Distributed and cloud
Event-driven services that reconcile dropped, duplicated, and out-of-order updates across regions. Graph and document data modelling, versioned APIs, and the observability needed to prove a consistency commitment.
Software engineer — Seattle, Washington
Builds systems where correctness is the hard part.
Seven years on distributed systems and machine learning over large graphs, with earlier work on firmware running under fixed memory and power budgets. The common thread is correctness under concurrency, failure, and scale.
Built systems at
- Amazon Web Services
- Samsung Research
- University of Arizona
- Baylor University
Latest
-
Started as founder and engineer at a stealth AI startup in Seattle, building an agentic platform over municipal records and permit filings.
-
Wrapped up at AWS Infrastructure Supply Chain after shipping the hardware system of record and its event-driven reconciliation service.
-
"Multivocal Study on Microservice Dependencies" published in the Journal of Systems and Software. Read on ScienceDirect
-
Paper on tuning fuzzy c-means in a VAE-based GNN approach presented at IEEE/ACM ASE 2024. Read on the ACM Digital Library
What I work on
Graph machine learning
Graph neural networks and variational autoencoders over large dependency graphs, with clustering to recover latent structure, and reproducible benchmarks to measure the result against.
Embedded and firmware
FreeRTOS firmware and device drivers in C and C++ over SPI, UART, and I2C. Bluetooth Low Energy and GATT control and data planes, rollback-safe over-the-air updates, and profiling under fixed memory and power budgets.
Selected work
Chosen for the difficulty of the problem, not how recent it is
Reconciling an event stream that drops, duplicates, and reorders
The hardware system of record carried a 24-hour consistency commitment while taking asynchronous updates from more than fifteen cross-region systems. An event-driven reconciliation service detected divergence and replayed it safely, so the commitment was something the system held rather than something the design asserted.
- 24 h
- consistency commitment, held
- 15+
- cross-region systems integrated
- 60% → 90%
- integration test coverage
Change impact as a graph traversal, not a lookup
Millions of hardware components, assemblies, and bills of material modelled as a graph, so that asking what a change breaks could be answered across products, sites, and programs before the change shipped. A rules engine evaluated site, region, and program policy from declarative configuration.
- 60%
- gain in hardware qualification efficiency
- 100+
- manufacturing sites onboarded without per-class code
Feature extraction across ten million transactions
Distributed pipelines extracting opcode and control-flow features from Solidity bytecode, then training anomaly-detection models over the result. At that volume the extraction step, not the model, decides whether the experiment is runnable at all.
- 65%
- improvement in analysis efficiency
- 10M+
- transactions processed
Sensor streaming that cannot miss a 15 ms deadline
Heart-rate, stress, and sleep tracking across more than 90 exercise modes, on a microcontroller whose memory and power budgets were fixed before the work started. Races and timing defects at that level leave nothing in a log, so they were found on the running instruction stream with JTAG and Trace32.
- 15 ms
- streaming cadence held
- 2M+
- devices shipped, no known production defects
- 99%
- rollback-safe update success, zero bricked devices
Publications
All 12 publications-
Multivocal Study on Microservice Dependencies
JSS 2025 volume 222, article 112334 Read on ScienceDirect
doi 10.1016/j.jss.2025.112334
-
Using Static Analysis to Aid Monolith to Microservice System Transformation: Tuning Fuzzy c-Means in a VAE-Based GNN Approach
IEEE/ACM ASE 2024 Read on ACM Digital Library
doi 10.1145/3691621.3694933
-
Benchmarking Micro2Micro Transformation: An Approach with Graph Neural Networks and Variational Autoencoders
Cluster Computing 2024 volume 27, number 4, pages 4171–4185 Read on SpringerLink
doi 10.1007/s10586-024-04526-z
-
A Multi-variant Benchmark for Microservice Systems in Software Engineering Research
ECSA 2024 pages 21–29 Read on SpringerLink
doi 10.1007/978-3-031-71246-3_3
-
Comparative Evaluation of Java Virtual Machine-Based Message Queue Services: A Study on Kafka, Artemis, Pulsar, and RocketMQ
Electronics 2023 volume 12, article 4792 Read on MDPI
doi 10.3390/electronics12234792
Writing
All postsReconciliation is a product feature, not a repair step
Asynchronous integration means updates arrive dropped, duplicated, and out of order. Treating the fix as a background chore is what makes consistency commitments impossible to keep.
Experience
Full CV-
May 2026 – present
Founder and Engineer current
Stealth AI startup, Seattle, Washington
An agentic AI platform that monitors municipal records, council agendas, zoning dockets, and permit filings across U.S. cities, converting them into early business signals for commercial real estate and multi-site operators.
-
Jan 2025 – Apr 2026
Software Development Engineer, AWS Infrastructure Supply Chain
Amazon Web Services, Seattle, Washington
Hardware system of record for AWS data centres and device manufacturing: graph modelling in Neptune, event-driven reconciliation, and a configuration-driven rules engine behind a GraphQL interface.
-
Mar – Dec 2024
Research Associate II, Software Engineering
CloudHubs Lab, University of Arizona, Tucson, Arizona
Led the lab's graph-learning line — graph neural networks and variational autoencoders over static-analysis dependency graphs — and released a multi-variant benchmark suite for architecture research.
-
Mar – Dec 2024
Full-Stack Developer, International Collegiate Programming Contest
ICPC Foundation, University of Arizona, Tucson, Arizona
Contest-management platform in Spring Boot, React, and PostgreSQL serving more than 50,000 students a year across 50+ regions, made multi-tenant so regional managers publish independently.
Held alongside the CloudHubs research appointment.
-
May – Aug 2023
Software Engineer, Summer Intern
Tyson Foods, Springdale, Arkansas
Unreal Engine virtual-reality quality-inspection training for 200+ manufacturing personnel, with 3D asset generation automated from two weeks to ten minutes.
A summer internship during the Baylor appointment.
-
Jan 2022 – Dec 2023
Graduate Research and Teaching Assistant, Computer Science
Baylor University, Waco, Texas
Distributed pipelines on Amazon EMR, a controlled evaluation of four JVM message queues, and an ACID transaction manager; taught software engineering and C++ to 125 students across five sections.
-
Jul 2018 – Dec 2021
Software Engineer, Wearable Platform
Samsung Research, Dhaka, Bangladesh
FreeRTOS firmware and drivers, the Bluetooth Low Energy and GATT planes, and rollback-safe over-the-air updates for wearables shipped to more than two million devices.
-
Jan – Jun 2018
Software Engineer
Divine IT Limited, Dhaka, Bangladesh
Django, DRF, and PostgreSQL enterprise resource planning modules across payroll, HR, inventory, sales, and procurement for 5,000+ daily users.
Contact
- Papers
- Google Scholar profile
- Profile
- LinkedIn profile
- Status
- U.S. Permanent Resident