Harsh ViraniHarshVirani

• 01Flag & FillFills every customs field, flags the ones it is unsure of.Agent · 2026Open case→fig. 01AgentA Hong Kong broker typed every declaration by hand. A 12-step LangGraph pipeline now reads the documents, fills the form and flags what a person should check. On a hand-labelled test set, field accuracy rose from 45.8% to about 88%.LangGraphExtractionEvalsdelivered
  01Agent · 2026Flag & Fill→Fills every customs field, flags the ones it is unsure of.
• 02PocketPulseA calm expense tracker, live at pocketpulse.me.Product · 2026Open case→fig. 02ProductAn installable expense tracker that works offline and stays out of your way. Designed, built and shipped in about three weeks.PWANext.jsOfflinelive
  02Product · 2026PocketPulse→A calm expense tracker, live at pocketpulse.me.
• 03Support agentAnswers only when it can cite a source. Otherwise, silence.Agent · 2025Open case→fig. 03AgentA support agent with one rule: grounded or silent. Every answer carries its source, and it says "I don’t know" instead of guessing.RAGFastAPIEvalspublic demo
  03Agent · 2025Support agent→Answers only when it can cite a source. Otherwise, silence.
• 04SSRI platformAn assessment product with real users and real payments.Product · 2025Open case→fig. 04ProductBuilt end to end as a freelancer, from assessments to payments. Live at assessment.silverspring.in, and I still maintain it.Next.jsPostgresPaymentslive
  04Product · 2025SSRI platform→An assessment product with real users and real payments.

• Thesisin preparation

  Agents in the Grid

  Deep-research LLM agents for office work: can an agent do real spreadsheet work reliably, and does that depend more on the model or on the harness around it? A planned study on SpreadsheetBench.

  2026 — 27Read the researchOpen case→

Papers

• P/01Paper · 2025

  The evolving ecosystem of large language models

  Decoder, encoder and encoder-decoder models compared across understanding and generation, with prompting, hallucination and code LLMs.

  Paper · 2025Read →PDF ↓
• P/02Paper · 2025

  AI-augmented low-code and no-code platforms

  How LLM agents bridge engineers and citizen developers: task-technology fit and real workflow integration.

  Paper · 2025Read →PDF ↓

Tap a skill to see where I used it.

1. 2025 — nowSilver SpringFreelance full-stack developer+
   • Built a retirement-readiness assessment platform on my own: questionnaire, automatic scoring, a personalised PDF report for each respondent, payments with discount codes and an admin area.
   • Live with real users and payments. I still fix issues and make changes on request.
   • Stack: Next.js, TypeScript, GraphQL (PostGraphile), PostgreSQL, Razorpay.
2. 2024 — 2025TechStaunch SolutionsSoftware engineer, AI & full-stack+
   • Worked directly with an e-commerce client as the active developer: React and Node.js features, ASP.NET API integration and real-time updates with SignalR.
   • Pages loaded about 40% faster after profiling and refactoring. Set up an AWS pipeline that cut a deployment from about 15 minutes to about 2.
   • Built RAG pipelines and chatbots (LangChain, LlamaIndex, FastAPI) that ran on the client's own servers with no external API calls, for clients with strict data-privacy needs.
   • Co-built the first version of an n8n workflow that found and ranked sales leads for the business development team.

• 2025 — nowM.Sc. AI & RoboticsHof University of Applied Sciences
  • Dean's List 2025 ↗
• 2021 — 2024B.Sc. Computer ScienceVeer Narmad South Gujarat University
  • MYSY Scholarship