Siddhartha – AI agent development, LLM, RAG

Developed a full-stack AI assistant for field service operations, enabling technicians to search technical manuals, equipment documentation, and troubleshooting guides using voice or text. The solution combined an advanced retrieval pipeline with a cross-platform web and mobile application, delivering grounded, citation-backed answers enriched with relevant images, diagrams, and page references to support faster issue resolution in the field.

• Designed and implemented a triple-hybrid retrieval architecture combining vector search, BM25, and structure-aware document navigation, fused with Reciprocal Rank Fusion to improve retrieval quality, deduplicate results, and increase answer relevance;
• Built a LangGraph ReAct agent with auditable retrieval and source-selection workflows, including configurable execution limits and graceful recovery from recursion failures;
• Developed a multimodal document ranking pipeline using a custom five-criterion grading rubric, combining LLM-based quality assessment with vector similarity scoring to improve retrieval accuracy across both text and image content;
• Engineered an end-to-end document ingestion pipeline featuring LLM-assisted semantic chunking, automatic question generation, multimodal document analysis, and unified text/image embeddings for cross-modal retrieval;
• Designed and implemented a low-latency, voice-first architecture with client-side wake-word detection, streaming speech-to-text, and persistent audio sessions to deliver a seamless conversational experience;
• Implemented a secure token-based authentication flow for voice services using short-lived credentials, ensuring API keys never reached the client while minimizing unnecessary token generation;
• Architected a scalable Turborepo monorepo with shared business logic, state management, API clients, and design tokens across web and mobile applications, while intentionally separating UI implementations to align with React and React Native best practices;
• Designed site-level data isolation directly within the retrieval layer by enforcing metadata-based filtering across search systems, ensuring technicians could access only documents relevant to their assigned locations.

An AI-powered document intelligence platform for processing complex commercial property records. The solution classified legal documents, extracted structured information from OCR-processed files (including handwritten annotations), consolidated multiple document versions into a single authoritative property record, and automated financial calculations through AI-generated executable code. A multi-stage verification pipeline with independent AI validation ensured the accuracy of extracted data, generated formulas, and calculated results.

• Architected a resumable, multi-stage AI processing pipeline covering document classification, agentic extraction, document consolidation, formula generation, and rent calculations, allowing production workflows to restart safely from any stage;
• Built a LangGraph-based extraction system with cumulative retry logic that preserved and merged valid results across attempts, ensuring retries progressively improved extraction quality without losing previously captured data;
• Designed a document consolidation engine that merged original agreements, amendments, and supplementary documents into a single authoritative record by applying amendment precedence and chronological ordering rules;
• Implemented a multi-stage AI verification pipeline with independent validation of extracted fields, generated business logic, and financial calculations, automatically triggering re-extraction or regeneration whenever inconsistencies were detected;
• Engineered an agentic code generation workflow that translated contractual clauses into executable Python business logic, validated through sandboxed execution, schema validation, and multiple verification tools;
• Developed robust OCR processing capable of correctly interpreting handwritten annotations, corrections, and margin notes by applying annotation-aware extraction rules and prioritizing annotated document sections during retrieval;
• Designed strongly typed Pydantic schemas for every stage of the pipeline, enabling reliable validation, structured outputs, and sequential extraction with cross-referencing across related legal and financial entities.

An agentic RAG platform that enabled operational teams to ask natural language questions about internal policies and receive grounded, citation-backed answers from enterprise documentation. The solution combined a LangGraph-based orchestration engine, automated document synchronization, semantic indexing, and LLM-assisted retrieval enrichment, with built-in evaluation and model promotion workflows to continuously measure and improve response quality.

• Architected a LangGraph-based orchestration engine with strongly typed intent routing, ensuring business logic was controlled programmatically while using LLMs exclusively for classification tasks;
• Built an incremental document synchronization pipeline with snapshot-based state management, supporting reliable create, update, relocate, and delete operations while maintaining consistency during partial synchronization failures;
• Implemented an LLM-driven semantic ingestion pipeline with intelligent chunking and automatically generated retrieval questions, significantly improving semantic search quality beyond simple keyword matching;
• Designed a parallel document evaluation workflow with distributed tracing, enabling scalable relevance grading while preserving full observability across the retrieval pipeline;
• Developed a challenger - champion evaluation and deployment framework using MLflow GenAI metrics, automated model comparison, artifact generation, and controlled production promotion based on quality thresholds;
• Implemented a multi-layer guardrail architecture combining graph-level constraints, structured output validation, source-grounded generation, citation verification, and automated safety evaluation before model deployment;
• Engineered secure infrastructure with isolated service identities and environment-specific credentials, enforcing least-privilege access and eliminating shared authentication across platform components.

An enterprise AI automation platform for processing customer emails and support requests. The solution automatically classified incoming emails, extracted structured information from attachments, matched customer records across business systems, and created support cases with minimal manual intervention. It also included a continuous feedback loop that monitored human corrections and used them to improve model performance over time through automated production evaluation.

• Architected a set of independent, production-grade data pipelines for email classification, entity extraction, model training, quality monitoring, and lifecycle management, with automated scheduling, health monitoring, and alerting;
• Designed a multi-stage classification engine combining deterministic business rules, LLM-based classification, and traditional machine learning models, optimizing for both accuracy and cost through prioritized execution;
• Built a resilient entity extraction pipeline with multiple fallback strategies, progressively combining PDF parsing, multimodal AI, and document intelligence services to maximize extraction accuracy while minimizing unnecessary AI usage;
• Implemented hybrid customer record matching using a combination of string similarity and embedding-based semantic search to improve account resolution despite inconsistent or incomplete data;
• Developed a closed-loop learning system that automatically captured human corrections and incorporated them into future model training, enabling continuous improvement without manual dataset labeling;
• Built a custom data quality framework with automated validation across every pipeline stage, proactively detecting silent regressions and model behavior changes before impacting production;
• Engineered a scalable integration framework for enterprise CRM operations using reusable batch-processing patterns, significantly simplifying the addition of new automated workflows.

Developed data science solutions focused on time series forecasting, anomaly detection, large-scale web scraping, and automated data extraction. Built data processing pipelines and predictive models to analyze complex datasets, detect unusual patterns, and transform unstructured web data into actionable business insights.

• Leveraged Large Language Models for product information extraction and recommendation systems from complex product page HTML;
• Incorporated fine-tuning of open-source LLMs and Retrieval Augmented Generation to improve accuracy and adaptability over rule-based methods;
• Built and deployed a bidirectional LSTM model to detect anomalies in receipt and POS data, reducing false positives;
• Created an unsupervised clustering-based text classification algorithm to match incoming data with an existing 1 billion record oracle dictionary;
• Trained and deployed machine learning models in the cloud using Docker, Databricks, and Azure Machine Learning;
• Created and presented interactive Power BI dashboards and reports to stakeholders to support data-driven decision-making.