Yash Vardhan Malik

Full-stack developer and data practitioner with hands-on experience in AI/ML, data analysis, and production engineering — shaped through multiple industry internships and real-world projects.

yash60066@gmail.com LinkedIn GitHub Portfolio

Experience

Machine Learning InternMay 2025 - Jun 2025

Centre for Railway Information Systems (CRIS)

Processed 450K+ freight records and engineered 20+ domain-specific features for ML-based time prediction pipelines on real Indian Railways operational data.
Boosted model accuracy by 27% using LightGBM with hyperparameter tuning; benchmarked 5+ algorithms (Random Forest, XGBoost, SVR) achieving R²: 0.78 (loading) and R²: 0.72 (unloading).

Data Analyst InternJul 2024 - Jul 2024

TRINITI

Completed coursework in Data Analytics, Statistics, and visualization tools (Excel, MySQL, Tableau, Python) — applying statistical methods to analyze, clean, and transform complex datasets, delivering actionable insights through reports and dashboards to support business decision-making.

Software Development InternMay 2024 - Jul 2024

General Software & Consultancy Services (GSCS)

Managed the AppKube cloud monitoring project overseeing 10+ AWS EC2 instances; reduced system monitoring time by 30% through automated dashboards and real-time alert pipelines.
Delivered EC2 reporting solutions serving 5+ enterprise client environments, improving operational observability and incident response efficiency across production infrastructure.

Projects

Creatorjoy RAG — YouTube Content Intelligencerender.com

Built a full-stack RAG platform where creators paste two YouTube URLs and ask why one outperformed the other — powered by a LangGraph StateGraph with 4-class query routing, fastembed ONNX BGE-small embeddings on pgvector HNSW, 3-stage transcript cascade (youtube-transcript-api → Supadata.ai → Groq Whisper V3), and SSE streaming with inline timestamp citations.

Engineered timestamp-aware chunking with deterministic intro_5s / intro_15s segments enabling metadata-filtered hook retrieval;
Deployed on Render + Vercel + Neon at $0.013/session (3.5× cheaper than GPT-4 equivalent) with LangGraph PostgresSaver.

DiagnoseAI — Codebase & Meeting Intelligence SaaSdiagnose-ai.vercel.app

Built and deployed a full-stack SaaS platform enabling developers to index any GitHub repository and query codebases via natural language — powered by a LangChain + RAG pipeline with pgvector (384-dim HuggingFace BGE embeddings) and Groq Llama 3.1 for context-aware answers with file-level references.

Engineered a meeting intelligence module transcribing audio (MP3/WAV/M4A up to 50MB) via Groq Whisper Large V3 and extracting action items, decisions, and topics via LLM with timestamps — scoped per project for contextual relevance.
Implemented AI commit summarization, team collaboration via invite links, Razorpay credit billing (1 credit = 1 file indexed), and Clerk auth — backed by Neon PostgreSQL, Uploadthing, and Vercel.

Multi-AI Hedge Fund Simulator

Built a multi-agent trading simulator encoding strategies from 15+ legendary investors into role-based agents (fundamentals, sentiment, risk arbitrage) with an agentic orchestration layer for signal-sharing and unified portfolio decision-making.

Backtested agent-driven strategies on historical market data — yielded 12% higher returns and 10% lower drawdown vs. baseline index strategies across multiple market regimes; demonstrated applied multi-agent system design with measurable financial impact.