Data systems · applied ML · research

Harshil Agrawal

I build data and AI systems that make it from an ambiguous business problem to a deployed, testable product. Then I examine how they behave in the real world.

Read selected work Inspect ML demo

Selected work

Systems with an evidence trail.

ALL PROJECTS →

Featured case / 01

Production LLM Platform

Conversational ordering for pharmaceutical distribution.

Owned the build of a WhatsApp inquiry platform that pairs Azure OpenAI intent classification with a Snowflake-backed product and conversation layer, designed for always-on ordering workflows.

System trace

WhatsApp inquiry→Intent classification→Product context→Snowflake logs

1,500+product catalog entries in scope

ROLE

Data & AI Engineering Intern
Prowess Consulting

Azure OpenAISnowflakeAzure App Service

Employer-owned work

Built during an industrial internship. The system is shown here at an architectural level.

02 / SALESVITALS

An analytics stack designed around operational constraints.

Built a FastAPI data product for pharmaceutical field representatives, then introduced a SQLite snapshot mode and GitHub Actions validation so insight delivery did not depend on recurring warehouse queries.

Delivery path

Source data→Versioned snapshot→FastAPI→Field insight

0recurring Snowflake query cost for snapshot delivery

FOCUS

Reproducibility, data access, and deployment discipline.

FastAPISQLiteGitHub Actions

Employer-owned work

Built during an industrial internship. The delivery decision and its constraint are shown here.

03 / PRIVACY-AWARE ML

Banking PII protection for sensitive text workflows.

Built a local-first service that extracts and anonymizes personally identifiable information in banking text, combining multilingual transformer models with schema-based validation to make each detected entity inspectable.

Protection path

Sensitive text→PII extraction→Schema validation→Anonymized output

0.866reported micro-F1 across 32,017 labeled entities

FOCUS

Privacy-preserving ML, local inference, and validation-aware entity extraction.

FastAPIReactHugging FacePII

View source ↗

Open source

Repository includes the FastAPI service, React interface, and local model setup.

Research direction

Forecasting, decision-making, and evaluation under uncertainty.

My current research explores a hybrid LSTM + Deep RL framework for decision support and automated trading, evaluated on NIFTY-50 backtests. The IEEE ICCICT 2026 paper is under publication.