Skip to content
View shivamtiwari3's full-sized avatar

Block or report shivamtiwari3

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
shivamtiwari3/README.md

Hi, I'm Shivam πŸ‘‹

πŸ“ Abu Dhabi, UAE | 🧠 AI Product Lead | πŸ› οΈ Builder-PM | πŸš€ Founder, Apli.ai (acquired)

I take AI products from demo to production and own what happens after: model choice, evals, cost per resolution, and the line where the agent stops and a human takes over.

Claude Code TypeScript Python Next.js MCP Voice AI Evals RAG

Start Here

  • πŸ”‘ keyway - Use any model in Claude Desktop. Bring your own key.
  • πŸŽ™οΈ stopbias - Does voice-agent stop latency survive the phone network? A study with known ground truth.
  • ☁️ HobbyCloud - A cloud for small software. The full PRD, written in the open.
  • πŸ—ΊοΈ Xpand - See what every area in the UAE is missing, and where to build next.
  • πŸ₯ Sanad - Dubai clinical capacity gaps, district by district. Live

Shipped in Production

Product What it is Result
🦠 Bugzy WhatsApp AI gut-health coach. Agentic RAG, multi-turn memory, hallucination guardrails 200K+ users in 6 months, 1.5x retention
🎧 Goodsy Agentic CX that processes refunds, order changes and subscriptions, with HITL escalation 65% of 24K monthly tickets resolved without a human, β‚Ή3Cr+/yr saved
βš–οΈ Model migration GPT-4o β†’ Llama 3.2 11B, gated on RAGAS faithfulness and relevancy evals 30% lower inference cost at 85%+ quality parity
πŸ‘“ Lenskart / John Jacobs GPT-4 face-shape eyewear recommendations, 500+ case eval suite before launch Conversion 3% β†’ 7.5%
πŸŽ“ Apli.ai NLP job matching as a ranking problem, 1,200+ institutions, 6.5L+ candidates Acquired by Edulab

Current Projects

Agents & AI Tooling

  • πŸ”‘ keyway - Tiny local BYOK gateway for OpenAI- or Anthropic-compatible providers. Website
  • 🧰 claude-skills - Claude Code skills for daily work: cold email, LinkedIn drafts, sprint summaries
  • πŸ“£ webengage-mcp - Track users, log events and trigger WebEngage campaigns in plain English from any MCP client

Voice AI

  • πŸŽ™οΈ stopbias - Offline full-duplex benchmarks vs real phone lines: G.711, PSTN passband, Opus, packet loss, line noise. 144 tests, preregistered Phase 2

Product Specs, in the Open

  • ☁️ HobbyCloud - Private internal apps with brokered credentials and spend ceilings. Problem, competition, 5-stage roadmap, pricing, validation plan

UAE Civic & PropTech Prototypes

  • πŸ—ΊοΈ Xpand - Service-gap intelligence for UAE real estate. Abu Dhabi AI PropTech Challenge
  • πŸ“‘ community-pulse - The first Xpand prototype: real OSM amenity gaps across 33 districts. Live
  • πŸ₯ Sanad - Bed and specialist gaps, 2026–2040 demand forecasts and planning briefs for Dubai. Live

Open Source Contributions

Merged upstream:

  • πŸ¦™ Ollama (182kβ˜…) - Close the thinking block before tool_use in the Anthropic API
  • πŸ¦• Deno (108kβ˜…) - Watched paths for deno test --watch, plus reported test locations
  • ⚑ Ruff (50kβ˜…) - Avoid strict TC001-003 behaviour when future annotations are on
  • 🎭 Backstage (34kβ˜…) - Replace removed tokens in the Table component
  • β–² Vercel AI SDK (27kβ˜…) - PDF and other non-image files reach the OpenAI Responses API as input_file. Co-authored with @aayush-kapoor, from my #14140
  • 🌳 Biome (26kβ˜…) - Skip unsupported <script> types in HTML
  • πŸ” Jaeger (23kβ˜…) - Guard addJitter against zero durations
  • πŸ€– SWE-agent (20kβ˜…) - Thread-safe patch hook for multi-worker batch runs
  • 🦜 LangChain.js (18kβ˜…) - PPTXLoader on the officeparser v6 API
  • πŸ”Ί Delta Lake (9kβ˜…) - Fix DeletionVectorDescriptor unique IDs
  • 🧠 exo - Allowlist instead of blocklist for the Bash PreToolUse hook

In review:

  • πŸ”₯ Python Fire (Google, 28kβ˜…) - 4 PRs: short flags swallowed by **kwargs, completion for class-grouped commands, logging for diagnostics, custom __repr__
  • πŸ”Ž Magika (Google, 19kβ˜…) - 3 PRs: a public get_content_type_info() API, CLI colours for unmapped groups, typos CI
  • πŸ•ΈοΈ graphify (124kβ˜…) - 8 PRs across extraction, markdown and language support

How I Work

  • Model choice is a P&L decision. Run evals before every swap, or you're guessing.
  • Design the escalation path before launch, not after the first angry customer.
  • Prototype in days with agents, then decide if it deserves a roadmap.
  • Write the spec in public when the thinking is the product.

Teaching

  • πŸŽ“ Lead Instructor, Google for Startups - AI product and prototyping for early-stage founders across India
  • πŸ›οΈ Visiting Faculty / Speaker - IIT Bombay, MESA School of Business, Masters' Union and 20+ events on GenAI and product

Connect

🌐 shivamtiwari.si Β· πŸ’Ό LinkedIn Β· 24K+ founders and PMs follow along

Pinned Loading

  1. webengage-mcp webengage-mcp Public

    MCP to use webengage with natural language within Claude desktop / Cursor / Antigravity

    TypeScript

  2. EMRMVP EMRMVP Public

    It's a emr management platform

    PHP

  3. Chiefofstaff Chiefofstaff Public

    Your EA to manage personal tasks and add more and more context

    HTML

  4. ai ai Public

    Forked from vercel/ai

    The AI Toolkit for TypeScript. From the creators of Next.js, the AI SDK is a free open-source library for building AI-powered applications and agents

    TypeScript

  5. langchain langchain Public

    Forked from langchain-ai/langchain

    The agent engineering platform

    Python

  6. ollama ollama Public

    Forked from ollama/ollama

    Get up and running with Kimi-K2.5, GLM-5, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

    Go