Skip to content
View roshankumar1204's full-sized avatar
🎯
Focusing
🎯
Focusing

Highlights

  • Pro

Block or report roshankumar1204

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
roshankumar1204/README.md

Hi, I'm Roshan Kumar

Full-stack engineer focused on building production-ready systems — workflow automation, AI-powered applications, and data pipelines that handle real-world messiness reliably.

Currently based in Bengaluru. Open to roles at the intersection of AI and systems engineering.


Stack

Languages  ·  Java · Python · JavaScript
Frontend  ·  React · TypeScript · Tailwind CSS · ReactFlow · TanStack Query · Recharts
Backend  ·  Node.js · FastAPI · Flask
Databases  ·  PostgreSQL · MongoDB · Redis
AI / ML  ·  LangChain · LLMs · scikit-learn · MediaPipe · BiLSTM
Infra  ·  Docker · AWS · Vercel · Render · Neon


Projects

AI Spend Observability System  ·  Live

Multi-vendor AI usage tracking — one trustworthy number for cost, usage, and identity across tools

The hard part of this system isn't the dashboard — it's making messy, inconsistent, late-arriving data from multiple vendors converge into something you can act on without silent failures.

  • Canonical event schema — every vendor (ChatGPT, Copilot, Cursor) maps into one AIUsageEvent shape; adding a new vendor is one new parser file, nothing else changes
  • Deterministic deduplication — hashed dedup_key per event; same event delivered twice is recognized and skipped, never double-counted
  • Late-cost backfill — billing lag treated as a patch on the existing row, not a new duplicate event
  • Drift detection — baseline key-paths recorded per vendor on first ingest; any silent field rename or removal raises DRIFT_DETECTED instead of quietly writing zeros
  • Proof-based identity resolution — the same human appears under different IDs across tools; identities are only linked when there's an explicit recorded proof type, never a fuzzy guess
  • Snapshot regression harness — pytest snapshots ensure fixing one vendor's parser can't silently break another's output

FastAPI PostgreSQL React TypeScript TanStack Query Recharts Neon Render Vercel


Workflow Automation Platform  ·  Live

Visual AI pipeline editor with drag-and-drop node canvas and real-time DAG validation

  • Node-based execution pipeline built with React Flow — dynamic, composable, stateful
  • FastAPI backend with DFS cycle detection (3-state node tracking) to validate directed acyclic graphs
  • Returns node count, edge count, and DAG status on every save
  • Workflow persistence, import/export, and execution metadata via REST APIs

React ReactFlow FastAPI Tailwind CSS


Soni  ·  sonii.app

Voice-activated AI assistant with context-aware responses and personalized auth

  • LangChain-powered NLP backend handling weather, Wikipedia, calendar, and web automation via LLM tool calls
  • Passcode-based authentication with PostgreSQL session management — scoped, persistent, secure
  • Google Cloud Text-to-Speech integration for natural audio output

Flask LangChain PostgreSQL Google Cloud TTS


Cloner  ·  cloner.live

End-to-end pipeline that converts any live website into clean HTML or React + Tailwind

  • HTML and CSS semantic analysis preserves layout, content, and design while stripping hydration scripts and trackers
  • Outputs a complete, ready-to-run project with merged components and minimal hallucination
  • Handles both static HTML export and full React component output

ISL Translator — Indian Sign Language

Multimodal AI pipeline for real-time gesture-to-Hindi translation

  • BiLSTM + MediaPipe for dynamic gesture recognition from video input
  • LLM-based language generation converts gesture predictions into natural Hindi sentences
  • Modular architecture with clean separation between vision, prediction, and generation — designed for API deployment

Python BiLSTM MediaPipe LLMs


Experience

Associate Software Engineer  ·  Mphasis  ·  Mar 2026 – May 2026
Built Python data pipelines for enterprise analytics — ingestion, cleaning, feature engineering, and a scikit-learn predictive model on top. Automated the full end-to-end flow, cutting analysis time from hours to minutes.

Software Developer  ·  Shubham Engineering  ·  May 2025 – Aug 2025
Built and deployed the company's production website. Managed digital compliance workflows for ESIC and PF documentation.


Education

MCA  ·  Presidency University, Bengaluru  ·  2024 – 2026  ·  CGPA 8.9
BCA  ·  Amity University, Ranchi  ·  2021 – 2024  ·  CGPA 7.65


Certifications

AWS SimuLearn: Cloud Practitioner  ·  AWS Training & Certification
Java Certified  ·  HackerRank


Contact

roshanhere03@gmail.com  ·  LinkedIn  ·  GitHub  ·  +91 73688 27906

Pinned Loading

  1. ai-usage-ledger ai-usage-ledger Public

    TypeScript 1

  2. rihzu-token rihzu-token Public

    This project is a crypto token built on the Internet Computer (ICP) blockchain using Dfinity’s SDK (Motoko) and Node.js. The goal is to explore decentralized token standards, smart contract develop…

    JavaScript

  3. sol_vault sol_vault Public

    TypeScript

  4. vectorshift vectorshift Public

    JavaScript