Projects

AI Projects You Can Try Live

Each one solves a real problem and is live to try: an AI assistant that cites its sources, a churn model that tells teams who's about to leave, a log analyser that caught 588 login attacks, and 4 more.

nexus-ai.taaranjain.comLive

01RAG · LangGraph Agent

Nexus AI

An AI assistant that shows its sources and confidence

Problem: Chat assistants give you an answer and hide the rest: whether it came from the web, your files or the model's memory, and how confident it is. That's exactly what matters when you're checking a policy or making a real decision.
What I Built: A six-node LangGraph pipeline routes each question to web search, your documents or the model, and streams every step with its time and token cost. Answers cite the exact file, page and line, and a calibrated decision model scores every option when you're choosing.
Result: Every answer is auditable: route, sources, timing and token cost stream live alongside the reply

0.4s

Decision board scored

~2×

Faster cold start

LLM providers

How it was measured: Local testing with real API keys. Cold start is the backend's import time before and after the change.

LangGraph
LangChain
Groq
FastAPI
Next.js
TypeScript
Hugging Face

Try the Live Demo Read the Case Study

decidex.taaranjain.comLive

02ML Churn Prediction

DecideX

Tells customer teams who might leave, why, and what to do

Problem: Customer teams usually find out a customer is leaving when it's too late, and a churn score on its own doesn't say what to do next, who should do it, or how much the account is worth saving.
What I Built: A decision app that reads account notes with an LLM and scores each customer with a calibrated churn model, then recommends what to do, who owns it and how urgently, with plain-English reasons, the revenue at risk and the smallest offer that brings the customer back under the risk line.
Result: Catches 85.6% of leavers on 1,409 held-out customers from the IBM Telco dataset (ROC-AUC 0.844, PR-AUC 0.653), with probabilities calibrated to an ECE of 0.018

85.6%

Leavers caught on held-out data

−71%

Missed churners after cost tuning

0.018

Calibration error (ECE)

How it was measured: Public IBM Telco dataset (7,043 customers), stratified 60/20/20 split, scored on the 1,409 held-out customers.

Python
scikit-learn
LightGBM
FastAPI
Next.js
TypeScript
SQLite

Try the Live Demo Read the Case Study

logvera.taaranjain.comLive

03ML Anomaly Detection

Logvera

Turns raw logs into incidents, attacks and root causes

Problem: Reading a raw log to find out what went wrong is slow, and most log tools need agents, accounts and setup before they show you anything. Engineers end up grepping by hand, and an attack spread across thousands of lines is easy to miss.
What I Built: Drop in a log or an archive and Logvera detects the format, learns message patterns with a Drain-style parser, and runs a five-model anomaly ensemble plus a rule-based security scan. An LLM layer writes summaries and answers questions with cited lines, seeing only redacted data, and every upload is encrypted and deleted after an hour.
Result: On a 34,492-line sample log it found 3 unusual periods, 8 security findings including 588 credential-stuffing attempts, and 84 distinct message patterns

Models voting on anomalies

588

Attacks found in sample log

Backend tests

How it was measured: A bundled sample of one day of e-commerce Nginx traffic with planted incidents, not a production log.

Python
FastAPI
scikit-learn
NumPy
Next.js
TypeScript
WebSockets

Try the Live Demo Read the Case Study

More Shipped Projects

All Case Studies