Jay Rathod

← Selected work

Data pipelines · LLM extraction

MCM: My Crypto Monitor

Backtests whether following crypto influencers on YouTube and Telegram would have made money or lost it.

Status
Live platform
Where
JMD Tech Solutions · Sep 2023 – Dec 2025
My role
Co-developer of the core backend and AI pipeline (team of two)
40,000+
YouTube videos and transcripts collected
100,000+
Telegram messages processed
100+
influencers discovered
8
return windows, 1 hour to 1 year

Overview

Crypto influencers make a lot of calls, and few people check how those calls actually turned out. MCM collects what they said, extracts each recommendation and measures what it would have returned. I co-developed the backend and AI pipeline in a team of two and worked directly with the client to plan features and how to build them.

Pipeline (refreshes about every 3 hours)
  1. 01 · Discover Playwright 100+ influencers across YouTube and Telegram
  2. 02 · Collect Videos · transcripts · messages Handles rate limits and dynamic content
  3. 03 · Store MongoDB Schema I designed
  4. 04 · Extract DeepSeek Recommendations pulled from text
  5. 05 · Price Binance · CoinGecko Historical and live prices
  6. 06 · Score ROI engine 8 windows; with and without 'hyperactive' coins

What I did

  • Built influencer discovery workflows with Playwright.
  • Built collection pipelines for YouTube videos, transcripts and Telegram channels, and designed the MongoDB schema.
  • Developed LLM-based extraction of recommendations from transcripts and messages.
  • Implemented the core ROI engine on Binance and CoinGecko price data.
  • Ran periodic manual validation in Excel and cross-checked the algorithm's maths with investors.

Engineering decisions

  1. Pick the model on evidence

    I compared DeepSeek, Gemini and GPT on output quality and cost and chose DeepSeek. I validated its extractions against the source videos and refined prompts where context was missing.

  2. Separate the moonshots

    A few 'hyperactive' coins with moonshot gains can make an influencer look brilliant. Results are shown both with and without them.

  3. Near real time, honestly

    Pipelines refresh on a cycle of about three hours, the practical limit for collecting and processing 100+ influencers.

Stack

  • Python
  • Playwright
  • MongoDB
  • DeepSeek
  • Binance API
  • CoinGecko API
  • Excel