I'm an AI Product Manager at Interview Kickstart who doesn't just write specs. I build. I prototype AI products hands-on (full-stack code or n8n/Zapier low-code), ship fast to validate ideas, and then drive the roadmap. My background spans product management, technical recruiting, and software engineering, which means I can scope an LLM feature, evaluate the engineering trade-offs, judge the talent to build it, and own the outcome end-to-end.

What makes me different: Most PMs talk about AI. I build multi-agent systems, RAG pipelines, and multi-modal SaaS products, then translate the build into product decisions, trade-offs, and measurable outcomes. Code or low-code, weekend prototype or production spec.

Associate Product Manager (APM-2) · Interview Kickstart

• Own the Software & Systems pod spanning 7 technical domains: full-stack, frontend, backend, cloud, embedded, test, and system design interview preparation programs
• Developing and scaling AI/GenAI curriculum and capstone projects that become hands-on student builds (multi-agent systems, RAG, NL→SQL)
• Engage directly with Subject Matter Experts (SMEs) across engineering disciplines to keep curriculum aligned with current industry hiring trends and emerging tech (GenAI, LLM tooling, cloud-native patterns)
• Scope end-to-end program requirements: define learning outcomes, design capstone specs, coordinate with SMEs and content teams, and track learner engagement

Repo →  ·  Live Demo →  ·  ⭐ 12   🍴 21

• Problem: Single-LLM chatbots hallucinate on operational queries (order lookups, invoice details) because they lack structured data access and verification
• AI approach: Designed a LangGraph supervisor routing to specialized sub-agents (catalog + invoice), each with parameterized SQL tools. Chose GPT-4o-mini (temperature 0) for deterministic, cost-efficient responses. Added identity verification gate, long-term memory (set-union writes), and human-in-the-loop escalation
• What shipped: Production-grade multi-agent system with 9 tools, 3-tier trust model, and Gradio UI, deployed on HuggingFace Spaces
• Outcome: 28 deterministic tests, CI/CD pipeline, anti-hallucination grounding via tools-only answer sourcing. Community-validated: most-forked AI project in portfolio

Repo →  ·  Live Demo →

• Problem: Call centers review <5% of calls manually. Inconsistent quality scoring, missed compliance violations, no audit trail
• AI approach: Built a 7-node LangGraph pipeline: local Whisper transcription → prompt injection detection (22 regex patterns) → PII redaction → LLM summarization → deterministic 5-dimension QA scoring (weighted rubric). Evaluated 3 LLM providers (GPT-4o, Gemini 2.0 Flash, Groq Llama 3.3 70B) for quality/cost/latency trade-offs
• What shipped: End-to-end call QA automation with compliance-grade audit logging, PDF/JSON exports, and Gradio dashboard (analyze, history, observability tabs)
• Outcome: Enables 100% call coverage vs. <5% manual. Security-first: PII masking, injection scanning, append-only audit trails. 22 unit + 2 integration + 2 security tests

Repo →

• Problem: Content creators juggle 4-5 separate AI tools (text, image, editing, resume) with no unified workspace
• AI approach: Evaluated and integrated 3 AI/media APIs: Google Gemini 2.0 Flash (text generation, resume review), Clipdrop (text-to-image), Cloudinary AI (background/object removal), weighing output quality, latency, and per-call cost. Chose Gemini via OpenAI-compatible SDK for flexibility
• What shipped: Full-stack SaaS (React + Express + NeonDB) with 6 AI-powered tools, Clerk auth with Free/Premium subscription tiers, community publishing hub, and Vercel deployment
• Outcome: 9 API endpoints, 10 feature pages. End-to-end product ownership: auth, monetization (plan gating), AI integration, and deployment

Repo →  ·  Live →

• Problem: Non-technical teams can't query complex educational session data without SQL knowledge or analyst support
• AI approach: Designed a 3-stage LLM pipeline using Groq Llama 3.3 70B: (1) entity extraction with fuzzy matching (Fuse.js), (2) context-aware SQL generation against a star schema, (3) executive summary. All stages at temperature 0 for deterministic outputs
• What shipped: Natural-language query interface with auto-visualization (Chart.js), SQL transparency, and intelligent chart-type selection, deployed on Vercel
• Outcome: Democratized data access for non-technical teams. Star schema design (1 fact + 4 dimension tables) with 9 architecture diagrams documenting the system

Repo →

• Problem: Business teams need ad-hoc answers from scattered documents (reports, catalogs, spreadsheets) but can't write queries, and generic chatbots hallucinate
• AI approach: Built a 24-node n8n workflow with a full RAG pipeline: document ingestion (type-aware chunking) → OpenAI text-embedding-3-large embeddings → Pinecone vector storage with namespace isolation → Cohere Rerank for precision → grounded answer generation with explicit refusal when the answer isn't in the source
• What shipped: Production-ready low-code RAG agent with hallucination prevention, streaming responses, and conversation memory, validated in days, not sprints
• Outcome: Demonstrated rapid AI prototyping velocity. From idea to working RAG agent without writing backend code. 4 external API integrations orchestrated visually

Also built with n8n + Zapier: Multi-Agent Healthcare Assistant (GPT-4o coordinator + 3 sub-agents, 26 nodes) · Automated Client Inquiry Responder (GPT-4o-mini, <2 min response SLA, ~$0.001/email)

Repo →

• Problem: PMs, engineers, and developers curious about Agentic AI and RAG lack a structured, hands-on starting point that doesn't require paid courses or deep ML background
• What I built: A browser-based interactive learning platform with 22 modules across 6 progressive phases (LLM fundamentals → tokens & context → prompting → RAG → agentic patterns → production & observability). Includes quizzes with deterministic shuffling, XP/streak gamification, per-module notes with markdown export, and a RAG evaluation playground
• Why it matters: Designed as a zero-barrier entry point for anyone getting started with AI. No signup, no payment, fully client-side (React + Vite, 96KB gzipped). Built to help the community learn the same concepts I apply in my AI product work
• Outcome: 22-module curriculum shipping as a single SPA. CI/CD, Docker support, and deployment configs for Vercel/Netlify/Cloudflare out of the box

I keep the full engineering stack sharp so I can prototype fast, evaluate technical trade-offs with my engineering team, and never be blocked on "is this feasible?"

Algorithmic thinking and consistent execution. A credibility signal for technical AI PM conversations.

B.Tech, Computer Science · Graphic Era University, Dehradun · GPA: 8.99 / 10

Let's connect. I'm always up for conversations about AI product building, multi-agent systems, and shipping AI that actually works.

LinkedIn · X / Twitter · LeetCode