Hitesh Goyal

> Applied AI Engineer | Speech Recognition · Vision · Language Models _

I've spent 3+ years building speech, vision, and language models, from research at Samsung R&D and NUS to production work at Tata Elxsi and Amplify Dental.

Open to freelance and full-time work

ABOUT

Signal over noise

Portrait of Hitesh Goyal, Applied AI Engineer

I start every machine-learning problem with the data: what it really represents, where it falls short, and what the problem needs. Then I build the simplest thing that could work, measure it on the metric that matters, and change course when the numbers say so.

This approach has worked across speech recognition, computer vision, biotech, and multimodal retrieval, using everything from classical methods and deep learning to generative models. The tools change with the problem. What matters is knowing which to use, how to use it, and when to try something else.

I carry that thinking through to execution, from data preparation and experimentation to evaluation, deployment, and usable interfaces. The goal is a reliable system built around real constraints like latency, compute, data quality, and measurable performance.

Areas I work in

Tools I use

Based in Bangalore, India. IST

PROJECTS

Selected projects

All projects

AudioBench

I built AudioBench for Amplify Dental to choose the right speech-to-text model. It tests models on noisy, accented, and distant-microphone audio and keeps the hand-checked transcripts and test setups, so the team never redoes that work.

Ask Krishna

Ask Krishna has two modes. Reflect lets you hold a question in mind and open the Bhagavad Gita to up to five random passages for guidance. Read Gita covers all 18 chapters.

Aftersight

I built Aftersight because my own photo library was disorganised and hard to search. It uses language models to organise images and find them by meaning, visible text, or metadata—not just filenames.

Pixtyle

I built Pixtyle to turn photos and videos into wallpaper-like or cartoon-like images using colour clustering, Fourier filtering, and edge detection.

Flodoro

I built Flodoro for people who want a Pomodoro timer, tasks, session history, and local music on one page without an account or application server.

OS Toolkit

I built OS Toolkit to analyse disk usage, copy and compare folders, and reuse common file operations from Python or the command line. I benchmarked its copying against copytree and robocopy.

LazyResume

I built LazyResume for people who already have a resume PDF and want to turn it into a structured profile, correct it, and regenerate a consistently formatted resume.

Fuzzy Hybrid Model for Stock Trading

In this stock-trading research project, I reduced the predictor's inputs from 96 values to 26, then combined a deep-learning price predictor with fuzzy buy, sell, or hold rules.

Detecting AI Generated Images

I built this PyTorch workflow for Kaggle's Detect AI vs. Human-Generated Images competition. It records each configuration, metric, plot, and checkpoint so I can compare experiments.