Hi, I'm Karan
kwakhare5·typescript/python·ai systems & full-stack
i'm karan, an ai & full-stack engineer based in pune.
autonomous ai systems, mcp servers, and low-latency tools are what i specialize in. builder of grocer (instamart mcp) and outpost.
high-agency developer workflows and agentic tooling are what i'm most into right now.
currently
- building grocer with swiggy builders club
- building indie dev tools outpost & git for prompts
- studying computer engineering @ adypu
previously
- full-stack developer intern @ ateion
- high school education @ wisdom high
#builds(4)
WhatsApp grocery assistant that verifies your order against live Swiggy Instamart inventory and auto-fixes out-of-stock items.
Quick-commerce carts constantly break when items go out of stock or pack sizes change, forcing you to start over.
A FastAPI backend turns WhatsApp messages into a strict order contract, checks live Swiggy Instamart inventory via MCP, and automatically suggests valid swaps before you confirm checkout.
- Direct Swiggy Instamart cart integration via Model Context Protocol (MCP)
- Automatic cart repair that replaces out-of-stock items without breaking dietary rules
- Explicit 1-tap WhatsApp checkout confirmation to prevent accidental purchases
Inventory replenishment decision engine for quick-commerce dark stores that forecasts stockouts and routes inter-store transfers with human-in-the-loop approval.
Quick-commerce dark stores lose margin to sudden stockouts during demand surges and perishable food expiry, while blind auto-reorders risk phantom stock or costly emergency courier runs.
A FastAPI and LangGraph 5-node state machine paired with Holt Linear demand forecasting that tracks physical FIFO inventory batches across 5 Mumbai hubs and proposes transfers or purchase orders behind a human approval gate.
- Holt Linear double exponential smoothing evaluated on a 3-day holdout using WAPE and MAE
- Physical FIFO batch ledger enforcing strict conservation of mass and realistic transit ETAs
- Level-2 autonomy gate requiring operator sign-off with stale-state detection and idempotency locks
Git-style version control for AI prompts with commit histories, token diffing, and multi-model evals.
Editing prompts in messy text files leads to broken outputs, lost versions, and unnoticed token cost spikes.
A Next.js and PostgreSQL platform that stores immutable prompt versions, renders word-by-word diffs in Monaco Editor, and runs automated tests across multiple AI models simultaneously.
- Monaco Editor diff viewer showing exact word and token changes between versions
- Automated test runner benchmarking accuracy, latency, and cost across models
- Native gfp CLI and webhooks for pulling versioned prompts straight into production code
Chrome extension that rewrites rough drafts into clear Slack, Gmail, and LinkedIn messages in under 200ms.
Drafting emails and workplace chat messages often sounds stiff, wordy, or blunt, causing unnecessary back-and-forth.
A Chrome extension using Shadow DOM for style isolation that streams text to Groq LLMs on Cloudflare Workers, replacing words directly in your text box without leaking API keys.
- Sub-200ms edge inference powered by Groq LPUs on Cloudflare Workers
- Shadow DOM encapsulation preventing style collisions in Gmail, Slack, and LinkedIn
- Offline tone engine with 30+ fallback rules that work even without an internet connection
#activity
github#stack
tooling#artifacts
gallery
// downtown, dubai
// gardens by the bay
// grand mosque, abu dhabi
// marina barrage park