Atul Kumar — Software Engineer · AI Agents, LangGraph, AWS, Apache Spark, Distributed Systems

Engineering Associate

TINA — Travel Insights & Navigation Agent

• Deflected 80% of travel & expense support queries from human agents by building TINA, an internal AI agent that resolves policy questions (missed flights, reimbursements, approvals) end-to-end.
• Cut average policy-answer latency to under 20 seconds across 40 policy documents by building a RAG pipeline that ingests document embeddings into AWS SageMaker and orchestrates multi-step retrieval with a LangGraph agent.

Travel & Expense Platform

• Enabled streaming export of 5M+ live transaction, report, and ledger records per request by engineering a server-side API that sustains ~50K records/sec within a 10-minute window, eliminating the timeouts and OOM failures from the prior batch approach.
• Eliminated daily manual reconciliation by the operations team across 80 global markets by building an auto/manual voucher creation flow for vendor (Amex) payments with auto-generation of ledgers, statements, and voucher reconciliation.
• Met Goldman Sachs' Tech Raise Bar quality standard for the greenfield Travel & Expense platform by authoring an end-to-end integration test suite covering AWS Step Functions, ECS tasks and services, S3, and Aurora.

AI-Driven Invoice Lifecycle Management (POC)

• Eliminated manual data entry on every invoice submission by building an OCR layer that pre-fills 100% of form fields from invoice attachments for downstream human verification.
• Built an intelligent reviewer-assignment engine that routes invoices to reviewers based on expertise, calendar availability, and criticality, with cycle-time metrics fed back into the model for continuous improvement.
• Prevented duplicate payments across $2M+ in average daily invoice volume by adding a duplicate-detection layer over historical submissions.

Slate — Inter-Affiliate Outsourcing Agreements

• Drove data-backed feature prioritization for the next platform build by integrating GS Analytics into the Slate frontend, capturing feature-level telemetry from 800 users across 10 workflows in the agreement-creation pipeline.

High-Throughput Spark Pipeline on AWS Glue (FFIEC 009)

• Engineered a high-performance Apache Spark 3 pipeline on AWS Glue that ingests and processes data from 12 upstream sources under strict latency and correctness budgets.
• Reduced new-pipeline build time across the team by ~75% by designing shared libraries that standardize data processing and validation.

Engineering Analyst

Large-Scale Data Pipeline Re-Architecture (TIC B)

• Cut end-to-end pipeline runtime by 83% (12 hrs → 2 hrs) by re-architecting the workload on Apache Spark 3.
• Designed a versioning system that enables exact reproduction of any historical calculation across 60 onboarded data products by capturing the precise inputs and logic used for any past run.
• Replaced manual hand-offs with a fully automated end-to-end data pipeline and downstream generation flow.
• Advanced downstream data availability by 9 days each cycle by removing legacy system dependencies and building a new data model.

Kafka-Driven NRT Pipelines (TIC SLT / SHCA / SHLA)

• Cut data-quality issue resolution from 4–5 days of manual email follow-up to near-real-time by building a UI over Kafka-driven NRT pipelines that surfaces issues as they occur.
• Advanced downstream data availability by 17 days each cycle by shifting end-of-cycle batch processing to daily incremental processing.

Technology Intern

Natural Language Query Parser for ML Studio

• Removed the learning curve for Amex's ML Studio Logstash by building an end-to-end natural language query parser that translates plain-English questions into executable Logstash queries — an early NLP-to-DSL system.