// AI Developer · Full-Stack Developer

Ashutosh Kumar Rao

Final-year IT undergraduate building agentic AI systems that don't fall apart outside a demo — LangGraph orchestration, hybrid RAG pipelines, and full-stack products shipped end to end. Currently contributing to Hive by Aden, a YC-backed AI agent framework — debugging real production failures across model providers and execution pipelines, not just fine-tuning on Colab notebooks.

YouTube Channel

1000+

GitHub Contributions

59+

Repositories

SWE / AI

Open To

[ ~ ] Three ways I help teams build and ship

What I do

[ * ] How the path has unfolded — roles, teams, and outcomes

My career & experience

Open Source Contributor

Hive by Aden

NOW

Mar 2026 — Present · Remote · YC-backed AI Agent Framework

Contributing to Hive's core agent framework alongside the core engineering team on a large-scale production codebase. Diagnosed and patched a critical 429 rate-limit silent failure spanning the backend-to-frontend pipeline — event loop, SSE transport, and React workspace components — with the fix currently under maintainer review. Also debugged unsupported model configs in the Cerebras provider and proposed improvements to model auto-detection for contributor API keys, reducing onboarding friction.

Open Source Contributor

GeminiCLI by Google

2026

Jan 2026 — Present · Remote · Tooling & Infra

Identified and raised multiple log-based issues that improved debugging clarity, error visibility, and developer experience across the CLI toolchain. Surfaced silent failures and inconsistent log outputs that were affecting contributor productivity and local development workflows.

[ / ] Highlighted builds worth the pixels

My Work

01

Forge

AI AgentDeveloper ToolCLI

Forge is a terminal-native AI coding agent built for developers who live in the command line, not a chat window. Describe what you want built, and Forge plans the work, writes the code, runs it, and ships it — all without leaving your terminal. Under the hood, a LangGraph-orchestrated Plan → Execute → Observe → Retry loop keeps the agent honest, self-correcting through a custom compiler-validation layer whenever a generated script breaks. Every prompt streams over a bidirectional WebSocket channel into a workspace executor that maps human intent onto real file-system operations across deep project directories. An auto-indexing ingestion layer chunks the entire codebase into dense embeddings, stored in Supabase pgvector, so every generated change is grounded in the actual structure of your repo instead of guessing at it.

TOOLS AND FEATURES

Python, LangGraph, LangChain, WebSockets, Supabase pgvector, Claude API

View on GitHub

02

CodeSense AI

AIRAGCode Review

CodeSense AI rethinks automated code review by grounding every suggestion in evidence instead of vibes. It's a hybrid RAG, multi-agent platform that pulls grounding context — internal style rules, PEP conventions, and project-specific patterns — directly from a retrieval layer before a single line of feedback is generated, so every review comes with a citation, not just an opinion. A language-evaluation router runs Tree-sitter AST parsing to turn raw file trees into structured code maps, which are then routed through Anthropic Claude and Groq models for deep contextual analysis. LangSmith tracing threads through the entire pipeline for full transparency into prompt performance, which helped push hallucination rates down to a negligible margin. The result is a code-review platform that reviewers can actually trust, not just skim past.

TOOLS AND FEATURES

FastAPI, LangChain, LangGraph, Tree-sitter, Sentence Transformers, Claude / Groq

View on GitHub

03

Chat-Cue

B2B SaaSMulti-TenantSupport Platform

Chat-Cue is a production-grade, multi-tenant B2B customer support platform built from scratch inside a Turborepo monorepo — the kind of architecture that has to hold up when multiple companies are relying on the same instance at once. Full data isolation and role-based access control keep every tenant's workspace strictly separate, while an optimized database schema eliminates the query bottlenecks and transaction conflicts that usually surface under real traffic. A multi-tier Stripe billing engine handles subscriptions end to end, AWS S3 manages file storage, and Sentry keeps a constant eye on errors across the stack for full observability. Real-time messaging runs over WebSockets with an SSE fallback, so conversations stay live and reliable even when the primary transport drops — because a support platform that lags is worse than no support platform at all.

TOOLS AND FEATURES

Next.js 15, Convex, Stripe, AWS S3, Sentry, Turborepo

View on GitHub

04

Scratch

AutomationVisual BuilderWorkflows

Scratch is a visual automation platform for people who want the power of code without writing much of it. Drag connected nodes onto a canvas — triggers, conditions, actions — and wire them together to build workflows that actually run: emails that fire on schedule, data that syncs across apps, tasks that trigger themselves. It's built around the idea that automation shouldn't require a developer on standby every time a business process changes. The visual builder makes the logic legible at a glance, so non-technical teammates can follow — and eventually build — the same pipelines an engineer would otherwise hand-wire. It's an early but genuinely useful step toward making internal tooling something teams can shape themselves.

TOOLS AND FEATURES

TypeScript, React, Visual Workflow Engine

View on GitHub

05

Ecommerce-2

E-CommerceFull-StackPayments

Ecommerce-2 is a full-stack ecommerce build made to survive contact with real customers, not just a portfolio demo. It's built on Next.js with secure Stripe payments wired in end to end — from cart to checkout to confirmation — alongside proper user authentication and session handling so accounts and orders stay exactly where they belong. Cart management is built to feel instant: quantities update, totals recalculate, and stock stays in sync without the janky reloads that plague a lot of hobby ecommerce projects. The whole thing was built with production deployment as the actual goal, not an afterthought — meaning the checkout flow, error states, and edge cases got as much attention as the product grid itself. It's ecommerce done the way it needs to hold up, not just the way it needs to look.

TOOLS AND FEATURES

Next.js, Stripe, Cart & Auth System

View on GitHub

[ : ] Stack, education, and languages

AGENTIC ARCHITECTURE

LangChain OrchestrationLangGraph Stateful WorkflowsMulti-Agent Communication Loops

GENERATIVE AI & RAG

Hybrid Retrieval (BM25 + Dense Search)Cross-Encoder RerankingCitation EnforcementOllama

ML / DL MODELS

TransformersBERTGPT SystemsLocal SLM Benchmarking

VECTOR STORES & DATABASES

Supabase pgvectorChromaDBFAISS IndexesPostgreSQLDrizzle ORM

CORE LANGUAGES & TECH

PythonTypeScriptJavaScriptC++HTML5CSS3FastAPINode.jsWebSocketsLangSmith Tracing

EDUCATION

B.Tech, Information Technology

Shri Ramswaroop Memorial Engineering College and Management, Lucknow · July 2023 — May 2027

LANGUAGES

English / ProfessionalHindi / Native

[ + ] certifications

Certifications

[ @ ] contact

Let's talk

Open to opportunities & collaboration

Lucknow, Uttar Pradesh, India

+91 90268 68343

aashutoshrao68@gmail.com

LinkedIn GitHub X

Ashutosh Kumar Rao

What I do

AI Development

Full-Stack Development

Open Source Contribution

My career & experience

Open Source Contributor

Open Source Contributor

My Work

Forge

CodeSense AI

Chat-Cue

Scratch

Ecommerce-2

Certifications

Generative AI with LangChain

CS50's Introduction to Artificial Intelligence with Python

Artificial Intelligence

Let's talk