Jacob
Kuriakose.
Structured inquiry.
Scalable impact.
I build AI that thinks, plans, and acts. Not just responds. My focus is Agentic AI and RAG systems powered by memory, tool-use, and multi-step reasoning.
Most of that work runs on AWS: fine-tuned small language models, multi-agent workflows, and retrieval pipelines that hold up once real people are using them. I have taken a model from 56% to 92% accuracy and kept a chatbot running for 2,000 students.
I am a Platform Engineer at Wipro in Plano, TX. Before that came an MS in Data Science at Arizona State University, a research assistantship, and a startup I co-founded.
What I build with.
LLM Fine-Tuning
Time Series Forecasting
RAG
Multi-Agent Systems
NLP
AWS
Machine Learning & AI
13Programming
6Frameworks & Libraries
11Cloud & Tools
11Experience & Education
Platform Engineer
Wipro · Plano, TX
- Joined full-time as a Platform Engineer after converting from the summer AI Engineer internship
AI Engineer Intern
Wipro · Plano, TX
- Fine-tuned a Granite 4.1 8B SLM for AT&T change risk management, taking accuracy from 56.5% to 92.3% and macro F1 from 0.50 to 0.92, using Claude as an LLM judge for explanation correctness and action relevance
- Built an end-to-end synthetic data generation and training pipeline from AT&T-provided schemas and sample records, leveraging Kimi K2.5 to generate 50K domain-specific change tickets for enterprise change management workflows
- Developed a change & configuration management agent connected to ServiceNow ticket data and CMDB network topology, enabling the fine-tuned SLM to generate risk rules, blast radius analysis, recommended actions, and rollback plans
- Deployed the agentic SLM workflow on AT&T's AWS dev instance and integrated it with Wipro's WINGS platform for testing and validation
AI Backend Developer
Arizona State University · Tempe, AZ
- Led migration of production-grade AI chatbot (Waterbot) to CISA AWS environment
- Provisioned AWS CDK stack (ECS Fargate, ECR, ALB, CloudFront, S3, RDS PostgreSQL, DynamoDB, Lambda), cutting manual setup ~40% and deployment downtime ~30%
- Built multi-stage RAG pipeline using AWS Bedrock Knowledge Base + GPT-4.1: safety moderation → intent detection → language detection → semantic retrieval → source citation linking
- Replaced a self-hosted Chroma vector DB with AWS Bedrock Knowledge Base as the retrieval layer; re-indexed embeddings and wired managed retrieval into ECS tasks for better scalability
- Engineered FastAPI backend with 13 REST & WebSocket endpoints; implemented per-session memory management and pluggable ModelAdapter pattern (OpenAI / Bedrock) for swappable LLM backends
- Developed real-time voice transcription via WebSocket with S3 storage and presigned URL downloads; added English/Spanish auto-detection using langdetect
- Implemented secure RDS integration via Secrets Manager & IAM roles; configured EC2 IAM role-based auth, eliminating static credentials
- Built CI/CD with GitHub Actions and pushed Docker images to ECR; switched to ECS CLI redeploys reducing release cycle ~85%
- Restored chatbot access for 2,000+ users by removing unintended CloudFront auth
Software Quality Engineer Intern
Red Hat · Remote
- Automated & executed software test suites using GoLang
- Reduced test execution time by 30% through optimization
- Designed detailed test plans in Polarion
- Used Docker for scalable deployment environments
NLP Intern
BraynixAI · Remote
- Enhanced legal document classification accuracy by 10% using Transformer models
- Built FastAPI ML services for document processing at scale
Master of Science
Arizona State University · Tempe, AZ
Field: Data Science
GPA: 4.00/4.00
- AI Backend Developer on Waterbot, a production RAG platform (azwaterbot.org)
- Courses: Software Security, Data Mining, Statistical Machine Learning, Database Management Systems
Bachelor of Technology
Dr. Akhilesh Das Gupta Institute of Technology and Management · Delhi, India
Field: Information Technology
GPA: 9.0/10
- Published 'Monument Tracker' Deep Learning research paper
Measured, not claimed.
Granite 4.1 8B on AT&T change risk
Fine-tuned a small language model on synthetic change tickets, judged by Claude for explanation correctness and action relevance.
Accuracy
% correct
▲ +35.8 points
| base | 56.5% |
|---|---|
| tuned | 92.3% |
Macro F1
0 – 1
▲ +0.42
| base | 0.50 |
|---|---|
| tuned | 0.92 |
Shipped systems
Production outcomes from the Waterbot platform, the synthetic data pipeline, and retail demand forecasting.
~85%
Faster releases
ECS CLI redeploys over full-stack deploys
2,000+
Users served
Waterbot RAG chatbot at azwaterbot.org
50K
Tickets generated
Synthetic change-management training data
38.9%
RMSE improvement
Exponential Smoothing over LSTM
Recent case studies.
Navia
Co-founded AI companion for neurodivergent adults, built around multi-agent orchestration, Pinecone memory and E2EE peer messaging
Explore case
ReAct-Based Medical Chatbot
Agentic medical AI chatbot with ReAct reasoning, dual search, and safety classification
Explore case
Waterbot
AI-powered educational chatbot for water literacy with multilingual & voice support
Explore case
Walmart Sales Forecasting
Retail demand forecasting using ML & Deep Learning
Explore caseInteractive
Beat the forecast.
Holt-Winters ES
the model from and try to beat it
Drag across the shaded area to draw where you think demand goes next. Then see how your line scores against the model.
0/10 days drawn
Synthetic weekly-seasonal series, not project data. The model is real: Holt-Winters exponential smoothing, fitted live on the visible history.
Professional Certifications
Generative AI with Large Language Models
Verify Certificate
Machine Learning
Verify Certificate
30 Days of Google Cloud
Verify CertificateOpen to full-time roles, freelance projects, and research collaborations. I'll get back to you within 24 hours.
Location
Plano, TX