🪙 CoinFlow — I Built a Full-Stack Micro-Tasking Platform with GitHub Copilot.

Redwan Shahriar Shubho — Fri, 05 Jun 2026 18:19:01 +0000

💡 What I Built

CoinFlow is a full-stack MERN micro-tasking platform with a real coin economy, Stripe payments, Firebase authentication, and a three-role system for Workers, Buyers, and Admins.

Live Demo: coin-flow-peach.vercel.app
GitHub Repo: github.com/redwanshahriarshubho/CoinFlow

📸 Before & After

❌ Before — No Backend At All

The entire server/ folder had no index.js. Every API call failed with network errors. The frontend looked beautiful but nothing worked.

✅ After — Fully Working Live App

🛠️ Tech Stack

Frontend: React 19, Vite 8, React Router 7, TanStack Query v5, Firebase v12, Stripe.js, Swiper, React Hook Form

Backend: Node.js, Express 5, MongoDB (native driver), JWT, Stripe SDK

Infrastructure: Vercel (frontend), Render (backend), MongoDB Atlas, imgBB

🧩 How It Works — The Three Roles

👷 Worker

Workers join the platform and get 10 free coins on signup. They browse available tasks, submit completed work, and get paid instantly when approved. Once they hit 200 coins ($10), they can withdraw via bKash, Nagad, Rocket, or Bank Transfer.

🏢 Buyer

Buyers post tasks with a coin reward per worker — the total cost is deducted when the task goes live. They review submissions through a dashboard and approve or reject each one. Unused coins are refunded automatically if a task is deleted.

🔐 Admin

Admins see platform-wide stats, manage all users, moderate any task, and approve worker withdrawal requests.

💰 The Coin Economy

Action	Coins
Register as Worker	+10 free
Register as Buyer	+50 free
Post a task	-(workers × coins_per_worker)
Submission approved	+payable_amount to worker
Delete task (unused slots)	+refund to buyer
Withdrawal minimum	200 coins ($10)
Exchange rate	20 coins = $1 USD

One design decision I'm proud of: when a submission is rejected, the task's required_workers count is incremented back by 1 so another worker can claim that slot. Without this, a single rejection would permanently block an open position.

app.patch("/submissions/:id/reject", verifyToken, verifyBuyer, async (req, res) => {
  await submissions.updateOne(
    { _id: new ObjectId(req.params.id) },
    { $set: { status: "rejected" } }
  );
  await tasks.updateOne(
    { _id: new ObjectId(sub.task_id) },
    { $inc: { required_workers: 1 } }
  );
  await notifications.insertOne({
    recipient_email: sub.worker_email,
    message: `Your submission for "${sub.task_title}" was rejected. Keep going!`,
    actionRoute: "/dashboard/task-list",
    time: new Date(), read: false,
  });
});

🤖 How GitHub Copilot Helped

Architectural Review

I used Copilot as a sounding board throughout. Describing the coin economy rules in plain English and asking "what could go wrong here?" surfaced edge cases before they became bugs.

Security Hardening

Copilot flagged an IDOR vulnerability — I wasn't verifying that the buyer making a request actually owned the task. It suggested adding buyer_email: req.user.email to every MongoDB mutation query:

await tasks.updateOne(
  { _id: new ObjectId(req.params.id), buyer_email: req.user.email },
  { $set: { task_title, task_detail, submission_info } }
);

MongoDB Index Design

Copilot reviewed my schema and recommended the full index set:

await users.createIndex({ email: 1 }, { unique: true });
await tasks.createIndex({ buyer_email: 1 });
await submissions.createIndex({ worker_email: 1 });
await notifications.createIndex({ recipient_email: 1, time: -1 });

Duplicate Submission Guard

const dup = await submissions.findOne({ task_id, worker_email });
if (dup) return res.status(400).json({ message: "You already submitted this task" });

✨ Features Shipped

Worker: Browse tasks, full detail page, submit work, paginated history, real-time notifications, withdrawal form

Buyer: Post tasks with live cost preview, review submissions via modal, approve/reject with instant worker payout, edit/delete with refund, purchase coins via Stripe, payment history

Admin: Platform stats, user management, task moderation, withdrawal approval queue

System: JWT + Firebase auth (Email + Google OAuth), role-based route guards, imgBB image upload, fully responsive UI

🚀 Try It Live

coin-flow-peach.vercel.app

Register as a Worker to browse and complete tasks, or as a Buyer to post tasks and review submissions.
Test Stripe with card 4242 4242 4242 4242 / any future date / any CVC.

Built with React, Node.js, MongoDB, Stripe, and Firebase. GitHub Copilot helped me think through security, edge cases, and performance throughout the build.

Nexus

Redwan Shahriar Shubho — Sun, 31 May 2026 04:57:14 +0000

This is a submission for the Hermes Agent Challenge

What I Built

Nexus is an autonomous research digest agent powered by Hermes.
In a single command, Nexus plans a search strategy, searches the web, reads the top sources, clusters findings into themes, and generates a clean HTML digest with citations and a what to read next queue. No hand holding. No prompting for each step.

Demo

$ python main.py AI safety and alignment

Nexus Research Digest
Topic: AI safety and alignment
Date: 2026-05-30
Model : openrouter/auto

[step 1] Thinking... Tool: search_web
[step 2] Thinking... Tool: search_web
[step 3] Thinking... Tool: fetch_url
[step 4] Thinking... Tool: cluster_findings
[step 5] Thinking... Tool: save_digest

Digest saved
Done. Check the output/ folder.

Code
GitHub: https://github.com/redwanshahriarshubho/Nexus
My Tech Stack
• Python 3.10+
• Hermes Agent via OpenRouter API
• scikit-learn (TF-IDF clustering)
• urllib (web search + fetching)
• Custom HTML renderer (no frameworks)

How I Used Hermes Agent

Nexus uses three core Hermes agentic capabilities:

Multi-step planning: Hermes doesn't just answer one prompt. It plans which search queries to run, decides which URLs to fetch, and sequences its own actions autonomously.

Tool use: Four custom tools are registered: search_web, fetch_url, cluster_findings, and save_digest. Hermes decides when to call each one and with what arguments — no hardcoded flow.

Autonomous completion: The agent runs for up to 20 steps, self-corrects when tools return errors or results are sparse, and always finishes by saving the digest.

The result is a clean HTML report with a summary, key findings, themed sections, citations, and a what to read next queue generated entirely by Hermes reasoning through the problem step by step.

DEV Community: Redwan Shahriar Shubho