Software Engineering + Applied AI @ IIT Kharagpur

I build reliable backend systems and applied AI products with measurable performance.

   

#include <iostream>
#include <string>
#include <vector>

class AyushGoel {
public:
    std::string university = "IIT Kharagpur";
    std::string focus = "Backend Systems, Algorithms & Applied AI";

    std::vector<std::string> toolkit = {
        "C++", "Python", "FastAPI", "PostgreSQL", "Redis", "Docker"
    };

    void sayHi() const {
        std::cout << "I build systems that are reliable, measurable, and useful.\n";
        std::cout << "Build -> benchmark -> improve -> repeat.\n";
    }
};

int main() {
    AyushGoel ayush;
    ayush.sayHi();
    return 0; // shipped successfully
}

I am a B.Tech. (Hons.) undergraduate in Chemical Engineering at IIT Kharagpur with a Micro-specialization in Artificial Intelligence and Applications.

My work sits at the intersection of backend engineering, algorithms, and applied AI. I enjoy turning ideas into systems whose reliability, latency, throughput, or accuracy can be measured - from LLM infrastructure and document intelligence to computer vision and robot motion planning.

Currently deepening my work in scalable APIs, distributed systems, system design, and performance-focused software engineering.

FastAPI Redis PostgreSQL Docker Async I/O

A production-minded gateway for routing, caching, rate limiting, fallback, and observability across LLM providers.

• Built three routing strategies across OpenAI and Gemini adapters with bounded retries and automatic fallback.
• Implemented exact-response caching, an atomic Redis Lua token bucket, and PostgreSQL p50/p95 metrics.
• Benchmarked 1,600 local requests across four concurrency levels, reaching 164 RPS with zero failures.

Repository · Live demo · API docs

FastAPI YOLO MediaPipe Computer Vision GitHub Actions

A review-first computer vision system that turns noisy frame-level detections into persistent behavioral alerts.

• Designed a five-rule AlertEngine with duration, grace-period, and cooldown controls.
• Improved throughput by about 19%, reducing average frame latency from 279.4 ms to 234.9 ms.
• Built candidate tracking using IoU, center distance, and area similarity; validated with 79 automated tests.

Repository

Python SciPy NumPy KDTree A* OpenCV

A deterministic motion-planning system combining probabilistic roadmaps, spatial search, collision checking, and A*.

• Constructed 502-node roadmaps with 2,233-5,312 collision-validated edges.
• Completed graph searches in 1.5-19.6 ms.
• Produced 21%-23% shorter successful paths than goal-biased RRT across three benchmark maps and 60 matched-seed runs.

Repository

• Software Development Engineer Intern - Wasserstoff
  Built a FastAPI OCR pipeline using Tesseract, Docling, OpenCV, Docker, caching, and concurrent processing. Reduced character error rate from 0.481 to 0.080 and repeat extraction time from 30.1 s to 0.4 s.

• Undergraduate Research Intern - University of Manchester & University of Liverpool
  Engineered forecasting data and evaluation pipelines over 225 Zarr stores and 29M+ grid-point predictions, adapting an 89.1M-parameter weather model while training only 0.23% of its parameters.

• Machine Learning Developer Intern - Chi SquareX Technologies
  Built an OpenAI Gymnasium options-trading environment and trained PPO/A2C agents with Stable-Baselines3, achieving a 54.05% backtest return over ten months of historical data.

• Undergraduate Researcher - Aerial Robotics Research Group, IIT Kharagpur
  Worked with ROS, Gazebo, MAVROS, drone simulation, perception algorithms, and autonomous-system tooling.

• Codeforces Expert with a peak rating of 1638 and 500+ algorithmic problems solved.
• Global ranks 1,863, 2,320, and 3,704 in Codeforces Rounds 1110, 1111, and 1115.
• Rank 22 in the Optiver Quantitative Trade-a-thon 2026.
• All-India Rank 7 in Round 1 of Convolve 4.0, the Pan-IIT AI/ML Hackathon.

_{Live, automatically updated, and clickable.}