Delhi / Jaipur | Applied AI Engineer | Building production GenAI systems across voice AI, RAG, memory, and MCP

I build AI systems that remember context, use tools, and survive real users. My work sits at the product layer of AI: clear jobs, backend reliability, useful evaluation, and workflows that make people sharper instead of hiding the thinking.

• Highlyt - AI-native reading and annotation product that turns PDFs, EPUBs, web articles, YouTube transcripts, and Kindle highlights into a connected knowledge graph queryable from Claude and ChatGPT through MCP.
• Rehearsal AI - Voice-powered interview preparation platform at Gradeless AI. I work on question delivery, voice interviews, scoring, feedback reports, RAG, memory, and backend reliability across 10+ institutions.
• Context Hub - Personal context layer for Claude, ChatGPT, Perplexity, Cursor, and other MCP clients. Deploys to Cloudflare Workers and keeps memory portable across tools.
• Multicast - MCP gateway that fans out tool calls across multiple HTTP MCP servers in parallel, reducing repeated model thinking cycles and making remote tool use faster.

• voicenotes-mcp - Custom MCP server for Voicenotes with search, create, edit, tags, stdio/HTTP support, OAuth 2.1 PKCE/DCR, and token auth.
• context-hub - Shared AI context layer across MCP clients, running on Cloudflare Workers and D1.
• core-sending-lab - Local email delivery simulator for queues, workers, retries, throttling, and delivery debugging.
• MCP Python SDK PR #3066 - Preserves query components in RFC 9728 protected-resource metadata URLs and resource validation.
• resend-python - Working through email infrastructure and SDK contribution paths.

• Highlyt mobile and web - Built document upload, highlighting, annotations, semantic color systems, graph-based idea linking, secure Supabase access, and MCP server variants.
• Rehearsal backend systems - Maintained FastAPI services, Supabase RLS, rate limits, CORS, OpenTelemetry, ARQ workers, stuck-recording recovery, report generation, and production/staging release checks.
• Jaipuria Intelligence - Academic data assistant using text-to-SQL, semantic search, query routing, caching, and natural-language exploration of institutional content.
• AI course generation pipeline - Planning, review, lesson generation, verification, and credit tracking with FastAPI, Next.js, Supabase, and multi-model LLM workflows.
• Memory infrastructure - Moved interview memory from Mem0 Cloud to a self-hosted Render + pgvector service while preserving memory categories and reducing infrastructure cost.
• CareAI - Flask and scikit-learn medical advisory system with symptom prediction, voice input, and health guidance.
• MelodyNet - CNN music genre classifier using mel-spectrogram preprocessing, TensorFlow, Flask, React, and Render deployment.

• Medium - Writing on LLM engineering, Claude workflows, AI-assisted development, voice AI, MCP, and production GenAI.
• Fieldwork - Substack field reports from building and shipping AI in production.
• Recent themes - context rot, MCP security, human-in-the-loop AI, learning systems, AI product reliability, agent evaluation, and engineering judgment.
• Published in - Towards AI, FAUN.dev, Built at Rehearsal, Bootcamp, CodeToDeploy, and personal publications.

• Building Highlyt - Turning reading artifacts into a typed knowledge graph that AI tools can query without losing source context.
• Building Rehearsal AI - Shipping voice-powered interview systems where feedback, scoring, memory, and follow-up questions improve real practice.
• Building MCP infrastructure - Context Hub, Multicast, Voicenotes MCP, and production OAuth/MCP patterns.
• Improving backend reliability - Async database paths, queue workers, RLS, observability, rate limits, and safer release workflows.
• Writing in public - Documenting the messy reality of GenAI product engineering.
• Contributing to open source - Testing MCP SDK behavior from real production server experience.

• Microsoft Certified: Azure AI Fundamentals.
• Meta Database Engineer Specialization.
• 3rd rank in university coding competition among 100+ participants.
• B.Tech in Computer Science and Engineering, AI specialization, JK Lakshmipat University.
• Recommendations from mentors and managers for technical ability, ownership, and reliability.

AI does not remove engineering judgment. It makes bad judgment show up faster.

Random facts:

• I care more about clear AI jobs than flashy demos.
• I like boring reliability work because users feel it immediately.
• I think memory systems need aggressive forgetting, not bigger context windows.
• I would rather build the feedback loop than collect another AI tool.
• The best AI products should make the user smarter, not just faster.