Hi, I'm Khushali. I build reliable backend and AI systems.

Newest · October 2026

LLM Gateway

83 tests · cache hits about 89x faster (simulated provider)

The problem

When several teams call LLM providers directly, nobody sees total cost, repeated questions are paid for every time, one outage takes everything down, and customer data can leak into prompts.

What I built

An OpenAI-compatible gateway: apps change only their base URL, and it adds authentication, rate limits, monthly budgets, PII redaction, caching, provider failover and per-team cost reporting.

What I learned

A load test exposed a database connection-pool bug that looked healthy in the latency numbers. I fixed it and pinned it with a regression test. Built in three days and runs locally.

Python, FastAPI, Postgres, Redis, SQLAlchemy, Docker Compose, pytest, Locust

View the code on GitHub

TriModel Financial Analyzer

84% stock trend prediction accuracy

What I built

Time-series, technical and BERT-based sentiment models, plus scalable data pipelines that collect, clean and align price data, indicators and financial news. Jan to Dec 2024, University of Colorado.

Impact

Improved prediction accuracy to 84%, and cut training time by 30% by reusing preprocessing outputs across experiments.

Python, BERT, Time-series models, Data pipelines

View the code on GitHub

Star War Simulator

Real-time multiplayer

What I built

A multiplayer, turn-based space battle game. Django and WebSockets carry the moves, and Redux keeps game state in sync across players with low latency.

Django, WebSockets, Redux

Movie Recommendation System

Two filtering methods, one recommender

What I built

A personalized recommender that blends content-based and collaborative filtering, with requirements documented in IEEE SRS format.

Recommender systems, Python

View the code on GitHub