AI Engineering • New Delhi / Gurugram
Bridging the gap between frontier foundation models and deterministic software. Specializing in autonomous multi-agent orchestration, grounded RAG pipelines, and real-time deep learning inference that ships.
Production stacks, neural toolchains & foundational architectures
Modern artificial intelligence is not about isolated Jupyter notebooks or cherry-picked accuracy metrics. It is about architecting self-healing tool-calling loops, strictly bounded vector retrieval, real-time edge computer vision, and reliable production software that stands up in enterprise environments.
Autonomous multi-step IT support agent integrating intent routing, grounded RAG policy retrieval, automated tier-1 incident resolution, human-in-the-loop escalation, and audit logging.
Smart roommate matching & co-living safety platform featuring Google Gemini AI rental contract lease analysis, 5-vector lifestyle compatibility radar charts, SHA-256 tamper-proof living pacts, and split bill UPI settlements.
Real-time YOLO11n deep learning computer vision pipeline engineered for helmet vs. no-helmet rider detection with edge inference optimization for automated traffic surveillance.
Context-aware browser AI companion (Manifest V3) that extracts on-page text, performs zero-shot reasoning, and interfaces with local Ollama instances and cloud OpenAI-compatible endpoints via a floating GUI.
End-to-end video intelligence engine converting YouTube transcripts into vector embeddings via FAISS, LangChain, and Ollama, paired with an asynchronous FastAPI backend for grounded semantic Q&A.
Local, grounded academic RAG system that parses syllabi and PDFs via PyMuPDF, performs semantic chunking and sentence-transformer vector search (all-MiniLM-L6-v2), and prevents hallucination via Ollama.
Predictive machine learning pipeline isolating critical agricultural yield drivers using XGBoost ensembles, rigorous feature selection, and SHAP explainability matrices for transparent agronomic decisions.
End-to-end environmental forecasting pipeline modeling multi-pollutant time-series patterns with Random Forest and XGBoost, packaged in Docker and deployed as a Streamlit application.
Data-driven vulnerability analysis of operating system kernels (2020–2025) across National Vulnerability Database (NVD) records, tracking severity distributions and CVE attack patterns.
Executive Tableau BI dashboard and interactive Story analyzing Amazon e-commerce performance, product category profitability, customer behavior, and regional delivery metrics.
Windows desktop utility built with PySide6 for dynamic switching between Claude Code configurations, custom LLM providers, base endpoints, and model parameters with secure credential management.
High-performance local area network file server featuring a Tkinter desktop dashboard and responsive web client supporting drag-and-drop uploads, media streaming, and live previews.
Custom desktop browser engineered with Python, PySide6, and QtWebEngine, incorporating persistent multi-tab routing, session management, and planned on-device quiz verification.
End-to-end KNIME data science workflow analyzing the Global Coffee Health dataset to model sleep quality, health issue severity, and circadian disruption from lifestyle attributes.
Machine learning benchmark comparing decision trees, random forests, and regression algorithms in KNIME for synthetic weather pattern classification and continuous temperature prediction.
Large-scale analysis of Indian national identity updates across states, identifying demographic shift ratios, biometric update lifecycles, and regional documentation trends.
Instant client-side search engine for university LMS curricula, enabling direct course ID queries, syllabus indexing, and seamless deep-linking to official academic modules.
Microcontroller automation system in C++ / Arduino managing school timetable alerts, automatic bell ringing schedules, LCD status telemetry, and servo-controlled door locks.
I'm an AI Engineer focused on bridging the gap between cutting-edge foundation models and production-grade software. Rather than training models that only live in isolated notebooks, I build autonomous agentic workflows, grounded RAG architectures, and real-time deep learning pipelines engineered for reliability, low latency, and zero hallucination.
My work spans the end-to-end AI engineering stack: from orchestrating local LLMs & tool-calling agents with LangChain and Ollama, designing dense vector embeddings (FAISS), and building high-throughput FastAPI backends, to fine-tuning YOLO11 vision systems and deploying explainable ML (SHAP) models with Docker.
AI Engineering Specializations
Autonomous agents, tool execution, multi-step reasoning & local Ollama
FAISS vector search, dense embeddings & zero-hallucination grounding
YOLO11 real-time detection, PyTorch neural networks & edge inference
XGBoost, SHAP interpretability, FastAPI backends & Docker containerization
Key design principles, reliability guardrails, and how I approach production artificial intelligence.
Through a multi-layer verification architecture: precision document ingestion (PyMuPDF), semantic chunking with overlapping context windows, dual-stage embedding search with FAISS, and strict prompt grounding constraints that restrict generation exclusively to verified retrieved context with source attribution.
Agents are built using modular planning loops with LangChain, typed Pydantic tool-calling schemas, dynamic local Ollama or cloud model routing, and automatic human-in-the-loop escalation protocols backed by comprehensive audit trails.
By pairing quantized lightweight architectures (like YOLO11n) with PyTorch tensor optimizations and asynchronous OpenCV video stream pipelines, achieving sustained real-time inference under 50ms per frame.
By deploying high-capacity gradient boosted ensembles (XGBoost) paired with SHAP (Shapley Additive Explanations) matrices, yielding rigorous mathematical explanations for individual feature contributions without sacrificing predictive accuracy.
