Kaushalendra
Singh
AI Engineer — Retrieval, Agents & LLM Infra
I build the backend and retrieval infrastructure behind LLM products — hybrid vector/graph retrieval, multi-agent orchestration with LangGraph, and the MCP layer that connects it to real tools. Currently building EpicWise's core context engine; also behind Evident, an open-source extraction and trust layer, and Buildify Labs, an agentic analytics platform.
Experience
EpicWise
May 2026 - Present
Canada startup
- Working directly with the founding team on EpicWise's core context engine - designing the multi-step retrieval and reasoning pipelines (LangGraph) that decide what context the system surfaces and why.
- Built hybrid retrieval combining vector search (Pinecone) and graph-based retrieval (Neo4j), so the engine surfaces relationships pure similarity search misses - not just similar text, but connected context.
- Authoring system prompts, execution routers, and tool-calling schemas; building the MCP integrations that connect EpicWise to a developer's existing tools instead of asking them to switch context.
- Built the internal tracing and observability layer (Langfuse + structured logging) the team uses to debug retrieval quality - the infrastructure that's invisible until it's missing, and why architecture decisions here aren't guesses.
Feb 2026 - Mar 2026
- Built the UI for an AI course platform using Next.js, TypeScript, and Tailwind CSS - improving the learning experience for 50K+ active learners.
- Replaced hardcoded data with Graphy's REST APIs and implemented SSR + code splitting, cutting page load time by 30% and bundle size by ~22%.
- Delivered course catalog, auth flows, and learner dashboards end-to-end within 6 weeks, shipping all features on time without QA rollbacks.
GRS Worker
Sep 2025 - Nov 2025
- Architected a full-stack worker management app with Next.js 15, Prisma ORM, and NextAuth - handling complex relational DB schemas and secure session management.
- Integrated Zod validation across all API routes, reducing form submission errors by 40% in the first month post-launch.
- Added Twilio SMS, QR-code check-in flows, and Framer Motion animations; cut average API response time by ~35% using Turbopack and Server Actions.
Open Source Contributor · Joplin
Mar 2026 - Apr 2026
- Joplin is a privacy-focused, open-source note-taking app with 45,000+ GitHub stars - opened 7 pull requests against its Desktop and mobile apps, with 2 merged into a mature, actively-maintained codebase under real reviewer scrutiny.
- Fixed a Desktop renderer crash triggered when closing a secondary window (#14849) — merged into Joplin's production codebase.
- Fixed OneNote .zip import handling when .one files sit at the root level of the archive (#14605) — merged.
- Opened fixes for Markdown↔Rich Text Editor switching (preventing HTML blocks from being escaped), duplicate tag syncing across devices, and truncated or corrupted .one file imports during OneNote migration.
- Contributed a mobile fix for title field border thickness on Android and investigated OAuth support for Joplin Server on Desktop.
Projects
Skills
Languages
AI / LLM Engineering
LangGraphBackend & Databases
Frontend
Contact
Ways to Connect
Open to backend/AI engineering roles, internships, and freelance work — especially anything involving retrieval, agents, or LLM infrastructure.
Usually replies within a day
Available for remote, freelance, or full-time
Quick Message
Or type it out here — goes straight to my inbox.
Writing
All articles
I Built an AI-Powered VS Code Extension That Understands Your Entire Codebase
How I went from find + mental models to a visual dependency graph, static flow analysis, and RAG-powered Q&A — inside VS Code.
Building Your First MCP Server
There is a specific kind of frustration that comes from tracking expenses through an app — you open it, tap through three screens. Here's how a simple MCP server fixes it.
Understanding MCP Architecture from First Principles
A practical breakdown of how AI systems talk to external tools like GitHub, Slack, and databases - and why MCP is the protocol that makes it clean.
···
Counting…
The best portfolio is running code.
Open to work — if my profile fits a role you're hiring for, email or ping me. I reply within a day.