Building the backend infrastructure for a multi-tenant fleet operations platform from zero to production. Core focus: real-time data pipelines, geospatial systems, and AI-powered automation for logistics.

As one of two engineers on the team, I architected and built the core backend infrastructure for a fleet operations platform now used daily by 15+ transportation companies managing 500+ vehicles.

Core technical work:

• Designed fault-tolerant ingestion pipelines processing 600K+ GPS datapoints/day via AWS Kinesis

• Built event-driven messaging architecture using Redis Streams/Pub/Sub with async Python workers

• Integrated OSRM routing engine with PostGIS for real-time geofencing and sub-200ms queries across millions of location records

• Developed OpenAI-powered WhatsApp automation that reduced manual dispatch overhead ~50%

• Created OLAP pipelines with ML feedback loop — capturing human corrections of AI-predicted statuses as labeled training data

• Built React/Next.js dashboard with WebSocket-driven live tracking

Stack: Python, FastAPI, PostgreSQL, PostGIS, Redis, AWS (Kinesis, Lambda, S3, ECS), React, Next.js, OpenAI API

Evaluating and improving LLM robustness through adversarial testing and RLHF training signal contribution.

Developed 150+ adversarial test cases evaluating LLM performance across code generation, mathematical reasoning, and logic domains. Identified systematic failure patterns that improved evaluation methodology.

Contributed to RLHF training pipeline by authoring reference solutions and iteratively refining model outputs. Also reviewed peer evaluations to maintain quality standards across the evaluation team.

TA for Michigan's core algorithms course, supporting 1000+ students per semester in graph algorithms, dynamic programming, and algorithmic complexity.

Debugged and reviewed C++ implementations of graph algorithms, dynamic programming, and priority queues across hundreds of student submissions over 4 semesters.

Led weekly problem-solving sessions on algorithmic complexity and optimization. Co-designed assessments and maintained consistent grading rubrics for course serving 1000+ students per semester.

Built hybrid cloud architecture bridging legacy IBM mainframe systems with modern AWS infrastructure for insurance quote automation.

Led development of a cloud-native insurance calculation platform as part of the University of Michigan MDP program (7-person team).

Key contributions:

• Designed hybrid architecture connecting IBM z/OS mainframe (DB2) with AWS via Apigee proxy and z/OS Connect — enabling real-time customer data access from legacy on-prem systems

• Re-engineered 700+ actuarial pricing models from Excel into Python/AWS Lambda, reducing quote generation from 10 minutes to 3.5 minutes (65% faster)

• Delivered AQUA (React.js + DynamoDB), eliminating manual data entry and improving underwriting workflow efficiency by 90%

Stack: Python, AWS (Lambda, API Gateway, DynamoDB, S3), React.js, IBM z/OS, Apigee