Amogh Arora ShriAmogh

Hi, I'm Amogh Arora

B.Tech in Artificial Intelligence & Machine Learning @ GGSIPU (2022–2026)
AI/ML Engineer | Model Trainer | Python Developer | FastAPI Enthusiast

Languages: Python, C/C++, SQL
Backend Development: FastAPI, Django, WebSockets, REST APIs, GraphQL
Databases: SQLite, PostgreSQL, MongoDB
Infra & Dev Tools: Git, GitHub, VS Code, Docker, Apache Airflow
AI & Integrations: PyTorch, TensorFlow, Keras, Scikit-learn, Pydantic, LangChain, LangGraph, NLP, RAG, MCP, RAGAS, APO, Agentic Frameworks, Transformers, LLM APIs
Web Scraping: Puppeteer, Selenium, Headless Browsers, DOM Extraction, OAuth Handshake
Core Concepts: Data Structures & Algorithms, Machine Learning, Deep Learning, DBMS, OS, OOPS

AI Engineer Intern - Fractics (Feb 2026 - May 2026)

Engineered RAG & Ingestion Pipelines: Integrated MCP servers into an existing RAG system utilizing bidirectional WebSocket streaming, and built an end-to-end RSS ingestion pipeline to extract and enrich structured data on startup funding, acquisitions, and founders.
Developed Advanced Prompting & Evaluation Systems: Implemented Chain-of-Thought RAG with real-time step-by-step UI streaming, created an APO system for prompt refinement using golden datasets, and evaluated the entire pipeline using RAGAS with custom domain metrics.
Built Authenticated Scraping Architecture: Developed a headless Puppeteer scraping pipeline for DOM-level extraction that handles OAuth login handshakes to retrieve refresh tokens, enabling automated and authenticated access to gated data sources.

Agentic AI Engineer Intern - AlgorithmX (Nov 2025 - Dec 2025)

Architected a scalable FastAPI backend with clean service-repository design, async ORM, and optimized DB interactions, and built a bidirectional WebSocket system enabling session pooling, real-time state sync, and multi-user event pipelines.
Developed a JSON-to-graph compiler that converted JSON into structured DAG, enabling automated code generation.
Engineered a multi-agent workflow using Microsoft Autogen, creating custom routing logic and orchestrated agent collaboration for automated code generation and validation.

Software Developer Intern – Duco Consultancy (Jul 2025 – Aug 2025)

Designed and implemented MongoDB schema for scalable data handling.
Built & deployed Express.js backend services, fully integrated with frontend.
Streamlined workflows by loading large datasets into MongoDB for scalability.

Software Engineer Intern – Arya.ag (Jul 2024 – Oct 2024)

Automated multi-page web data extraction with Selenium & CSV export.
Achieved 96% accuracy with RandomForest on a multilingual dataset via custom preprocessing + Google Translate API.
Applied image processing (contour-based masking + classification) to identify crop types.

Fine-tuned Qwen2.5-1.5B via LoRA adapters on the Spider dataset using Kaggle GPUs, implementing completion-only loss masking and Cosine scheduling to reduce training loss to 0.02.
Implemented 4-bit QLoRA quantization (NF4 precision), successfully reducing model GPU memory footprint by 73% and optimizing query generation latency during local inference runs.
Developed a two-layer validation framework combining local SQLite execution row-matching with a Gemini LLM Judge to audit query semantics, achieving a 90% execution accuracy (a 10% absolute increase over the base model’s 80% baseline).

Architected a query-aware Hybrid RAG engine combining BM25 lexical retrieval, dense semantic search (Sentence-Transformers), and Cross-Encoder re-ranking (ms-marco-MiniLM-L-6-v2) in ChromaDB, integrating natural language intent and date extraction to eliminate temporal hallucinations.
Engineered a multi-agent state machine using LangGraph featuring an iterative self-correcting validation loop for Pydantic schema enforcement, a ComparativeAgent for multi-paper synthesis, and an AuditAgent that computes automated hallucination scores (0–100%) for factual grounding.
Developed an end-to-end full-stack AI platform using FastAPI and React (Vite), engineering asynchronous REST endpoints and an interactive UI for live PDF document ingestion, side-by-side literature review matrix rendering, and multi-turn grounded QA.

Designed a production-grade ELT pipeline on 10,000 real industrial sensor records (AI4I 2020), implementing a Kimball star schema with surrogate-key dimensions, staging-layer feature engineering (Kelvin-to-Celsius conversion, mechanical power derivation, overstrain metrics), and idempotent upserts for full reprocessability.
Orchestrated the 5-stage pipeline with an Apache Airflow DAG using XCom-based batch lineage, automated retries, and 8 data quality gates (null checks, range validation, referential integrity) with a circuit-breaker pattern that halts downstream loads on critical failures.
Built a live FastAPI observability dashboard serving real-time warehouse metrics, an interactive SQL sandbox with preloaded queries and analytics views computing rolling Z-score anomaly detection across sensor telemetry.

Engineered an end-to-end RAG pipeline with document ingestion, chunking, embedding, vector similarity search, and cross-encoder re-ranking, retrieving top-K high-relevance chunks and integrating an LLM to re-generate grounded, context-aware answers based on re-ranker scores for improved accuracy and reduced hallucinations.
Designed a scalable Django backend with authentication-protected routes, session-based user isolation, PostgreSQL persistence, and modular app architecture, ensuring clean separation between auth, ingestion, and query workflows.
Implemented asynchronous processing and scalability patterns using Celery for background indexing, job-based progress tracking, and non-blocking request handling, making the system resilient to heavy document loads and multi-user concurrency.

⭐️ From ShriAmogh