Nishchal Kumar Singh

Technical Skills

Python, C/C++, JavaScript, SQL

FastAPI, Flask, REST API, Node.js, PostgreSQL, MongoDB

LangGraph, LlamaIndex, RAG, ChromaDB, CrewAI

PyTorch, TensorFlow, Scikit-learn, NLP, Transformers, LLM Fine-tuning (QLoRA, GRPO)

Git, Docker, Postman, Linux (WSL), Google Colab, Hugging Face

DSA, OOP, DBMS, Operating Systems, Computer Networks

Tool-Augmented Reasoning Agent (TARA)

Source Code ↗

PyTorch · Unsloth · Hugging Face · Flask · Docker · SymPy

Fine-tuned Qwen2.5-3B using Unsloth 4-bit QLoRA and GRPO, improving GSM8K accuracy from 74.2% → 84.7% through autonomous error-correction — all on a single 16 GB GPU.
Engineered a ReAct state machine generating 17.3K execution-aligned trajectories via 40K+ live tool executions; achieved a 90% execution success rate across 8 custom GSM8K metrics.
Designed verification and reflection rewards that recovered from initial reasoning failures in 50% of evaluation cases; deployed as a Dockerized Flask/Gunicorn inference service.

LogiBot — AI-Powered Logistics Router

Source Code ↗ Live Demo ↗

FastAPI · React · LangChain · Gemini · Pydantic · PostgreSQL

Engineered an AI-powered logistics backend using a custom Genetic Algorithm with ORS distance matrices, live traffic & weather APIs to optimise routing sequences under 2.5 s.
Architected a LangChain + Gemini agentic copilot with 90% tool-selection accuracy, enabling natural-language parsing, database updates, and dynamic route re-optimisation.
Built a React.js / Leaflet frontend with dual-mode NLP/map input and a 4-layer traffic heatmap; optimised AI state synchronisation to render 50+ stop routes smoothly.

Attention-Based Image Caption Generator

Source Code ↗

TensorFlow · MobileNet · LSTM · Bahdanau Attention · Flask · Docker

Built an encoder-decoder pipeline using MobileNet encoders and custom Bahdanau attention layers; achieved a 25% accuracy gain and 30% reduction in inference time.
Optimised inference via beam search and an eager TensorFlow graph warm-up, eliminating cold-start delays and guaranteeing API responses under 2 s.
Architected a production-ready Docker / Gunicorn deployment using a single-worker multi-threaded config to prevent TF segfaults and enforce a strict 3 GB memory limit.

Machine Learning Library from Scratch

Source Code ↗

Python · NumPy · Google Colab

Built Linear & Logistic Regression, KNN, Decision Trees, K-Means, and N-layer Neural Networks entirely from scratch — zero high-level ML libraries.
Used only NumPy for all matrix operations, gradient computations, and model logic, deepening low-level understanding of ML fundamentals.
Tested and documented on Google Colab with reproducible notebooks for each algorithm.

SmartMail — RAG-Powered Email Assistant

Source Code ↗

Python · RAG · LLM · LangChain

Built an automated email response system using Retrieval-Augmented Generation (RAG) to extract context from email chains and generate accurate, on-topic replies.
Integrated LLM-backed semantic retrieval to surface relevant prior context before generation, enabling smart inbox automation without manual templates.
Won 1st place among 1300+ participants at HackFest IIT (ISM) Dhanbad's GoFloww Atom Mail challenge.

MNIST Digit Classifier from Scratch

Source Code ↗

Python · NumPy · Neural Networks

Achieved 97.2% test accuracy on MNIST by implementing a full neural network — forward pass, backpropagation, and weight updates — entirely from scratch.
Modularised each component (layers, activations, loss functions, optimisers) into separate, reusable files; model architecture and optimisers are fully plug-and-play.
The NeuralNetwork class dynamically chains all layers and performs complete forward and backward propagation without any deep learning framework.