Manish Kumar Maurya

Project: Multi use AI Assistence

Designed and implemented a multi agent Agentic AI Platform that serves multiple enterprise use cases (FAQs, order tracking, recommendations, refunds) through a single conversational interface.

The system uses embedding-based intent classification, agent orchestration, Retrieval-Augmented Generation (RAG), and LLMs, and is deployed on Kubernetes with authentication, scalability, and observability built in.

Tools & Technologies Used:

Languages & Frameworks

• Python, FastAPI

AI / LLM

• OpenAI GPT models(gtp-4o-mini)
• Embedding-based intent classification
• Retrieval-Augmented Generation (RAG)

Data & Caching

• Redis (response & intent caching)
• Vector Database

Security

• JWT Authentication
• PII masking
• Rate limiting

DevOps & Cloud-Native

• Docker
• Kubernetes
• Kubernetes Secrets
• CI/CD

Observability

• Langfuse

Project: Intelligent Invoice Processing (IDP)

• Design and implement an AI-driven Intelligent Document Processing (IDP) solution to automatically extract structured json invoice data from unstructured vendor-specific PDF invoices document using Large Language Models (LLMs), ensuring high accuracy, schema compliance, and minimal manual intervention for finance workflows.

• Architected a modular invoice ingestion pipeline to process multi-page, vendor-specific PDF invoices with robust text extraction and normalization
• Designed a schema-first extraction approach for invoice fields such as Invoice Number, Vendor Name, Invoice Date, Line Items, Tax, and Total Amount
• Engineered constraint-based prompts to enforce strict JSON output, null handling, numeric normalization, and ISO date formats for financial accuracy
• Integrated OpenAI GPT models as the core extraction engine, optimizing for cost, latency, and token usage through prompt control and content truncation
• Implemented defensive JSON parsing and fallback mechanisms to ensure pipeline reliability for financial systems
• Designed APIs to expose structured invoice data for downstream systems.

• Reduced manual invoice data entry effort by 70%
• Improved invoice extraction accuracy by 40% compared to OCR + regex solutions
• Accelerated invoice processing turnaround time by 50%
• Enabled scalable automation for high-volume finance and accounting operations

Tools & Technologies

Python, OpenAI GPT(gpt-4.1-mini), Prompt Engineering,, Pandas, dotenv, Langfuse (Observability)

Project: Context aware LLM optimization plateform

Designed and implemented an enterprise-grade Retrieval-Augmented Generation (RAG) platform to contextualize Large Language Model (LLM) for improved accuracy, cost, and latency. The solution focuses on intelligent context management rather than model selection, enabling scalable, cost-efficient, and reliable LLM adoption for enterprise knowledge systems.

Roles and Responsibilities:

• Architected a multi-stage RAG pipeline incorporating semantic chunking, query rewriting, reranking, and context compression to improve response accuracy and efficiency.
• Implemented semantic chunking based on document structure (policy-level sections) to preserve business meaning and improve retrieval recall.
• Built a high-recall vector retrieval layer using embeddings and Chroma Vector DB to ensure relevant knowledge consistently
• Developed an LLM-based query rewriting mechanism to normalize user queries into enterprise policy language and handle ambiguity (e.g., SLA response vs resolution).
• Integrated a cross-encoder reranking layer to improve precision by selecting the most semantically relevant context before answer generation.
• Implemented context compression to extract only query-relevant information from retrieved documents, significantly reducing prompt size while maintaining correctness.
• Enforced strict answer grounding and guardrails to prevent hallucination
• Designed and implemented a metrics and observability layer to track token usage, latency, and cost per request.
• Conducted comparative analysis (with vs without context compression), demonstrating measurable improvements in system efficiency in term of cost and latency.
• Languages: Python
• LLMs: OpenAI GPT 4o model models
• Embeddings: OpenAI Embeddings
• Vector Database: Chroma
• Reranking: Sentence-Transformers Cross-Encoder
• Metrics: Token accounting, latency measurement, cost analysis

Project: FALCM (Edward Jones)

• Modernized a Fortune 500 financial services platform for better advisor/branch experience.
• Designed and implemented Java 17 + Spring Boot microservices on AWS.
• Built REST APIs, CI/CD pipelines,
• Contributed to architecture, development, and integration of large-scale systems.
• Managed team of 12 members .
• Mentored junior developers for improved coding skills and understanding of best practices, leading to increased overall code quality within the team.
• Collaborated with cross-functional teams to design and implement innovative technical solutions, fostering streamlined communication channels and efficient problem resolution.
• Developed technical solutions.

Project: M&H Digital Transformation (AXA)

• Led end-to-end delivery of integration systems for insurance digital transformation.
• Designed and delivered 8+ APIs interacting with core systems.
• Set up build & release pipelines with DevOps teams.

Project: CXP (Kroger)

• Led design and delivery of customer experience platform.
• Coordinated with business teams for loyalty, pricing, and promotions modernization.
• Earlier Projects (IBM, Etisalat, Skidata, Infinite, Capgemini, Torry Harris)
•