Tharun C

Who I Am

About

I'm a final-year AI & Machine Learning undergraduate in Bengaluru, and I care about the same thing on both sides of the stack: does it actually work in production, not just in a demo.

That's meant fine-tuning transformer models on real-world CVE data, building fare-prediction microservices, and shipping voice pipelines that don't fall over when the LLM returns malformed JSON — alongside REST APIs, full-stack apps, and clean, scalable architectures.

I'm currently looking for Software Engineering / AI internships where I can work on real systems with a real team.

🎓

AI & ML Student

Dayananda Sagar University, expected 2027

📍

Based in Bengaluru

Open to relocation for the right role

⚡

AI + Full Stack

From model fine-tuning to shipped UI

🎯

Currently seeking

SWE / AI-ML internships, 2026–27

Toolbox

Skills

AI & ML · Primary Focus

Python

PyTorch

HuggingFace Transformers

Scikit-Learn

Languages

Java

TypeScript

JavaScript

SQL

Web & Backend

React

Next.js

Node.js

FastAPI

Databases & Tools

PostgreSQL

Firebase

Git

AWS

Featured Work

Projects

01 · Security · ML

Neural Vulnerability Detector

Fine-tuned CodeBERT for security analysis, scoring F1 0.782 across 190K+ CVE functions (BigVul) and catching 741 of 1,026 vulnerabilities with only 128 false alarms.

F1 0.782 · Recall 0.722 · Precision 0.853 · 250ms inference

124M-param CodeBERT, weighted cross-entropy + tuned threshold for 16:1 imbalance

Shipped 3 ways: FastAPI, GitHub Action, VS Code extension

Python · PyTorch · FastAPI · ONNX · CUDA

02 · AI · Voice

InterviewX

AI-powered mock interview platform with a full real-time voice loop and per-user isolation.

Gemini 2.5 Flash generates role-specific questions on demand, difficulty tuned via prompt design

Full voice interview loop: Deepgram STT → Gemini evaluation → TTS, with malformed-JSON recovery

Per-user isolation: 3-collection Firestore schema, HTTP-only session cookies

Next.js · Gemini AI · Deepgram · Firebase

03 · Mobile · ML

Tripzo

AI-powered ride booking application with live, transparent fare prediction instead of static pricing.

Fare prediction on 7 real-world factors via Scikit-Learn microservice

4 Google Maps APIs replace static pricing with live traffic/weather-aware fares

Transparent breakdown + confidence score, surge up to 2.0x

Razorpay signature verification + Clerk auth on normalized PostgreSQL (NeonDB)

Independent Node.js/Express API and Python/FastAPI ML service, decoupled for scale

Rides created only after payment confirms, closing the unpaid-booking gap

React Native (Expo) + Zustand app talking to a fully typed TypeScript backend

React Native · Node.js · FastAPI · Scikit-Learn · PostgreSQL

Open Source

Collaborations

Collaboration · 2 contributors

AutoHeal

Self-healing ML pipeline that detects concept drift and automatically retrains itself — no human in the loop.

Rolling accuracy fell to 33.6% avg (from a 96.7% clean baseline) across 5 runs before drift was caught

Auto-retrained and recovered accuracy to 81.6% with zero manual intervention

Outperformed scheduled retraining (36%) and no-monitoring (27%) baselines at 40% during-drift accuracy

My part: Designed the baseline comparison framework (no-monitoring vs. scheduled vs. AutoHeal) and led the statistical analysis across runs; built the full visualization suite.

with @ShreyaSindhe

Python · scikit-learn · Pandas · Matplotlib

Beyond The Classroom

Journey

2022 · Getting Started

First steps into AI & ML

Started learning AI & Machine Learning fundamentals, pairing every new concept — from classic ML to neural nets — with a small project to make it stick.

2023 – 2024 · Certifications

Built the foundations

Earned AWS Academy Graduate (Cloud Foundations), IBM SkillsBuild (Cloud Computing Fundamentals), and NPTEL Foundations of Deep Learning (IIT/IISc — Elite, 82%), all alongside coursework.

2025 · Open Source

Shipped it into real workflows

Built and shipped a GitHub Action that auto-comments CWE vulnerability findings on pull requests — taking the vulnerability detector out of a notebook and into real developer workflows.

2026 · Now