Ayush Kumar

Ayush Kumar

Full-stack × ML · production systems, research, and the occasional drone


About

I like owning a problem from end to end: the model, the systems around it, and the interface someone eventually clicks on. At Cornerstone OnDemand, that has meant building CSOD Harness, an LLM engineering assistant that indexes our codebase, links it to the infrastructure beneath it, and coordinates several models to trace bugs across repositories, using 56% fewer tokens along the way. It has also meant shipping .NET APIs and React dashboards for an enterprise job scheduler, and a Rust-backed rendering pipeline that produces thousands of PDFs per batch across 22+ production regions. Outside work, I build full products on my own. Kima is an event-sourced operations platform, covering ledgers, payroll and approvals, that runs entirely on Cloudflare’s edge. MailForge is a Go delivery engine that sends 100 emails a second. I care just as much about how things look, and I’ve designed and built the web presence for VinHack and StudyHub. My research follows a single conviction: models should propose and deterministic logic should decide. That idea runs through a document-extraction system that records where every value came from, a pipeline that reconstructs 3D scenes from ordinary photos and video, and a browser extension that composes music for the page you’re reading without your browsing ever leaving your device. Before Cornerstone I worked at Deloitte, Samsung R&D and Titan Smart Labs. At VIT, I led engineering for a campus app used by 30,000 students, ran a national hackathon, and led the software team behind an autonomous drone that took first place at an international challenge. If you’re building agents, AI/ML systems, solid backends, or beautiful frontends, I’d love to hear about it.

Skills
Python
Go
Rust
TypeScript
JavaScript
C++
SQL
PyTorch
LLM Agents
LangGraph
LangChain
RAG
LoRA/PEFT
Computer Vision
React
Next.js
Node.js
FastAPI
.NET
PostgreSQL
Redis
Docker
AWS
Cloudflare Workers
Linux
Experience

Junior Engineer

Cornerstone OnDemand

Dec 2025 to Present
  • 1) Built CSOD Harness, an internal LLM engineering assistant with Roslyn code indexing, a code–infrastructure knowledge graph and multi-model agent orchestration, cutting token use by 56% in multi-repo debugging
  • 2) Built a mass PDF-generation pipeline (Python transformation engine feeding a Rust parallel renderer), deployed across 22+ production regions and generating thousands of PDFs per batch
  • 3) Shipped REST APIs, job lifecycle and retry logic, and React operational dashboards for the CSX Scheduler, a distributed job-orchestration platform on .NET and AWS EC2

SDE Intern, AI & Backend Engineering

Deloitte USI

Jun 2025 to Aug 2025
  • Built FastAPI microservices for an AI-enabled internal productivity tool, containerized with Docker at 90%+ PyTest coverage

Research Intern, ML & Generative AI

Samsung R&D Institute India

Sept 2024 to Jun 2025
  • Built a customer-segmentation pipeline with RFM modeling and CLV estimation (BG/NBD, Gamma-Gamma), plus a GenAI workflow that turns segments into executable SQL

Algorithm Intern, Smart Wearables

Titan Smart Labs

Jun 2024 to Aug 2024
  • Built preprocessing pipelines for multimodal smartwatch time series (heart rate and accelerometer) and improved sleep and stress classification, validated with leave-one-subject-out CV
Projects

Explainable Financial Advisory Engine

A deterministic quant scoring engine (fundamental, technical and risk metrics) combined with investor behavioral profiling and retrieval-grounded news sentiment. A local Llama 3.2 explains each recommendation but never touches the decision scores. Backtested at 18.5% annualized vs. 13.7% for the S&P 500, with 88.3% stability across 231 weight permutations. IEEE ICACKE 2026, Best Paper

Python
Llama 3.2
Sentence-BERT
yfinance

StudyHub

An exam-prep platform for VIT students, with past papers, notes, videos and a question bank. It runs on a Go backend with MongoDB vector search, Redis and R2. A 20-tool MCP server lets students study from the real question bank in Claude, ChatGPT or Cursor.

Go
Next.js
MongoDB
Redis
MCP

Schema-Aware Document Extraction

Runs OCR-free Donut and an OCR-grounded LLM in parallel and reconciles them field by field with a deterministic evidence policy that logs every value’s source. Reaches 0.980 F1 on multi-page contracts, beating a 7B VLM. Per-style LoRA adapters for LayoutLMv3 (4.5 MB each) improve layout macro-F1 by up to 0.19.

Donut
LayoutLMv3
LoRA/PEFT
LangGraph
Qwen2.5-VL

Z-Shift: 2D Media to 3D Assets

A modular four-stage pipeline that turns images, video, synced multi-camera captures and live feeds into 3D meshes and point clouds using MASt3R. Restoring capture order recovers 6.6× pairing locality, and reruns on the same GPU produce byte-identical output.

Python
PyTorch
MASt3R
FastAPI
FFmpeg

Web2Music

A Chrome extension that composes ambient music for the page you’re reading. An on-device NLI stage in the classifier cascade sends 38% fewer pages off-device at the same accuracy. Audio starts 36 ms after the mood commits, and measured-seam loop cutting drops audible seams from 74% to 12%.

Chrome MV3
ONNX Runtime Web
MusicGen
FastAPI
PostgreSQL

Kima: Internal Operations Platform

A solo-built platform for tasks, ledgers, attendance, expenses and payroll, running entirely on Cloudflare’s edge. Its event-sourced core folds balances from an append-only ledger, makes accruals idempotent, and routes every approval through one engine behind a unified inbox.

TypeScript
Next.js 15
Cloudflare Workers/D1/R2
Drizzle ORM
Auth.js

MailForge: Mass-Mailing Platform

A self-hosted email campaign platform with a Go API, a Redis-queued worker pool, JWT-scoped multi-tenancy, and rate-limited delivery at 100 sends a second with retry-backoff.

Go
Next.js
PostgreSQL
Redis
Docker
Education

B.Tech

Vellore Institute of Technology, Vellore · 9.35/10

Sept 2022 to Jun 2026