Ayush Kumar Poddar

AI Researcher

Bangalore, Karnataka, India1 yr 9 mos experience

AI EnabledAI ML Practitioner

Key Highlights

Built advanced AI systems from scratch.
Improved OCR accuracy for ancient manuscripts.
Developed data-driven solutions for telecom industry.

Stackforce AI infers this person is a skilled AI/ML Engineer specializing in Natural Language Processing and data-driven solutions.

Contact

Skills

Core Skills

Natural Language Processing (nlp)Machine LearningData Science

Other Skills

PyTorchFastAPIA/B TestingXGBoostTeamworkMarketing StrategyData EngineeringAgentic AI DevelopmentLangChainModel Context Protocol (MCP)Retrieval-Augmented Generation (RAG)SQLComputer VisionApplication Programming Interfaces (API)Large Language Model Operations (LLMOps)

About

I'm a final-year undergrad at IIT Roorkee who builds AI systems, the kind that go from a blank file to a working product people can actually use. I've worked on language models that read ancient manuscripts, multi-agent systems that research and reason on their own, and machine learning pipelines that help businesses make smarter decisions. My work spans the full stack, from training and fine-tuning models to deploying them as production-ready APIs using FastAPI, Docker, and AWS. I work with Python, PyTorch, LangChain, and the modern GenAI stack, building everything from RAG pipelines and LLM-powered applications to data science solutions with real business impact. I'm drawn to problems that are genuinely hard, where the solution isn't obvious and the stakes are real. I think deeply about not just what to build, but why it matters. Part engineer, part tinkerer, part curious generalist. Let's connect if you're working on something meaningful, or just want to talk about AI, startups, or what's actually worth building.

Experience

1 yr 9 mos

Total Experience

10 mos

Average Tenure

Current Experience

Indian institute of technology, roorkee

Research Intern

May 2025 – Feb 2026 · 9 mos · Roorkee · On-site

Building production-grade LLM pipelines for Sanskrit manuscript digitization — one of the hardest NLP domains due to degraded OCR and scarce training data.
Fine-tuned LLaMA-3.1-8B with LoRA (r=16, α=32) + 4-bit QLoRA on a single A100 — improved BLEU-4 40% (18.3 → 25.6) at 3x lower training cost
Designed a multi-stage LangChain + LangGraph pipeline for OCR correction, retrieval, and reasoning with modular, scalable architecture
Built a TrOCR + LLM post-correction pipeline that cut character error rate from 23% → 14% across 12K samples
Benchmarked 6 post-OCR strategies (rule-based, n-gram LM, seq2seq, prompt-based LLMs, RAG, hybrid) with automated eval pipelines; deployed via FastAPI for reproducible experiments

Natural Language Processing (NLP)PyTorchMachine Learning

Quantiq analytics

Data Science Intern

Jan 2025 – Apr 2025 · 3 mos · Remote

Applied ML to real business problems — churn prediction and retention optimization on production-scale telecom data.
Built an XGBoost + logistic ensemble churn pipeline on 180K records; AUC 0.89, with SHAP attribution to surface top business drivers
Ran 3 sequential A/B experiments on notification timing and copy — personalised timing yielded +12% 7-day retention (p<0.001), adopted into product roadmap

A/B TestingXGBoostData ScienceMachine Learning