Projects
Systems I have designed, built, and shipped — at Nokia Bell Labs (2022–present) and during my PhD. Each one spans the full stack: algorithm design, infrastructure, and a working end-to-end deployment.
In progress
ElasticServe — SLO-aware LLM inference serving control plane
Personal project on NVIDIA DGX Spark · 2026
Building an adaptive, multi-tenant inference-serving control plane that routes mixed workloads across vLLM/SGLang model endpoints: admission control and backpressure, SLO-aware weighted-fair scheduling, adaptive concurrency, and model-residency policy, measured against static baselines on goodput (TTFT/TPOT SLOs), tail latency, fairness, and cache efficiency, with Prometheus/Grafana observability throughout.
Nokia Bell Labs
XRNet: network-aware XR delivery over 5G
NSF “Breaking Low” program (~$3.5M) · Oct 2025 – present
Multi-institution project with the University of Michigan (PI), USC, Columbia, Duke, and Nokia Bell Labs (Co-PI), targeting motion-to-photon latency requirements considered infeasible on today’s Internet. My contribution is one of the first open end-to-end implementations of the 3GPP Release-18 PDU Set Marking RTP header extension (TS 26.522) in a production WebRTC stack — so every video packet tells the network how important it is.
- Implemented the extension across three codebases: Unity’s C# application API, its native C++ plugin, and a modified, rebuilt libwebrtc — every RTP packet carries standards-compliant frame-importance and sequencing metadata usable by 5G schedulers, negotiated via SDP.
- Built a real-time H.264 NAL-parsing classifier (RFC 6184 STAP-A/FU-A aware) that maps keyframes, parameter sets, and reference frames to PDU Set Importance values during packetization of a 1080p/60 fps stream.
- Delivered the full interactive cloud-XR loop: a Unity-rendered 3D scene streamed over DTLS-SRTP to a stock browser, user input returned through a WebRTC DataChannel, Node.js signaling, and STUN/TURN traversal for operation across a 5G core.
- Validated at the wire level with Wireshark — byte-exact extension contents on every packet of the encrypted stream — and on a three-node Linux tc/netem testbed whose QoS classifier reads the importance marking directly from packet bytes: PDU-Set-aware handling kept the interactive stream smooth under loss and jitter that froze importance-blind delivery.
- Demonstrated the complete end-to-end system live at the University of Michigan (June 2026), validating the program’s central approach in practice.
Edge–cloud scalable 3D Gaussian Splatting delivery
Patented · Jun 2025 – Mar 2026
- Architected an edge–cloud 3DGS scene-delivery system that folds live network state into splat selection and prioritization, jointly optimizing network- and application-level performance.
- Resolved 3DGS scalability limits for compute- and memory-constrained client devices and for simultaneous multi-user access under network contention.
- Approved by the Nokia Patent Board; assessed by expert reviewers as a first-of-its-kind cross-layer XR streaming approach, relevant to Nokia’s MantaRay RIC and AI-RAN product lines.
ML-driven spectrum defragmentation for optical transport networks
Demonstrated at OFC 2025 · Mar 2024 – Mar 2025
- Developed an automated, ML-driven spectrum-defragmentation method that improves energy efficiency in optical transport networks.
- Built the automation pipeline end to end and demonstrated it live at the Optical Fiber Communications Conference (OFC 2025).
PSASlicing: SLA-aware reinforcement learning for O-RAN slice management
IEEE GLOBECOM 2024 · Dec 2023 – Aug 2024
- Designed a perpetual, SLA-aware reinforcement-learning method for O-RAN network-slice management that sustains SLA compliance over long operating horizons — RL that has to keep working in production conditions, not just in a training run.
End-to-end 5G/O-RAN system and RIC applications
Jun 2022 – Dec 2022
- Implemented a complete 5G end-to-end system with OpenAirInterface, Open5GS, and O-RAN components.
- Developed xApps (near-RT RIC) and rApps (non-RT RIC) hosting ML-based radio resource management agents for RAN slicing.
PhD systems work (UNCC / NJIT)
FedVision — federated video analytics with edge computing
IEEE OJ-CS 2020, 45+ citations
- Designed a federated video analytics system that jointly optimizes on-device inference and computation offloading across an edge network, trading off detection accuracy against end-to-end latency.
- Built the testbed on NVIDIA Jetson TX2, nvidia-docker, OpenWRT, and desktop GPUs; used learned surrogates (neural processes) and black-box optimization to pick models, frame resolutions, and offloading rates at runtime.
FLEX — trading edge computing resources for federated learning
Demoed at IEEE INFOCOM 2021
- Designed and implemented a blockchain marketplace (Ethereum smart contracts in Solidity) where edge nodes sell compute to federated learning tasks, including the TensorFlow-side workflow and interfaces between edge, global, and chain nodes.
Earlier systems
- Decentralized video analytics — distributed testbed with heterogeneous edge hardware; consensus and incentive design accounting for both compute speed and network latency (2021).
- SDN for power distribution systems — learning-based network management and DoS mitigation for distributed grid controllers, on ONOS/Ryu SDN controllers with NS-3, OPAL-RT, and OpenDSS co-simulation (2019–2020).
- Underwater communication system — closed-loop underwater fault monitoring with custom RF/optical channel models in NS-3 and an OPAL-RT–GNU Radio bridge, prototyped in a lab seawater setup (2018).
- Edge datastore for distributed vision analytics — key-frame/feature store written in Go with REST APIs and SDN-fed network awareness (ACM/IEEE SEC 2017 poster).