M.Tech CSE · IIIT Hyderabad

Vikash Kumar

I'm a Software Engineer

I build software that works at scale and models that learn: from training small Transformers from scratch to shipping a satellite-powered lake forecasting app. Currently deepening my work in language models and retrieval.

View projects GitHub LinkedIn
0Projects
0GATE CS percentile
0IIIT-H PGEE '25

A bit about me

Portrait of Vikash Kumar

I'm a software engineer pursuing an M.Tech in Computer Science and Engineering at IIIT Hyderabad. Before that I worked in industry as an intern at Odessa Technology and as a software engineer at GirnarSoft Finserv, where I shipped a payout module, faster APIs and a compliance-driven KYC integration.

These days I spend my time on language models, retrieval and geospatial machine learning: pretraining Transformers from scratch, building RAG systems over official documents, and working on EcoTwin, a research project that forecasts the ecological future of India's lakes.

I like systems that are well tested, measurable and explainable, and I enjoy going down to the low level too: a BitTorrent client and a Unix shell are on my list of things built from scratch.

🎓

Studying

M.Tech in CSE at IIIT Hyderabad, July 2025 to present.

🧠

Focus

LLMs, retrieval-augmented generation and geospatial ML.

💼

Previously

Software Engineer at GirnarSoft Finserv; Software Engineer Intern at Odessa Technology.

Education

Experience

Sept 2022 – June 2023
Gurugram, India
  • Fully functional new module developed for the payout mechanism (improved the work efficiency of multiple teams)
  • Created highly optimised multi-purpose reusable APIs (improved the response time from 18 sec to 2 sec)
  • Implemented a high-priority KYC feature integration following new government mandates, ensuring compliance and operational continuity under strict deadlines

Software Engineer Intern

Odessa Technology
Mar 2021 – June 2022
Bengaluru, India
  • Wrote highly maintainable code using the C# ASP.NET framework backend to incorporate new features
  • Created procedures and wrote queries for DDL and DML operations on the database for migration, backup and optimisation using SQL Server; designed and developed reports and documents using SQL Server

Projects

Things I've built, from low-level systems to language models and a live ecosystem-forecasting app.

Live · Research 2026

EcoTwinEcosystem Forecasting for India's Lakes

A digital twin of lakes: satellite imagery in, auditable 3-year forecasts out. Built in a team of three with Ramachandra Prasad P, IIIT Hyderabad.

0Lakes
0Indian states
0Embeddings
3-yrForecasts
  • Geospatial pipeline over Sentinel-2 imagery for 829 lakes across 26 states (2016–2025), producing 8,227 foundation-model embeddings with zero pipeline failures
  • Decoder-free forecasting using trajectory-based analog retrieval over pgvector, giving auditable 3-year forecasts with confidence-aware abstention
  • Full-stack app (FastAPI, React, PostgreSQL) with automated PDF reports, demonstrated on five Ramsar wetland case studies
PrithvipgvectorFastAPIReactPostgreSQL
Live Website GitHub ecotwin-lakes.vercel.app
LLM · Indic NLPAug – Sep 2026

GPT-style Language Models from Scratch

Two small Hindi and Nepali language models, built the hard way: from an empty file.

  • Implemented a decoder-only Transformer from scratch; pretrained separate Hindi and Nepali 25M-parameter LLMs on ~650M tokens each
  • 12-stage corpus pipeline with MinHash/LSH near-dedup and decontamination; SentencePiece BPE tokenizers trained from scratch
  • Mixed-precision training, bit-exact checkpoint resume and an LLM evaluation suite; validation perplexity 26.2 (Hindi), 41.5 (Nepali)
  • Reasoning fine-tune: 0% → 41–43% on seen data, 6–20% on held-out splits
PyTorchCUDATransformersSentencePiece
GitHubViserionBuild/nanogpt-hi-ne
RAG · GenAIApr – May 2026

Government Services RAG Chatbot

Ask questions about Indian government services; get answers grounded in the official PDFs.

  • End-to-end Retrieval-Augmented Generation chatbot over 143 official PDFs across five services (passport, voter ID, driving licence, income tax, Ayushman Bharat)
  • PyMuPDF → chunking → MiniLM embeddings → FAISS → top-5 retrieval, with a relevance cutoff that returns "no relevant documents" instead of guessing
  • Groq LLM API with citation-backed answers, model fallback and conversation memory
  • Streamlit UI with an English/Hindi interface and PDF download
PythonFAISSGroqStreamlit
GitHubIIITH-2025-27/infobridge
Multimodal · Deep LearningFeb – Mar 2026

MERGE: Emotion-Driven Lyrics Generation

Looks at a face, feels the emotion, writes the song lyrics, in rhythm.

  • Multimodal flow: a face image (MTCNN + DenseNet-121) or text (RoBERTa) is classified into six emotions, then an emotion-conditioned GPT-2 writes lyrics
  • Rhythm-aware reranking picks candidate lines by syllable meter and rhyme for better structure
  • Full-stack web app (FastAPI, React, MongoDB) with OTP-verified signup, forgot-password, saved chat history and loading and error states
DenseNet-121GPT-2RoBERTaFastAPIReact
GitHubIIITH-2025-27/NLP_Project
Networking · SystemsSep – Oct 2025

BitTorrent Implementation

A peer-to-peer file sharing system with a tracker, seeders and verified chunks.

  • Multi-threaded tracker (users, groups, file and seeder registry) and peers that both download and seed over TCP
  • 512 KB chunks with per-chunk SHA-1 verification and multi-seeder download with failover; finished downloads re-register as seeders
  • POSIX threads with mutex-protected shared state and a command-line client for accounts, groups, upload and download
C++POSIX ThreadsTCP SocketsSHA-1
GitHubviserion999/bit_torrent
Operating SystemsAug – Sep 2025

POSIX Shell Implementation

A tiny Unix shell: parse it, fork it, pipe it.

  • Interactive shell in C++ that parses commands, runs pipelines and command sequences, and keeps persistent history
  • Process management: foreground and background jobs, process groups, and Ctrl+C and Ctrl+Z forwarding
  • Pipes and I/O redirection, plus built-ins such as cd, ls, echo, pwd, history and recursive search
C++POSIX APIsfork / execvpReadline
GitHubviserion999/posix_shell

Technical Skills

⌨️Languages

PythonC++CJavaC#ASP.NET

🧠Generative AI & NLP

LLMsTransformersHugging FaceFine-tuningRAGAgentic AILLM evaluation

🗄️Databases

PostgreSQL / PostGISVector searchMySQLSQL Server

🛠️MLOps & Tools

DockerFastAPIGitGitHubLinuxJupyter NotebookVS CodeBitBucketJiraConfluenceReplicon

Achievements

🏆0

Secured in IIIT-H PGEE '25

💯0

GATE CS 2024 percentile

📊0

GATE DA 2025 percentile

Let's build something great.

Have a role, a project or an idea you'd like to talk about? My inbox is open.

viratvikash151@gmail.com LinkedIn GitHub HackerRank