QQuantprims

About Quantprims

Quantprims is an infrastructure that exposes verified financial-calculation primitives (Sharpe, max drawdown, walk-forward, Monte Carlo, optimization, and more) to AI agents like Claude Desktop, Cursor, and Codex through an MCP server. A public REST API that lets any AI client or framework call the same primitives is in active development; the launch date will be announced separately.

Why I built it

One day, while validating FX strategies on a separate project (fx-signal), I noticed that Claude wrote a slightly different Sharpe ratio or max drawdown implementation every single time I asked. Even for Sharpe alone, pyfolio, empyrical, and ffn use subtly different formulas, and walk-forward optimization shifts numerically depending on anchored vs rolling vs purged.

Rather than letting the AI keep recalculating, I decided to ship the verified calc infra that the AI calls. I pivoted the FX backtest engine I had spent 9 months on (with 303 unit tests and multiple rounds of Codex adversarial review) into an MCP server exposed to AI builders. A public REST API for arbitrary AI clients and frameworks is in active development and will be announced separately.

How we got here (transparency)

Solo side project. The path has narrowed a few times:

  • Mid-2025 through early May 2026: built and operated a browser-only FX backtest site (fx-backtest.com), with no-code sample data + CSV upload so beginners could try it.
  • 2026-05-14: briefly pivoted to a Tauri native desktop app (GUI-first mental model).
  • 2026-05-16: locked in MCP-first with a free desktop GUI as a helper (aligned with the AI builder target).
  • 2026-05-18: focused on AI calc infrastructure positioning. Dropped the desktop GUI (after confirming that the founder, as the #1 customer, mostly uses it through AI), shipped the MCP server first, and put the public REST API into active development (Pattern C).

This may look like a flurry of pivots, but it is the validation phase with zero data on the ground. The existing web build (/simulator) is frozen for existing-user continuity, and new development focuses on the MCP server (available in trial) and the public REST API (in active development).

What we offer, what we don't

This service is an infrastructure for calculation primitives over historical data. It does not recommend specific symbols, currency pairs, or trade timings. It is designed not to fall under investment advisory or agency business as defined in Article 28, Section 3 of the Japanese Financial Instruments and Exchange Act. A lawyer-signed opinion confirming non-applicability is in progress.

  • What we offer: 8 calc primitives (run_backtest, calc_sharpe, calc_maxdd, run_walkforward, run_montecarlo, optimize_grid, optimize_genetic, compare_strategies) through the MCP server, returning numbers only, with audit logs and reproducibility. Access via the public REST API will be announced after development completes.
  • What we do not offer: buy/sell signals, personalized investment advice, data feeds (OHLC arrays are user / AI agent input), investment solicitation, or referrals to specific brokers.

Investment decisions are the user's (or the user's AI agent's) responsibility.

Founder = primary customer

The founder is the #1 customer of Quantprims. I use this infrastructure myself every day to automate FX strategy validation through Claude and Codex, and I share the results on social media (dogfood narrative).

Even with zero external paid users, the founder uses it daily — so this is a structurally stable path, not a failed pivot. At the same time, we are opening the same infra to the AI builder community.

Current launch status: self-serve Stripe checkout and an auto-issuing license dashboard are in development. During early access, the founder issues licenses by hand. The planned pricing (7-day free trial, then ¥980/month) will carry over, but actual paid plan enrollment will only apply to users who complete the official checkout flow after self-serve goes live.

Developer

jissocyu (実装中, jissōchū) — a solo side-project developer. The handle is wordplay on 「実装中」 ("in the middle of implementation").

Tech stack

  • Calc core (shared/): TypeScript ESM nodenext. Indicators, backtest engine, statistics, DSL, walk-forward, Monte Carlo, plus genetic / grid optimization.
  • MCP server: Node 22 + @modelcontextprotocol/sdk + stdio protocol, 8 tools.
  • REST API: AWS Lambda (Node 22) + API Gateway REST v1 + Cognito Identity Pool + IAM auth + WAF.
  • Web dashboard: Next.js 14 App Router + TypeScript + Tailwind CSS + shadcn/ui.
  • Auth and billing: Cognito + Stripe + DynamoDB (License / UserLicenseSentinel / processed-events tables).
  • Hosting: AWS Amplify Hosting in the Tokyo region.

Kill criteria (transparency)

As a solo side project, I have drawn a hard "pivot or shut down if there is no result in 3 months" line:

  • 2026-06-17 (1 month): 5 external design partners actively using Quantprims via Claude / Cursor / ChatGPT — partial adoption maintained, recruiting strengthened.
  • 2026-07-17 (2 months): 3 paid pilots OR ¥10k MRR — otherwise narrow to dogfood / free maintenance only.
  • 2026-08-14 (3 months): 10 paid users OR ¥30k MRR + 30 weekly MCP runs — if not met, paid MCP is withdrawn and only free maintenance continues.

Service information