Skip to content
View harshbhx's full-sized avatar
🌐
focusing
🌐
focusing
8 followers · 13 following

Achievements

Achievement: QuickdrawAchievement: YOLOAchievement: Pull SharkAchievement: Pair Extraordinaire

Achievements

Achievement: QuickdrawAchievement: YOLOAchievement: Pull SharkAchievement: Pair Extraordinaire

Highlights

Organizations

@dualmindlabs

Block or report harshbhx

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
harshbhx/README.md

Harsh Bhanushali — co-founder of DualMind Labs. “I build AI agents & automations — and the products around them.” CS student at Parul University in Vadodara, India. In 2025 I co-founded DualMind Labs with Rishi A. Bangoriya, and our first product — DualMind Arena, blind head-to-head AI model evaluation — is live. Open to work, worldwide: freelance, contract, full-time, internships · remote · NDA-friendly. Vadodara, Gujarat, India · IST (UTC+5:30). In this issue: No. 1 The Arena story · No. 2 Selected work · No. 3 Working together · No. 4 The toolbox · No. 5 Letters. Links to dualmindlab.tech.

01 Arena — jump to No. 1, The Arena story 02 Work — jump to No. 2, Selected work 03 Hire — jump to No. 3, Working together 04 Toolbox — jump to No. 4, The toolbox 05 Letters — jump to No. 5, Letters

Email me · LinkedIn · dualmindlab.tech

No. 1 — The Arena story: what we run in production

Cover story, DualMind Labs: DualMind Arena, live. Blind, head-to-head AI model evaluation. Public AI benchmarks are easy to game: cherry-picked prompts, brand bias, results nobody can audit. Two models answer the same real prompt with their names hidden; the community votes before the identities are revealed, and the votes feed a live leaderboard. Box score: 5,000+ requests absorbed in one traffic spike (Bots + real users; Cloudflare CDN took the load before it reached the origin.); #1 on Google for “DualMind Labs” (dualmindlab.tech); Real-time login anomaly detection (IP capture, risk scoring and an email alert to the account owner on every suspicious event.); Docs full user documentation, published (docs.dualmindlab.tech). Next: Expanding from model evaluation into AI agent evaluation. Visit arena.dualmindlab.tech. Fig. 2 — How a battle works, drawn as a request path: 1. User prompt (one real prompt, two models) → 2. Cloudflare CDN (absorbs traffic spikes) → 3. .NET Web API (stateless · scales out) → 4. Cloudflare AI Gateway (rate limits · cache · logs) → 5. Model A / Model B (identities hidden) → 6. Community vote (on the two answers) → 7. Supabase · PostgreSQL (the live leaderboard). Why it is built this way: CDN first, stateless API — Spikes land on Cloudflare, not the origin; scaling the .NET API out is a config change. AI Gateway over direct calls — One place for rate limits, caching and per-request logs across every model provider. Supabase on purpose — Leaderboard ranking needs relational queries. Names are revealed only after the vote. Docs: docs.dualmindlab.tech.

No. 2 — Selected work: what I built, and why

Feature 1 of 5: PixelProof — Chrome extension · AI image detection. Status: built for Next Byte Hacks V2. Is this image real? A Chrome extension that answers while you browse. Flags AI-generated and manipulated images on any page, explains the verdict in plain English and cross-checks the claim against live fact-check data. Fig. 3, how a scan works: 1. Detect — Reality Defender API; 2. Explain — Google Gemini, in plain English; 3. Cross-check — Google Fact Check API. Stack: Chrome MV3, Vanilla JS, Gemini, Reality Defender, Fact Check API. Link: pixelproof.dualmindlab.tech. Feature 2 of 5: ModVigil — Reddit moderation automation · Devvit. Status: built for the Reddit Mod Tools Hackathon 2026. A second pair of eyes for Reddit moderators. Turns a subreddit’s raw mod log into a live dashboard and runs the routine safety work automatically — humans still make every final call. Fig. 4, what it automates: Spike alerts — modmail when removals hit 3× the hourly baseline (checked every 15 min); Progressive warnings — educational note → final warning → flagged for human review; Ghost-mod detection — surfaces moderators inactive for 30+ days; Weekly digest — week-over-week report every Monday. Stack: Reddit Devvit, TypeScript, React, Redis. Link: github.com/harshbhx/modvigil.

Feature 3 of 5: AI Component Builder — Generative UI · Cursor SDK. Status: built at a Cursor workshop in Rajkot. Describe a component in plain English; get working React back. A chat panel streams generated React components from the Cursor SDK into a live preview with syntax-highlighted code. Fig. 5, from prompt to preview: 1. Prompt — chat panel; 2. Generate — Cursor SDK, streamed over SSE; 3. Preview — sandboxed iframe + code view. Stack: Next.js 14, TypeScript, Cursor SDK, SSE, Tailwind. Link: github.com/harshbhx/cursor-rajkot-workshop. Feature 4 of 5: AI Support Agent — Support automation · function calling. Status: built. Answers the routine questions; hands people the rest with context. Claude / GPT-4o with function calling classifies every inbound query by intent and urgency, resolves the routine ones and escalates the rest with the context already drafted. Fig. 6, how a query is routed: 1. Classify — intent + urgency; 2. Resolve — auto-reply, logged to the CRM; 3. Escalate — human review, context pre-drafted. Stack: Node.js, Claude, GPT-4o, Supabase, Webhooks.

Feature 5 of 5: MealMind — Restaurant ecosystem · 3 Flutter apps. Status: built. One payment, three apps, under a second. Customer, chef and admin apps on one Firebase backend: a confirmed Razorpay payment reaches the kitchen screen and the admin dashboard in under a second. Fig. 7, one payment, every screen: 1. Pay — Razorpay, in the customer app; 2. Confirm — webhook → Cloud Function; 3. Sync — one Firestore write, every app — reaching the customer, chef and admin apps in under a second. Stack: Flutter ×3, Firestore, Firebase Auth, Cloud Functions, Razorpay. House rules — how I build. 1. Ship early: Real users surface problems planning never will. 2. Observability is architecture: If you can’t see it in production, it isn’t done. 3. Pick the right tool: The trade-off matters more than familiarity. 4. Own the outcome: The deliverable is the running, monitored system — not the commit. 5. Communicate clearly: Async first, unprompted updates, no surprises. — Harsh, Vadodara, India.

No. 3 — Working together: freelance · remote · under NDA

Freelance, remote, under NDA. I take on AI and product work for clients worldwide — scoped up front, shipped in weekly steps. Services and typical timelines: AI agents & automations (1–4 weeks) — RAG, agents and multi-model systems on OpenAI + Claude, behind Cloudflare AI Gateway; Mobile apps (2–6 weeks) — Flutter for iOS & Android — auth, Firebase, Razorpay / Stripe, store release; Backend APIs (1–3 weeks) — .NET Web API or Node.js — auth, security alerting, fully documented; Full-stack MVPs (3–8 weeks) — Mobile + backend + AI + security, deployed on Azure + Cloudflare; how I work: NDA-friendly — Happy to sign one — client work stays private. Remote, worldwide — IST (UTC+5:30), async-friendly, replies within 24 hours. Your code, delivered — Weekly deliverables and full source handoff, no scope surprises. Email me about a project: harshbhx@gmail.com. /now — what I’m doing now. Building: AI agents & automations — my Gen AI focus. At DualMind Labs: Taking DualMind Arena from model evaluation into AI agent evaluation. Studying: Computer Science at Parul University, Vadodara — after a diploma from Darshan University. Open to work, worldwide — freelance first, any time zone: Freelance, Contract, Full-time, Internships; remote · NDA-friendly · Vadodara, India. Email me about a project: harshbhx@gmail.com · IST (UTC+5:30).

No. 4 — The toolbox: what I ship with, and how often

The toolbox — index, part 1 of 2, A–Z: every tool I ship with, and the work in this issue it appears in. AI (models, gateways, SDKs): AI Gateway (Cloudflare) — in DualMind Arena (Fig. 2); Claude — in AI Support Agent (Fig. 6); Cursor SDK — in AI Component Builder (Fig. 5); Gemini (Google) — in PixelProof (Fig. 3); OpenAI — in AI Support Agent (Fig. 6). Build (languages & frameworks): C#; Dart — in MealMind (Fig. 7); Flutter — in MealMind (Fig. 7); JavaScript — in PixelProof (Fig. 3); .NET — in DualMind Arena (Fig. 2); Next.js — in AI Component Builder (Fig. 5); Node.js — in AI Support Agent (Fig. 6); React — in ModVigil (Fig. 4) and AI Component Builder (Fig. 5); TypeScript — in ModVigil (Fig. 4) and AI Component Builder (Fig. 5). The toolbox — index, part 2 of 2, A–Z: every tool I ship with, and the work in this issue it appears in. Data (where state lives): Firebase — in MealMind (Fig. 7); PostgreSQL — in DualMind Arena (Fig. 2); Redis — in ModVigil (Fig. 4); Supabase — in DualMind Arena (Fig. 2) and AI Support Agent (Fig. 6). Run (cloud, edge, delivery): Azure; Chrome MV3 — in PixelProof (Fig. 3); Cloudflare — in DualMind Arena (Fig. 2); Devvit (Reddit) — in ModVigil (Fig. 4); Docker; GitHub Actions; Workers (Cloudflare).

My GitHub numbers: all-time contributions; commits, pull requests, issues and repositories over the last 12 months; and streaks, refreshed daily Languages across my repositories by code size, as counted by GitHub, refreshed daily

My GitHub contribution calendar for the last 12 months, refreshed daily

No. 5 — Letters: by email, from any time zone

Freelance or NDA project in mind? Write to me. “I’m open to remote work worldwide — freelance first, contract and full-time too — and happy to work under NDA. Tell me what you’re building.” — Harsh. Email harshbhx@gmail.com — opens a pre-filled email about a role. An airmail envelope addressed to Harsh Bhanushali, Vadodara, Gujarat, India; also linkedin.com/in/harshu-dev, dualmindlab.tech and Discord harshu911. The stamp is Harsh’s portrait over “Vadodara”; the postmark reads “Open to work, worldwide”, IST (UTC+5:30).

harshbhx@gmail.com
LinkedIn · Dev.to · Discord: harshu911

Where I am — Vadodara on GitHub’s interactive map 
{
  "type": "FeatureCollection",
  "features": [
    {
      "type": "Feature",
      "geometry": {
        "type": "Point",
        "coordinates": [
          73.1812,
          22.3072
        ]
      },
      "properties": {
        "title": "Harsh Bhanushali",
        "description": "Vadodara, Gujarat, India · open to work, worldwide",
        "marker-color": "#C9643C",
        "marker-symbol": "star"
      }
    }
  ]
}
Loading

Colophon: reprinted daily by GitHub Actions, with a live status check of dualmindlab.tech, the Arena, the docs and PixelProof

Plain-text version

Harsh Bhanushali — Co-founder, DualMind Labs · Vadodara, Gujarat, India

I build AI agents & automations — and the products around them.

CS student at Parul University in Vadodara, India. In 2025 I co-founded DualMind Labs with Rishi A. Bangoriya, and our first product — DualMind Arena, blind head-to-head AI model evaluation — is live.

  • Focus: Generative AI — AI agents & automations
  • Open to work, worldwide: Freelance, Contract, Full-time, Internships · remote · NDA-friendly
  • Based in: Vadodara, Gujarat, India · IST (UTC+5:30)

No. 1 — The Arena story: DualMind Arena (live at arena.dualmindlab.tech)

Public AI benchmarks are easy to game: cherry-picked prompts, brand bias, results nobody can audit. Two models answer the same real prompt with their names hidden; the community votes before the identities are revealed, and the votes feed a live leaderboard.

  • 5,000+ requests absorbed in one traffic spike — Bots + real users; Cloudflare CDN took the load before it reached the origin.
  • #1 on Google for “DualMind Labs” — dualmindlab.tech
  • Real-time login anomaly detection — IP capture, risk scoring and an email alert to the account owner on every suspicious event.
  • Docs: full user documentation, published — docs.dualmindlab.tech
  • CDN first, stateless API: Spikes land on Cloudflare, not the origin; scaling the .NET API out is a config change.
  • AI Gateway over direct calls: One place for rate limits, caching and per-request logs across every model provider.
  • Supabase on purpose: Leaderboard ranking needs relational queries.
  • Stack: Azure, .NET Web API, Vanilla JS, Supabase · PostgreSQL, Cloudflare CDN, Cloudflare AI Gateway, GitHub Actions
  • Request path: User prompt → Cloudflare CDN → .NET Web API (stateless) → Cloudflare AI Gateway (rate limits · cache · logs) → Model A / Model B (identities hidden) → Community vote → Supabase · PostgreSQL (leaderboard)
  • Next: Expanding from model evaluation into AI agent evaluation.

No. 2 — Selected work

  • PixelProof — Chrome extension · AI image detection · built for Next Byte Hacks V2 — pixelproof.dualmindlab.tech · Devpost · Source
    • Is this image real? A Chrome extension that answers while you browse.
    • Flags AI-generated and manipulated images on any page, explains the verdict in plain English and cross-checks the claim against live fact-check data.
    • Detect (Reality Defender API) → Explain (Google Gemini, in plain English) → Cross-check (Google Fact Check API)
    • Stack: Chrome MV3, Vanilla JS, Gemini, Reality Defender, Fact Check API
  • ModVigil — Reddit moderation automation · Devvit · built for the Reddit Mod Tools Hackathon 2026 — Source
    • A second pair of eyes for Reddit moderators.
    • Turns a subreddit’s raw mod log into a live dashboard and runs the routine safety work automatically — humans still make every final call.
    • Spike alerts: modmail when removals hit 3× the hourly baseline (checked every 15 min)
    • Progressive warnings: educational note → final warning → flagged for human review
    • Ghost-mod detection: surfaces moderators inactive for 30+ days
    • Weekly digest: week-over-week report every Monday
    • Stack: Reddit Devvit, TypeScript, React, Redis
  • AI Component Builder — Generative UI · Cursor SDK · built at a Cursor workshop in Rajkot — Source
    • Describe a component in plain English; get working React back.
    • A chat panel streams generated React components from the Cursor SDK into a live preview with syntax-highlighted code.
    • Prompt (chat panel) → Generate (Cursor SDK, streamed over SSE) → Preview (sandboxed iframe + code view)
    • Stack: Next.js 14, TypeScript, Cursor SDK, SSE, Tailwind
  • AI Support Agent — Support automation · function calling
    • Answers the routine questions; hands people the rest with context.
    • Claude / GPT-4o with function calling classifies every inbound query by intent and urgency, resolves the routine ones and escalates the rest with the context already drafted.
    • Classify (intent + urgency) → Resolve (auto-reply, logged to the CRM) → Escalate (human review, context pre-drafted)
    • Stack: Node.js, Claude, GPT-4o, Supabase, Webhooks
  • MealMind — Restaurant ecosystem · 3 Flutter apps
    • One payment, three apps, under a second.
    • Customer, chef and admin apps on one Firebase backend: a confirmed Razorpay payment reaches the kitchen screen and the admin dashboard in under a second.
    • Pay (Razorpay, in the customer app) → Confirm (webhook → Cloud Function) → Sync (one Firestore write, every app)
    • Stack: Flutter ×3, Firestore, Firebase Auth, Cloud Functions, Razorpay

House rules

  1. Ship early. Real users surface problems planning never will.
  2. Observability is architecture. If you can’t see it in production, it isn’t done.
  3. Pick the right tool. The trade-off matters more than familiarity.
  4. Own the outcome. The deliverable is the running, monitored system — not the commit.
  5. Communicate clearly. Async first, unprompted updates, no surprises.

No. 3 — Working together (freelance · remote · under NDA)

I take on AI and product work for clients worldwide — scoped up front, shipped in weekly steps.

  • AI agents & automations (1–4 weeks) — RAG, agents and multi-model systems on OpenAI + Claude, behind Cloudflare AI Gateway
  • Mobile apps (2–6 weeks) — Flutter for iOS & Android — auth, Firebase, Razorpay / Stripe, store release
  • Backend APIs (1–3 weeks) — .NET Web API or Node.js — auth, security alerting, fully documented
  • Full-stack MVPs (3–8 weeks) — Mobile + backend + AI + security, deployed on Azure + Cloudflare
  • NDA-friendly: Happy to sign one — client work stays private.
  • Remote, worldwide: IST (UTC+5:30), async-friendly, replies within 24 hours.
  • Your code, delivered: Weekly deliverables and full source handoff, no scope surprises.

Now

  • Building: AI agents & automations — my Gen AI focus
  • At DualMind Labs: Taking DualMind Arena from model evaluation into AI agent evaluation
  • Studying: Computer Science at Parul University, Vadodara — after a diploma from Darshan University
  • Open to: Freelance first · contract · full-time · internships — remote, worldwide, NDA-friendly

No. 4 — The toolbox

  • AI: OpenAI, Claude, Gemini, AI Gateway, Cursor SDK
  • Build: TypeScript, JavaScript, Dart, Flutter, C#, .NET, Node.js, Next.js, React
  • Data: Supabase, PostgreSQL, Firebase, Redis
  • Run: Azure, Cloudflare, Workers, Docker, GitHub Actions, Chrome MV3, Devvit

No. 5 — Letters

Freelance or NDA project in mind? I’m open to remote work worldwide — freelance first, contract and full-time too — and happy to work under NDA. Tell me what you’re building. Email me about a project.

Email: harshbhx@gmail.com · LinkedIn · GitHub · Dev.to · dualmindlab.tech · Discord: harshu911

Pinned Loading

  1. dualmindlabs/DualMind_UI dualmindlabs/DualMind_UI Public

    JavaScript 1

  2. pixelproof pixelproof Public

    Chrome extension that detects AI-generated & deepfake images in real time — with plain English explanations, AI source detection, and one-click fact checking.

    HTML

  3. modvigil modvigil Public

    Real-time moderation analytics and automation for Reddit, built on Devvit: activity heatmaps, spike alerts, progressive auto-warnings and mod health scoring.

    TypeScript

  4. tarang-fm tarang-fm Public

    Tarang FM: an AI radio station in your language. Multi-voice shows and live call-ins, powered by Sarvam Bulbul v4 Flash, Saaras v4 and Sarvam-105B.

    TypeScript