Shashank Pasumarthi

AI Engineer with close to 3 years of experience building production-grade GenAI applications for enterprise clients in financial services and telecom — working end-to-end from ingestion pipeline to deployment on AWS.Hands-on with RAG pipelines — document extraction using pdfplumber and unstructured, custom chunking strategies, embedding with text-embedding-3-small, and semantic retrieval with FAISS IndexFlatL2.Built conversational AI systems using LangChain and ConversationSummaryBufferMemory — managing multi-turn context with a rolling summary buffer across sessions keyed by session_id.Designed and exposed FastAPI async endpoints for AI query handling and streaming, integrated with SQL for structured data retrieval alongside vector search. Worked on production reliability — circuit breakers, retry with exponential backoff, fallback responses, Docker containerisation, AWS ECS rolling deploys, and health-check endpoints.Comfortable with GPT- 4-turbo for generation (128k context), prompt engineering for accuracy and hallucination reduction, and output validation to keep responses grounded in retrieved context.

AI Engineer – Generative AI Solutions at One IT Corp (2023-09 – 2026-03)

AI Engineer at One IT Corp (2024-12 – 2026-03)

Project Name: Enterprise Knowledge Assistant using Retrieval-Augmented Generation (RAG) Built an enterprise Generative AI knowledge assistant for Charles Schwab that lets internal teams query a large corpus of financial policies, compliance documents, and product manuals through natural language. The system retrieves relevant chunks from a FAISS index and uses GPT-4-turbo to generate grounded, citation-backed answers — with no general knowledge bleed, since the model is explicitly constrained to the retrieved context.

• Built the document ingestion pipeline from scratch — extracted text and tables from PDFs using pdfplumber and unstructured (which also detected embedded images; we sent those to GPT-4V to generate text descriptions before embedding), chunked content into 800-token segments with 150-token overlap, and stored embeddings in a FAISS IndexFlatL2 index (~12,000 vectors).
• Handled query retrieval using text-embedding-3-small to embed incoming questions, ran FAISS search with k=5 and an L2 distance threshold of 0.75 — chunks with L2 > 0.75 were discarded, and if all five fell above the threshold the system returned a fallback message rather than passing weak context to the model.
• Designed the prompt structure (role + instructions + retrieved context + question) and used GPT-4-turbo (temperature 0.0) for answer generation, followed by a post-generation validation step that checked whether the response was actually grounded in the retrieved chunks.
• Implemented ConversationSummaryBufferMemory to maintain multi-turn context — kept the last 8–10 turns verbatim plus a rolling LLM-generated summary of older turns, so the model had coherent conversation history without blowing up the context window.
• Set up storage and security — documents on AWS S3 with IAM role-based access, credentials managed through AWS Secrets Manager, no document content written to logs.
• Added production reliability features — circuit breaker (5 consecutive OpenAI failures triggered a 60-second pause), retry with exponential backoff on transient errors, and pre-written fallback responses that gave users a clear message instead of an error.
• Deployed using Docker containers on AWS ECS with a rolling deploy strategy and separate configs for dev, staging, and prod. FAISS index was versioned separately from the application to allow rollback without redeployment.
• Evaluated performance on a curated test set of ~80 queries — measured groundedness, answer relevance, and retrieval recall. In production, tracked two signals: fallback trigger rate and average FAISS L2 distance per query, both as drift indicators.
• Tech Stack: Python, LangChain, FastAPI, OpenAI GPT-4-turbo, text-embedding-3-small, FAISS, pdfplumber, unstructured, AWS S3, AWS ECS, IAM, Secrets Manager, Docker, SQL

AI Engineer at One IT Corp (2023-09 – 2024-11)

Project Name: Intelligent Customer Support Copilot using Large Language Models Built an AI copilot for Verizon customer support agents that suggests real-time responses during live calls. The system pulls relevant troubleshooting articles from a FAISS knowledge base, fetches the customer’s live account data from SQL, and passes both through a structured prompt to GPT-4-turbo — which generates a 1–3 sentence suggestion the agent can use or adapt on the spot.

• Set up a two-source retrieval architecture — FAISS for static content, and SQL for customer-specific data (order status, device, plan) fetched fresh at the start of each call.
• Designed a 6-section prompt that combined agent role, retrieved knowledge context, live customer context from SQL, conversation history, task instruction, and output format — tuned temperature to 0.1 for suggestions and 0.0 for call summaries.
• Implemented ConversationSummaryBufferMemory keyed to session_id — last 8–10 turns stored verbatim with a rolling LLM summary of older turns appended. Session memory was cleared at call end to prevent data from one call leaking into the next.
• Built FastAPI async streaming endpoints so suggestions appeared token-by-token to agents rather than after a full generation wait — important for live call latency.
• Enforced PII guardrails at the prompt level — the model was explicitly instructed not to include customer name, phone number, or account identifiers in any output. Responses were validated before returning.
• Handled the migration from ada-002 to text-embedding-3-small in January 2024 — re-embedded all ~8,000 knowledge article chunks and rebuilt the FAISS index with no service downtime, getting a 5× cost reduction and better retrieval quality as a result.
• Worked on production reliability — circuit breaker on OpenAI failures, FAISS index versioning for rollback, rate limiting on API endpoints, and logging that captured query patterns without recording any customer content.
• Collaborated with the business team on the k and threshold values using a controlled query test set — pushed back on a request to increase k from 3 to 10 with data showing it introduced noise without improving answer accuracy.
• Tech Stack: Python, LangChain, FastAPI, OpenAI GPT-4-turbo, text-embedding-3-small, FAISS, SQL, ConversationSummaryBufferMemory, Docker, AWS S3, AWS ECS

Master of Science (M.S.) in Computer Science – Northwest Missouri State University, USA (2021 – 2023)

Bachelor of Technology (B. Tech) in Electronics & Telecommunication Engineering – Sreyas Institute of Engineering and Technology (JNTUH), Hyderabad (2016 – 2020)