Prathiba Lakshmi

Conversational AI Systems, LLM Security & Semantic Search
• Designed and deployed enterprise conversational AI systems using FastAPI, Python, MySQL, and LLMs, supporting
large-scale issue reporting and workflow automation.
• Implemented LLM security guardrails to mitigate prompt injection vulnerabilities, preventing unauthorized access
to sensitive employee and system data.
• Built multi-format document intelligence pipelines supporting PDF, Word, Excel, Google Sheets, and embedded
images using Gemini Pro Preview and Gemini 2.5 Flash.
• Integrated vector search (Pinecone) for semantic retrieval, improving answer relevance and contextual grounding.
• Coordinated Dev Prod deployments using GitHub Actions, Jenkins, AWS ECS, and validated APIs using Postman
and production logs.

Semantic and Keyword Search Optimization for NLP Systems
• Built and enhanced slang & emoji interpretation modules for conversational AI platforms using curated databases
and Meta-based search workflows.
• Implemented hybrid semantic + keyword search for slang and keyword-only retrieval for emojis, reducing incorrect
fallbacks and improving moderation accuracy.
• Designed payload orchestration, reranking logic, and conversation-history-aware query routing.
• Supported vector-based RAG workflows using embeddings and Pinecone indexes for contextual response
generation.
• Ensured production readiness through API testing, controlled merges, and deployment coordination.

Docker Vulnerability Remediation
• Investigated and fixed OpenSSL and libssl vulnerabilities across Python Docker images.
• Implemented multi-stage Docker builds using secure Debian variants (Bookworm, Trixie).
• Improved container security posture without breaking ML dependencies.

• Payroll Intelligence
  • Designed and implemented an LLM-powered payroll extraction pipeline to convert unstructured documents
  (payslips, contracts) into structured data, improving accuracy and reducing manual review.
  • Built a schema-driven entity extraction framework using prompt engineering and strict validation to ensure factual,
  non-hallucinated payroll data.
  • Integrated extraction with database persistence, request tracking, and modular ingestion utilities, improving
  workflow reliability, scalability, and audit readiness.
  Invoice OCR Automation
  • Designed an OCR pipeline using Tesseract, image preprocessing, regex, and LLM-based extraction.
  • Supported multi-language, multi-format invoices with ERP-ready JSON outputs.
  • Reduced manual invoice processing effort through end-to-end automation.

Generative AI RAG Chatbot
• Developed Generative AI RAG chatbots using LangChain, BERT, Pinecone, and GPT-based models for
document-driven question answering.
• Implemented document chunking, cleaning, and embedding pipelines for unstructured PDFs and research
documents, improving retrieval quality.
• Engineered robust user query classification model achieving 98% accuracy with BERT.Built query classification and
retrieval evaluation workflows, improving response accuracy and reducing hallucinations in production chatbots.
• Integrated pre-trained Large Language Models (LLM) such as Jina V2 and GPT-3.5 Turbo for prompt engineering,
answer generation, and embedding generation stored in Pinecone VectorDB index

Prompt Injection Mitigation in Production LLM System
• Eliminated prompt-based instruction bypass attempts by introducing layered input and output guardrails in a
production LLM system.
• Reduced security exposure and compliance risk by preventing unauthorized access to employee and internal data
through architectural controls.
• Improved predictability and safety of LLM responses by enforcing instruction hierarchy and prompt isolation
mechanisms.

• Developed and fine-tuned predictive models, including Logistic Regression and Neural Network, for Breast Cancer Prediction, achieving a 92% increase in accuracy and enhancing predictive capabilities for early detection
• Attained accuracy rate of 92% in predictive performance that aids in early breast cancer detection or risk assessment
• Reduced overfitting in ML models by applying regularization techniques and systematically evaluating model complexity.

• Scraped and structured data from 500+ Amazon product pages using BeautifulSoup and Selenium, capturing over 10,000+ customer reviews and product attributes. Stored data in PostgreSQL for structured querying, enabling 95% faster access to key insights for analysis.
• Applied SMOTE to balance class labels across 5,000+ records, engineered features using scikit-learn, and developed a custom FCNN in PyTorch achieving 87% prediction accuracy in identifying product relevance for specific user needs improving recommendation precision and guiding stock strategy through POC implementation.

Bank Loan Analysis

• Developed an interactive Power BI dashboard by extracting and transforming data using SQL queries from a bank loan dataset of 10,000+ records. Visualized key metrics such as loan approval rate, default rate, and credit score distribution, improving risk analysis and enabling 15% better decision accuracy.
• Utilized SQL for complex data filtering and joins, and performed detailed loan trend analysis to uncover insights into demographic patterns, repayment behavior, and high-risk segments, contributing to an 8% reduction in loan defaults through data-driven policy recommendations.

• Tutored over 300 students in various IT subjects and Soft Skills and Communication training, achieving success rate of over 90%
• Designed & delivered training modules across diverse skills across various age groups with average satisfaction rating of 93%
• Provided one-on-one mentoring to help individuals reach their full potential within the organization.

• Achieved year-on-year scale up of 40% growth and student success rate
• Transformed Matriculation curriculum with data analytics, elevating student engagement and retention by up to 25%