Tayyaba Tabassum

Senior AI Product Engineer specializing in Agentic Workflows and RAG Architectures.

Based in: Karachi, PakistanCurrently: Senior AI Engineer, Next Generation InnovationsExperience: 4 years

Tayyaba is an expert AI Product Engineer with a proven track record of architecting end-to-end document intelligence systems and graph-based RAG platforms using AWS Bedrock and LangChain. She excels at bridging the gap between product vision and technical execution, delivering high-performance conversational AI and automated workflows for global clients.

Experience

Senior AI Engineer

Next Generation InnovationsKarachi, PakistanworkJun 2025 – Present

Summary

Architected and built Arro, an AI-powered document intelligence system, end to end - owning system architecture, technical decisions, and delivery from design through production. Built Arro's Python pipeline to extract, classify, and structure large volumes of unstructured email and multi-format documents using AWS Bedrock LLMs (Claude, Llama), with named-entity extraction and confidence-scored validation. Designed guardrails for LLM outputs: confidence thresholds, schema validation, and data monitoring to keep AI-generated results accurate and safe to act on, reducing manual review effort. Architected natural-language search and conversational querying on AWS OpenSearch and Knowledge Bases, plus a RAG pipeline that turns unstructured content into decision-ready JSON for business stakeholders.

What I did

Developed an AI-first application where document parsing is the primary requirement.

PythonAWS BedrockClaudeLlamaAWS OpenSearchKnowledge BasesRAGJSONDocument parsing

Product Lead

Early-Stage StudiosRemote, UKworkDec 2024 – Jun 2025

Summary

Owned the entire product as Product Lead, translating product vision into technical plans, priorities, and releases. Architected the platform's foundational system design, setting the technical base the product was built on. Designed and built a graph-based RAG system combining knowledge-graph relationships with semantic retrieval to deliver more accurate, context-aware LLM answers.

What I did

Enabled responses to hierarchical and complex questions by implementing a graph structure.

Results

Completed the application in 6 months, split into two 3-month phases.
Achieved approximately 70% higher accuracy with graph-based RAG compared to vector-based RAG.

Graph-based RAGKnowledge-graphSemantic retrievalLLM

Software Engineer

WavetecKarachi, PakistanworkJun 2023 – May 2025

Took a multilingual, user-facing conversational AI product on WhatsApp from concept to launch, integrating OpenAI and Whisper APIs to automate voice-based banking workflows for end customers. Achieved sub-10-second response times through LangChain- and MCP-based agentic context management and token optimisation. Delivered an avatar-based helpdesk assistant powered by Google Gemini (Vertex AI) and Dialogflow, with real-time face detection for interactive, human-like support. Owned CI/CD pipelines, Git-based workflows, and Docker deployments, implementing automated release and monitoring processes to keep production AI services reliable.

WhatsApp APIOpenAIWhisper APILangChainMCPGoogle GeminiVertex AIDialogflowCI/CDGitDocker

Skills

Technical

Python

RAG

Anthropic Claude

Meta Llama

LangChain

OpenAI GPT

Google Gemini (Vertex AI)

Prompt Engineering

AWS (Lambda, Bedrock, S3, SageMaker, OpenSearch)

Vector databases (OpenSearch, Pinecone, FAISS)

Multi-Agent Workflows

Guardrails & Evaluation

LangGraph

Knowledge Graphs / Graph RAG

Hugging Face Transformers

Model Context Protocol (MCP)

AI Safety

TensorFlow

LoRA/QLoRA fine-tuning

PyTorch

React

Node.js

Document parsing

Team Management

Data Visualization

Architectural Decision Making

LoRA/QLoRA

Projects

Personalised Adaptive Learning System (EdTech)

Lead Developerpersonal

Designed and built a student-focused learning platform that estimates each learner's knowledge state using LSTM autoencoders and reinforcement learning, then adapts content recommendations through a continuous, data-driven feedback loop.

LSTM autoencodersReinforcement learningPython

Zapier Dataset Fine-tuning

personal

A practice project involving the fine-tuning of Qwen 2.5 on the Zapier dataset using data chunks to handle task complexity.

Qwen 2.5LoRA/QLoRA

Arro

personal

A document analysis tool currently used by internal clients for their daily workflows.

LLMGraph charts

Education

Bachelor of Science

FAST - National University of Computer and Emerging SciencesComputer Science

Recognition

Awards

Dean's List of Honor

FAST - National University of Computer and Emerging Sciences

Hire TayyabaReach out about a role, a contract or a conversation.For recruiters

Tayyaba's twin is AI, it can make mistakes.

More people on Kanvis

Browse more people Create your own Kanvis