Ask Rino

Ask Rino is an AI companion allowing Amazonians within Devices and Services to operate on sensitive financial data and automate repeated workflows.

Ask Rino has been used by 400+ finance managers at Amazon across 41,000+ conversations, featured at internal FinTech conferences, actively pushed by the VP (Tobias Straub) into cross-org adoption, and referenced by the Amazon CFO (Brian Olsavsky). Now a VP-level Kingpin goal with a 100% onboarding target.

As the highest code contributor, I led Ask Rino from 0-1 and built core infrastructure that powers it:

• Architected one of the first remote MCP server infrastructures at Amazon, a centralized gateway connecting finance services directly into the AI platform. 2 MCP servers onboarded, 175+ tool invocations in 3 months.

• Extended the MCP protocol to a novel third transport layer to allow Ask Rino to securely operate on users' local systems (think OpenClaw).

• Built agentic text-to-SQL workflows that write, self-correct, validate, and execute queries on sensitive financial databases autonomously. Backed by an eval stack with 93%+ accuracy across 93 test queries.

• Secured Ask Rino's citation sources and links with CloudFront Signer allowing restricted access within Amazon. Supports 9 file types for user upload and chat.

• Solved a critical ThrottlingException error due to Claude's rate limits by proposing a novel model cycling method that scaled throughput to 600 requests per minute, bypassing the need for VP/SVP-level provisioned capacity approval.

Query API

• Built the Query API for blending calculations across monthly and lifetime device cost reports - 2,100+ calls in its first 4 months, now central to reporting and downstream services.

• Optimized latency by 98%, reducing deepest-level lineage calculation from 33.5s to 546ms.

Building controllable expertise selection interface using LACE model for reviewer paper assignment on Openreview with Sheshera Mysore and Andrew McCallum.

Self discover based style transfer for long form text with Gregory Druck.

Whistler: Financial Anomaly Detection

• Cut financial anomaly detection runtime 50% by moving from S3 to DynamoDB for Amazon Device, saving 18,000 file creations.

• Migrated AWS Step Functions infrastructure from Console to AWS CDK leading to scalability and code reusability.

Whistler Auditor: Centralized Cross-account Logging

• Accelerated failure tracking by 85% through the streamlined processing of 2k+ cross-account logs per month into Amazon Redshift.

Custom Product Hierarchy: Hierarchical Relationship Linking

• Designing and building a property-graph based custom hierarchy data model for 8k+ Amazon ASINs using Amazon Neptune, enabling insight generation.

Engineering co-lead for the Alexa Prize Taskbot challenge under Prof. Hamed Zamani at the Center for Intelligent Information Retrieval (CIIR) to build conversational agents for recipe and DIY tasks.

Alexa Prize Taskbot Challenge is a global competition to create next-generation conversational AI experiences that delight customers by addressing their changing needs as they complete complex tasks.