LLM & RAG Engineer (Generative AI)

UTKARSH SINGH

AI/ML Engineer LLM, Generative AI & RAG Systems Immediate Joiner

+91-956******* *******************@*****.*** linkedin.com/in/utkarsh-singh-32b29b216 github.com/utkarshsingh2122

PROFESSIONAL SUMMARY

AI/ML Engineer with 2+ years of experience building LLM and generative AI applications for production. Designed, built and owned a self-hosted LLM/RAG platform (Ollama, Qdrant, FastAPI, React) as the sole engineer at a 40-person accounting firm, including an OCR and LLM-based invoice extraction pipeline and a source-grounded SOP assistant. Also built ML-based phishing detection, AI agent SFT data annotation and a human-in-the-loop AI supervisor, and automated Azure and AWS resources with Terraform and Jenkins. Microsoft Azure certified (AZ-900, AZ-104, AZ-204, AZ-140).

TECHNICAL SKILLS

AI/ML and LLM Engineering: RAG pipelines, LangGraph, LangChain, Generative AI, LLM application development, LLM integration, embeddings (Jina AI), vector retrieval (Qdrant), chunking and text preprocessing pipelines, Redis chat history, self-hosted LLM inference (Ollama, vLLM), LLM fine-tuning (LoRA, QLoRA), Google Gemini API, structured LLM output, OCR (Tesseract), pdfplumber, document layout routing, prompt design Machine Learning and Data Analysis: Machine learning, deep learning, neural networks, PyTorch, TensorFlow, NLP, time series analysis, predictive modeling, feature engineering, ML-based phishing and threat detection, Python data analysis, NumPy, Pandas, Matplotlib, statistical analysis, data structures and algorithms Programming Languages: Python, Go (Golang), JavaScript, TypeScript, Java, SQL, Bash, C/C++, HTML5/CSS3 Backend Development: FastAPI, Node.js (Express), REST API design, OpenAPI/Swagger, Pydantic, SQLAlchemy, JWT authentication, RBAC, webhooks, event-driven patterns, LiveKit integration (STT/TTS flow) Frontend Development: React (Vite), TypeScript, SCSS, responsive and accessible UI, reusable component design Cloud and Infrastructure: AWS resource automation (Amazon EC2, Amazon S3, AWS IAM, Amazon VPC), Microsoft Azure (Resource Manager, Storage, Service Bus, Integration Services, Virtual Desktop, Azure SDK for Java/Python), Azure resource provisioning and automation, Linux (Ubuntu) administration, NVIDIA GPU server and on-prem AI hosting, Nginx reverse proxy, network isolation DevOps and Automation: Infrastructure as Code (Terraform) for AWS and Azure, Jenkins (Groovy pipelines), GitHub Actions, CI/CD, deployment strategies, versioning (SemVer), Docker, Docker Compose, Prometheus, Grafana, Bash scripting

Databases and Messaging: PostgreSQL, MongoDB, Redis, SQLite, Qdrant, Azure Service Bus (queues/topics) Tools and Platforms: Git, Linux, Swagger UI, cURL, Streamlit, VS Code, IntelliJ, PyCharm PROFESSIONAL EXPERIENCE

AI Platform Engineer Ratiobox India Private Limited, Delhi, India June 2026 – Sept 2026

• Architected and built the company's self-hosted AI/RAG platform as the only engineer at a 40-person accounting and bookkeeping firm, owning design, development, testing, deployment and production support.

• Delivered the full stack (FastAPI, React, PostgreSQL, Qdrant, Ollama, Docker, Nginx) covering document RAG, invoice extraction and authentication on an on-prem Ubuntu GPU server.

• Implemented JWT authentication, 4-role RBAC (Admin, Accountant, Staff, Viewer), audit logging and Docker network isolation to protect confidential client data.

• Tuned retrieval quality, container resource allocation and monitoring thresholds after launch, using real production usage and Prometheus/Grafana metrics.

Software Engineer Capgemini Technology Services India Limited, Bangalore, India Oct 2024 – May 2026

• Automated Azure resource provisioning with Terraform and built Jenkins (Groovy) CI/CD pipelines, improving deployment speed and consistency across environments.

• Developed cloud solutions on Azure Integration Services and troubleshot and optimized CI/CD workflows with cross-functional DevOps and development teams.

• Built an AI Human-in-the-Loop Supervisor (FastAPI, SQLite, LiveKit) where an AI receptionist answers known queries and escalates unknown ones to a human, with supervisor answers feeding back into the knowledge base.

• Built an end-to-end phishing detection application (Python, ML, Streamlit, Docker) that classifies URLs and text messages in a live prediction interface.

• Built reusable, responsive and accessible React/TypeScript/SCSS front-end components for client-facing modules.

• Annotated and evaluated multi-step AI agent trajectories (CLI, browser, API) for Supervised Fine-Tuning datasets on the OpenClaw x Meta project.

PROJECTS

Ratiobox Private AI Platform (Ratiobox) FastAPI, React, PostgreSQL, Qdrant, Ollama, Docker, Nginx, Prometheus, Grafana

• Built a self-hosted RAG platform for accounting and internal knowledge workflows used by a 40-person non-technical team: upload, text extraction, chunking, embeddings, Qdrant indexing, semantic retrieval and source attribution.

• Added metadata-driven filtered retrieval (client code, document type, access group) with client-level authorization to keep client knowledge separated.

• Developed an SOP AI Assistant on local LLM inference (Ollama) that returns context-grounded answers with source references, plus a React admin portal for documents, AI queries, invoice review, audit logs and system health.

• Deployed behind Nginx with Docker network segmentation and firewall controls so PostgreSQL, Qdrant, FastAPI, Prometheus and Grafana are not reachable from the LAN or public network.

• Integrated Prometheus and Grafana to track request counts, latency, ingestion failures and service health, with query logging and health endpoints for traceability. Invoice Field Extraction Pipeline (Ratiobox) Python, pdfplumber, Tesseract OCR, LLM structured extraction, FastAPI

• Built a pipeline extracting invoice number, dates, parties, line items, tax treatment, totals and payment terms for AP/AR matching, VAT compliance and reconciliation.

• Routed each document to text-layer parsing, table extraction or LLM extraction by layout, with Tesseract OCR fallback for scanned invoices; enforced a strict JSON schema on LLM output.

• Added arithmetic validation (line items vs. subtotal, subtotal plus tax vs. total) and normalization (ISO 8601 dates, ISO 4217 currency), handling reverse-charge VAT, credit notes, partial payments and multi-page invoices.

• Recorded extraction provenance (source file, method, model version) for auditability. Azure & AWS Infrastructure Automation: One-Click Deployment (Capgemini) Terraform, Jenkins, Groovy, Azure Integration Services, Amazon EC2, S3, IAM, VPC, React

• Developed Terraform scripts and translated them into Groovy for Jenkins CI/CD, enabling one-click provisioning across multiple Azure services.

• Applied the same pipeline-driven approach to AWS: automated EC2 virtual machine provisioning and configuration, S3 bucket creation and management, IAM users, roles and permissions, and VPC networking.

• Replaced manual console setup with repeatable, pipeline-driven deployments on both clouds, improving deployment speed and consistency across environments and reducing manual effort. Human-in-the-Loop AI Supervisor (Capgemini) FastAPI, SQLite, SQLAlchemy, Python, JavaScript, LiveKit

• Designed the data model (KnowledgeBase, HelpRequest, Notification) with a pending-to-resolved/unresolved lifecycle, input validation, idempotent knowledge-base upserts and a documented Swagger API; added a LiveKit stub for future voice.

Phishing Detection from URLs and Text Messages (Capgemini) Python, Machine Learning, Streamlit, Docker, Git

• Built a feature-extraction pipeline (URL structure, urgency language) and trained a classifier, then integrated it into a live Streamlit app, containerized with Docker, returning Phishing/Legitimate results in real time. AI Agent SFT Data Annotation: OpenClaw x Meta (Capgemini) OpenClaw, AI Agents, CLI, Browser Automation, APIs, LLM Evaluation

• Simulated real-world workflows using an autonomous AI agent (OpenClaw) and annotated multi-step trajectories

(CLI, browser automation, APIs) with step-level Correct/Incorrect/Redundant evaluation to improve LLM reliability in real-world task execution.

• Evaluated trajectories for correctness, efficiency and completeness to produce Supervised Fine-Tuning datasets

(Jan 2026 – Apr 2026).

EDUCATION

B.Tech in Computer Science GLA University, Mathura, Uttar Pradesh Aug 2020 – May 2024 CERTIFICATIONS

Microsoft Azure Fundamentals (AZ-900) • Azure Administrator Associate (AZ-104) • Azure Developer Associate

(AZ-204) • Azure Virtual Desktop (AZ-140) • HackerRank Java (Basic) and Problem Solving (Basic)

Contact this candidate