Machine Learning Engineer · Delhi, IN
Teaching noisy data to reason & decide.
I'm Yash Vasudeva - building end-to-end predictive models, deep learning pipelines, and empirical AI benchmarks that bridge research rigor and production systems. B.Tech IT @ BPIT, GGSIPU (8.75 CGPA) →
Fine-tuned version of Qwen2.5-3B-Instruct optimized for Data Science and ML coding tasks using QLoRA on 16.4K instruct examples with Unsloth. Achieved an 86% win rate over the base model with +109.5% BLEU and +66.9% ROUGE-L score gains.
Enterprise-grade, latency-aware 6-stage adaptive RAG platform with Apple HIG web UI. Features dual-filter extraction, Qdrant HNSW + BM25 hybrid retrieval, calibrated RRF (k=20), continuous confidence estimation with top-1 consensus guard, blended cross-encoder reranking, and 97.3% abstention accuracy.
AI-powered resume intelligence platform and analysis engine. Evaluates resumes with zero fluff, extracts structured factual profiles via Docling and PyMuPDF, identifies ATS vulnerabilities, critiques bullet points, and benchmarks candidates against 12+ tech roles and job descriptions.
AI-powered hackathon judging platform. Upload a CSV of team submissions and a multi-agent pipeline - five parallel judge agents, a bias auditor, and a chief judge - scores every project automatically, generates written feedback, and produces a ranked leaderboard.
Controlled paired experiments across 20 Newsgroups and IMDB show classical preprocessing lifts topic F1 (+0.043) while hurting sentiment (-0.011), and sentence embeddings hinge on truncation and morphology.
An empirical look at when model capacity stops paying for itself, and when it actually becomes a liability.
A controlled experiment across 4 open-weight LLMs (82M to 410M) testing long-form narrative coherence reveals that parameter count does not predict collapse - knowledge distillation does.
Have a problem worth modeling?
Open to Machine Learning Engineer, Data Science, and AI Engineering opportunities.