.git ~/workspace/odds-prediction-mispricing

(Or point SIGNAL_ROOT and ODDS_ROOT at wherever yours live.)

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run the server (defaults to port 8000):

.venv/bin/python -m server.main

Open the browser demo:

open web/index.html

Run the test suite:

pytest

Repo map

alphavoice/
  README.md            this file
  LICENSE              license choice (see top-of-file note)
  docs/
    architecture.svg   system diagram: voice to spoken answer
    adr/               architecture decision records for the five design decisions
    alexa-integration.md production path from the web sim to a real Alexa+ agent
    demo-script.md       60-second demo script
    friction-log.md      builder friction log (template)
    product-feedback.md  DRAFT feedback on the MCP SDK
  evals/               tool-routing eval harness (questions, router, report)
  server/              the MCP server and tool implementations
  tests/               engine and contract tests
  web/                 browser simulation of the voice experience

What is new in the hackathon window

The quant engines pre-date this project: the signal families, the ridge combiner, the mispricing math, and the Hawkes fitting already existed. Built during the hackathon window: the MCP server itself, the four tool schemas (including the composed market_brief path), the voice interaction design, the summary/detail contract, the speech rules, the routing eval harness, and the demo. The code in server/, the tests in tests/, the evals in evals/, and the simulation in web/ are the evidence.

Honesty notes

• Any simulated or synthetic data in the demo is labeled as simulated.

• mispricing_check is paper only. Nothing in this repo places real trades or executes orders.

• The web experience is a simulation of the voice interaction, labeled as simulated.