Shreshtth Kumar Agarwaal

AI research engineer · Amberg, Germany

Engineer and master’s student.

I am in the master’s in artificial intelligence at OTH Amberg-Weiden. The work I care about is the overlap of software engineering, AI systems, and applied research.

I studied computer science and engineering at BML Munjal University. The first internships were software and data: application workflows at MetLife, then ETL and dashboards at KPMG India.

At GlobalLogic I built GenAI into a real product, AuditAI, on React and FastAPI. At Sftwtrs.ai I worked as an AI Research Engineer on call automation that has to survive a production phone line.

Work

Selected work.

01 · GlobalLogic · internship
AuditAI
An audit platform that drafts the assessment, then waits for a person before the report ships.
- React
- FastAPI
- spaCy
- Presidio
Audit work repeats the same assessment, scoring, and write-up, and two reviewers can score the same answer differently.
I built the MVP. A domain holds its own question bank. An LLM scores a response, suggests a maturity level, writes a summary, and asks a follow-up when the answer is thin. A reviewer still signs off.
One configurable domain model, so IT, compliance, and finance audits do not each need a new app.
Names and other personal data are stripped with spaCy and Microsoft Presidio before the model sees the text.
The report stays a draft until someone reviews it.
02 · Sftwtrs.ai · client work
AI Call Automation
A production voice system for customer calls, built for a client, not as a side project.
- LiveKit
- SIP
- FastAPI
- PostgreSQL
- Supabase
- Docker
The client needed phone calls handled in real time, on the telephony they already use, without a long pause between turns.
I worked on the call path: LiveKit for the media session, SIP and Tata SmartFlo for the dialer, and a FastAPI service that keeps the conversation in PostgreSQL through Supabase. Docker carries the production setup.
Speech stays on a streaming path, with barge-in, so the caller can interrupt.
Session state lives in the database, not in the process, so a call can be recovered.
03 · University policy search
EduQuery
Students ask a question and get an answer taken from the college’s own documents.
- Flask
- React
- Pinecone
- Hugging Face
- AWS
Policy documents are long, and a model that answers from memory will invent a rule the college never wrote.
Documents are split, embedded with Hugging Face, and stored in Pinecone. Flask retrieves the closest passages and puts them in the prompt. React is the chat, with Appwrite for sign-in. The app was deployed on AWS.
The model only sees retrieved passages, so an institution-specific answer has a source.
Sign-in sits in front of the documents.
Documents /
Chunks /
Embeddings /
Pinecone /
Flask /
Answer

Path

Work experience.

June 2025 – January 2026
AI Research Engineer
Sftwtrs.ai
- AI call automation for client work, with LiveKit, SIP, and Tata SmartFlo.
- Backend services and the database behind the calls: PostgreSQL, Supabase, and Docker.
- Production infrastructure, technical notes, and day-to-day work with the engineering team.
January 2025 – May 2025
GenAI and Full Stack Development Intern
GlobalLogic
- Built AuditAI in React and FastAPI, with a workflow that can be reconfigured per audit domain.
- Anonymization with spaCy and Microsoft Presidio, and a human review step before a report is final.
June 2024 – August 2024
Data Analysis and Visualization Intern
KPMG India
- ETL pipelines that pulled, cleaned, and joined data from more than one source.
- Power BI dashboards and DAX over customer and sales data.
May 2023 – July 2023
Full Stack Development Intern
MetLife Global Shared Services
- Angular, .NET MVC, and Microsoft SQL Server inside an existing ERP.
- Automated team budget allocation, monthly accrual, and the report people actually opened.

Education

The degrees.

March 2026 – March 2028
M.Sc. Artificial Intelligence for Industrial Applications
OTH Amberg-Weiden · Amberg, Germany
August 2021 – July 2025
B.Tech. Computer Science and Engineering
BML Munjal University · Gurugram, India

Earlier

Other projects.

Sound Slice
A Flutter app that separates a song into stems. Audio goes up, Demucs runs through an inference API, and the stems come back.
- Flutter
- Firebase
- Demucs
Neural Image and Video Style Transfer
A deep learning project that transfers a visual style onto images and video.
- Neural style transfer
Explainable AI for COVID Detection
A CNN on chest X-rays, with Grad-CAM so you can see where the model was looking. It is a study on the COVID-19 Radiography Database, not a diagnostic tool.
- Python
- Grad-CAM
- 20,000+ images

Toolkit

Technical toolkit.

Contact

Say hello.

Selected work.

01 · GlobalLogic · internship

AuditAI

An audit platform that drafts the assessment, then waits for a person before the report ships.

React
FastAPI
spaCy
Presidio

Audit work repeats the same assessment, scoring, and write-up, and two reviewers can score the same answer differently.

I built the MVP. A domain holds its own question bank. An LLM scores a response, suggests a maturity level, writes a summary, and asks a follow-up when the answer is thin. A reviewer still signs off.

One configurable domain model, so IT, compliance, and finance audits do not each need a new app.
Names and other personal data are stripped with spaCy and Microsoft Presidio before the model sees the text.
The report stays a draft until someone reviews it.

02 · Sftwtrs.ai · client work

AI Call Automation

A production voice system for customer calls, built for a client, not as a side project.

LiveKit
SIP
FastAPI
PostgreSQL
Supabase
Docker

The client needed phone calls handled in real time, on the telephony they already use, without a long pause between turns.

I worked on the call path: LiveKit for the media session, SIP and Tata SmartFlo for the dialer, and a FastAPI service that keeps the conversation in PostgreSQL through Supabase. Docker carries the production setup.

Speech stays on a streaming path, with barge-in, so the caller can interrupt.
Session state lives in the database, not in the process, so a call can be recovered.

03 · University policy search

EduQuery

Students ask a question and get an answer taken from the college’s own documents.

Flask
React
Pinecone
Hugging Face
AWS

Policy documents are long, and a model that answers from memory will invent a rule the college never wrote.

Documents are split, embedded with Hugging Face, and stored in Pinecone. Flask retrieves the closest passages and puts them in the prompt. React is the chat, with Appwrite for sign-in. The app was deployed on AWS.

The model only sees retrieved passages, so an institution-specific answer has a source.
Sign-in sits in front of the documents.

Documents /
Chunks /
Embeddings /
Pinecone /
Flask /
Answer

Work experience.

June 2025 – January 2026

AI Research Engineer

Sftwtrs.ai

AI call automation for client work, with LiveKit, SIP, and Tata SmartFlo.
Backend services and the database behind the calls: PostgreSQL, Supabase, and Docker.
Production infrastructure, technical notes, and day-to-day work with the engineering team.

January 2025 – May 2025

GenAI and Full Stack Development Intern

GlobalLogic

Built AuditAI in React and FastAPI, with a workflow that can be reconfigured per audit domain.
Anonymization with spaCy and Microsoft Presidio, and a human review step before a report is final.

June 2024 – August 2024

Data Analysis and Visualization Intern

KPMG India

ETL pipelines that pulled, cleaned, and joined data from more than one source.
Power BI dashboards and DAX over customer and sales data.

May 2023 – July 2023

Full Stack Development Intern

MetLife Global Shared Services

Angular, .NET MVC, and Microsoft SQL Server inside an existing ERP.
Automated team budget allocation, monthly accrual, and the report people actually opened.

Other projects.

Sound Slice

A Flutter app that separates a song into stems. Audio goes up, Demucs runs through an inference API, and the stems come back.

Flutter
Firebase
Demucs

Neural Image and Video Style Transfer

A deep learning project that transfers a visual style onto images and video.

Neural style transfer

Explainable AI for COVID Detection

A CNN on chest X-rays, with Grad-CAM so you can see where the model was looking. It is a study on the COVID-19 Radiography Database, not a diagnostic tool.

Python
Grad-CAM
20,000+ images

Technical toolkit.

ShreshtthKumarAgarwaal

Engineer and master’s student.

Selected work.

AuditAI

AI Call Automation

EduQuery

Work experience.

AI Research Engineer

GenAI and Full Stack Development Intern

Data Analysis and Visualization Intern

Full Stack Development Intern

The degrees.

M.Sc. Artificial Intelligence for Industrial Applications

B.Tech. Computer Science and Engineering

Other projects.

Sound Slice

Neural Image and Video Style Transfer

Explainable AI for COVID Detection

Technical toolkit.

Programming Languages

AI / ML

Backend

Frontend

Databases and Data

Infrastructure and Tools

Say hello.

ShreshtthKumarAgarwaal

Engineer and master’s student.

Selected work.

AuditAI

AI Call Automation

EduQuery

Work experience.

AI Research Engineer

GenAI and Full Stack Development Intern

Data Analysis and Visualization Intern

Full Stack Development Intern

The degrees.

M.Sc. Artificial Intelligence for Industrial Applications

B.Tech. Computer Science and Engineering

Other projects.

Sound Slice

Neural Image and Video Style Transfer

Explainable AI for COVID Detection

Technical toolkit.

Programming Languages

AI / ML

Backend

Frontend

Databases and Data

Infrastructure and Tools

Say hello.