Software Engineer

Hello I'm Kunal Sharma

Software Developer & Android Enthusiast | INTP-A | Pragmatic, delivery-oriented | "If I gotta eat rocks, just tell me a measure"

2+

Years of
Experience

Top 0.1%

CodeChef &
GeeksforGeeks

2

Research
Publications

1,200+

Code
Commits

Services

What I can do for you

01

Web Development

Responsive front-ends in React and TypeScript, backed by Node.js and RESTful APIs. Clean, maintainable code and interfaces that stay fast.

02

Backend & Systems

Services in Rust (Axum, Diesel) and Python, with MySQL, MongoDB and Redis behind them. Containerised with Docker and deployed on GCP.

03

Desktop & ML Tooling

Native desktop apps with Tauri 2 and Rust, wrapping real ML workloads. PyTorch and NVIDIA NeMo pipelines that run entirely offline.

04

AI Benchmarks & Evals

Freelance work: cybersecurity challenges built from real memory-safety bugs, SWE-bench-style coding tasks, terminal-agent benchmarks and RLHF rubrics, each with a machine-verifiable test harness.

Why hire me?

A developer who ships

Three roles and two published papers so far.

Jun 2026 - Present

Software Engineer Current

AfterQuery · Remote

  • Author SWE-bench-style coding tasks from real bugs in private repositories, each with a machine-verifiable test harness and reference solution.
  • Turn real memory-safety bugs into cybersecurity evaluation challenges for frontier LLM agents.
May 2026 - Jun 2026

AI Systems Evaluation Intern

AirDawg Labs · Remote

  • Built benchmarks for Terminal-Bench 2.0, evaluating coding agents in real terminal environments.
Feb 2025 - Jun 2025

SDE Intern

Xelron AI · Bengaluru

  • Built internal tools and evaluation pipelines running coding tasks in isolated Docker containers with fixed resource limits.
  • Designed prompt-specific rubrics and pairwise model rankings feeding RLHF alignment workflows.

My Work

Projects

Desktop

Srtforge Studio

A fully offline, CUDA-accelerated subtitle pipeline shipped as a native Windows app. A Tauri 2 shell drives a Python ML worker over a custom IPC protocol.

RustTauri 2ReactPyTorchNVIDIA NeMo
Read the paper
Desktop

MPV Parakeet Transcriber

A Lua script for the MPV media player that generates subtitles for whatever is playing, using NVIDIA's Parakeet ASR model.

LuaMPVParakeet ASR
ML

Smart Grid Energy Prediction

Load forecasting under uncertainty using fuzzy logic instead of a rigid linear fit, with membership functions mapping temperature and time to predicted load.

PythonFuzzy LogicScikit-learn
ML

Bone Fracture Classification

Four CNN architectures compared on 1,287 X-rays across 10 fracture types. DenseNet121 came out ahead on AUROC. Published as a 2024 paper.

TensorFlowKerasDenseNet121
Read the paper
ML

Alzheimer Recognition

A neural-network classifier for Alzheimer recognition from medical imaging.

PythonNeural Network
Web

Web-based Expense Tracker

A web application for tracking and categorising expenses.

Web

Research

Published work

Two papers, with the results and the pipelines behind them.

Paper 2026 Sole author

CUDA-Accelerated Offline Subtitle Generation Pipeline Combining Neural Vocal Separation and ASR

An offline subtitle pipeline that isolates dialogue before transcribing it, so nothing is uploaded and the music never reaches the recogniser.

  • Separating vocals first cut word error rate from 6.57% to 5.26% across a 24-episode season, improving every episode.
  • Separation also cancels the accuracy penalty of int8 quantisation, making the cheapest precision setting the best local operating point.
  • Runs at RTF 0.16, about 6x faster than real time on a laptop GPU.
Read the paper Source
input.mkv ffmpeg melband-roformer parakeet-tdt faster-whisper geminioptional netflix timing rules .srt
Processing pipeline. Dataset-wide: 5.26% WER vs 6.57% for raw-audio Whisper.
Paper 2024 with Gaurav Kumar Singh

Utilization of Deep Convolutional Neural Networks for Bone Fracture Classification

A head-to-head comparison of CNN architectures for reading fractures off X-rays, rather than one model reported in isolation.

  • 1,287 X-rays across 10 fracture types, split 80/10/10 for training, validation and test.
  • A custom CNN benchmarked against ImageNet-pretrained DenseNet121, InceptionV3 and MobileNetV1.
  • DenseNet121 led on AUROC. Its dense skip connections extract more discriminative features.
Read the paper Source
1,287 X-rays · 10 classes 256×256 + augmentation four architectures compared sequential CNN DenseNet121 InceptionV3 MobileNetV1 AUROC comparison DenseNet121 · best AUROC
Study design. DenseNet121 led on AUROC across the 10 fracture classes.
Google ScholarFull publication list

Contact

Let's work together

Have a project in mind? Tell me about it and let's discuss how I can help.

Emailkunalsh926@gmail.com
Phone+91 6309132981
LocationIndia