Adithya Naik-Security Intern

An LLM agent that scores resumes against a job description with forced structured output

• A self-directed project built specifically to gain hands-on experience with tool-calling, prompt engineering, and workflow automation – not just to claim familiarity with them.

What it does

• Given a job description and a resume, the agent returns a structured JSON object (not a paragraph to interpret) containing an overall score, sub-scores for skills match and experience relevance, a list of specific named gaps, and a one-line summary – ready to sort, filter, or feed into another system automatically.

What I did

• Used the LLM's tool-calling API to force structured JSON output via a declared schema, rather than parsing free text with regex, guaranteeing the response always matches the expected format.
• Applied deliberate prompt engineering: explicit scoring criteria per field, a few-shot example of the expected reasoning style, and ordering the schema so the model's reasoning field comes before the numeric scores, so reasoning conditions the score instead of just justifying it after the fact.
• Built a Streamlit UI for uploading resumes and viewing a ranked, sortable results table.
• Wrapped the scoring logic in a Flask API and connected it to an n8n workflow: a form submission triggers the agent via webhook, which then sends an email notification with the result, tested against a real SMTP sandbox (Mailtrap).
• Built a small Lovable app that posts structured data to an external webhook, to get direct hands-on experience with a no-code builder rather than secondhand familiarity.
• Included a mock API client so the entire pipeline can be built, run, and tested end to end without spending any API credits.

Tech stack: Python, Gemini API (tool-calling), Streamlit, Flask, n8n, Lovable, JSON Schema

Predicting risky drug combinations using graph and sequence models

• Pairs Graph Neural Networks with Recurrent Neural Networks to flag harmful drug interactions before they happen.

What it does

• The model represents drugs and their properties as a graph structure and combines this with sequential modeling to predict whether a given pair of drugs is likely to produce a harmful interaction, aimed at supporting safer prescribing decisions.

What I did

• Combined Graph Neural Networks (GNNs), to capture relational structure between drug properties, with Recurrent Neural Networks (RNNs), to capture sequential/sequence-based patterns, rather than using either architecture alone.
• Improved prediction accuracy by 12% over baseline approaches.
• Built and evaluated the model in PyTorch, using scikit-learn for supporting analysis of drug properties.
• Framed the project around a concrete pharmaceutical research use case: flagging harmful combinations early, rather than treating it as a purely academic exercise.

Tech stack: Python, PyTorch, Graph Neural Networks (GNNs), Recurrent Neural Networks (RNNs), scikit-learn