Deep Rathi

Wrote the AdaBoost (SAMME) classifier that now ships in the linfa-ensemble crate, version 0.8.1 on crates.io: the boosting loop with adaptive sample reweighting, per-model alpha weighting and weighted majority voting, behind a typed ParamGuard hyperparameter builder. Merged into rust-ml/linfa after maintainer review. Also fixed an ndarray version drift that was breaking CI across the workspace sub-crates.

An LLM-driven data-analysis agent that takes a raw dataset and a question in plain language, then plans and runs the analysis: cleaning, feature preparation and generated insight. Built for the IIM Ahmedabad AI Hackathon 2024, where it took 1st place, and later written up and published in AIP Conference Proceedings.

Low-latency C++ client for the Deribit derivatives exchange REST API: authentication, order placement, edit and cancel, open orders, positions and order book. The request path is built for latency — a pooled curl-handle transport, preallocated 16 KB response buffers, an order-book cache, TCP keepalive, a 10-minute DNS cache and branch-prediction hints on the hot path. Every trading and market-data request is timed with std::chrono, with round-trip latency aggregated in a fixed-capacity circular buffer reporting median and 95th percentile. A local Boost.Beast WebSocket server broadcasts order-book updates to subscribed clients.

Intelligent personalized learning platform powered by GROQ LLM API. Generates custom study plans in under 5 seconds based on user goals and learning style. Includes progress tracking, adaptive difficulty adjustment, and exportable study materials. Leverages fast inference for real-time educational content generation.

SadTalker wrapped behind a serving API over a 13-persona library, returning an avatar video from a voice clip in under 5 seconds. Adds persona selection, idle and blink handling and an inference entry point suited to being called from a product rather than a notebook.

Algorithmic system that generates unlimited data structures and algorithm problems with automatic validation. Features difficulty scaling, constraint generation, and solution verification. Includes comprehensive test case generation and performance analysis. Useful for interview preparation, competitive programming, and algorithm education.

Collaborative filtering-based recommendation engine with advanced feature engineering and model optimization. Implements matrix factorization, neural collaborative filtering, and hybrid recommendation approaches. Optimized for both accuracy and computational efficiency with production-ready data pipelines.