Meet Babariya
Contact
← All work

Product · AI copilot × copy trading on Hyperliquid

QuantPilot

A non-custodial social-trading layer for Hyperliquid perpetuals. It ranks traders by risk- and funding-adjusted performance with a luck filter, lets users copy them without handing over withdrawal rights, and earns a small builder fee on every routed fill.

Role
Solo builder — research, quant, backend, product
Year
2026
Status
Private · demo on request
Links
—

The wedge

Trading front-ends on Hyperliquid earn builder fees on the volume they route. The leaders are Telegram bots and copy-trading apps. When I studied them, they shared three blind spots:

  1. Leaderboards sort by raw P&L, which rewards leverage and luck.
  2. Custody fear: many bots hold user keys.
  3. No AI layer to explain what a trader is actually doing.

QuantPilot targets all three, and charges 0.03% instead of the 0.05% an incumbent charges.

Ranking that survives luck

A trader with a great month is usually lucky. The ranking worker scores accounts on risk-adjusted and funding-adjusted returns (perp funding can quietly dominate P&L), then applies a Deflated Sharpe Ratio filter. That corrects the Sharpe for how many candidates were screened and for non-normal returns, so the top of the board is less likely to be the best of many coin flips. The worker discovers accounts and has run on live mainnet data.

Architecture

Fig. 01 TypeScript product shell, Python quant brain, one schema

TypeScript · product shell

Telegram botgrammY · onboarding, alerts, one-tap copy
Signing web appNext.js · user approves an agent wallet once
Execution APIHono · /trade, /positions, builder-code fills, AES-GCM agent keys

Python · quant brain

Ranking workerrisk- & funding-adjusted metrics, Deflated Sharpe
AI copilotFastAPI · explains traders & positions (next phase)
PostgreSQLone schema, read and written by both languages
Hyperliquid L1 agent wallet can trade · cannot withdraw

Users approve an agent wallet once. It can place trades but cannot withdraw, so QuantPilot never holds custody. Agent keys are encrypted at rest with AES-GCM.

I started in Python and moved the product surface to TypeScript deliberately. Builder revenue scales with routed volume, not latency, so distribution and UX matter more than raw speed. A single TS stack (bot, web, API) on a mature Hyperliquid SDK ships faster. The quant code stayed in Python, where it belongs, and both languages share one PostgreSQL schema. I verified that integration end to end.

Where it stands

  • Built: onboarding and signing, execution with builder code (verified on testnet), and the ranking engine (run on mainnet data).
  • Next: the AI copilot. It explains a trader’s style, exposure and drawdowns in plain language before you copy them.