LLMs at the edge
Splitting a small LLM between a smartphone and an edge server: latency- and energy-aware partitioning
Run part of a small language model on the phone and the rest on a nearby edge server, and choose the split point that gives the fastest, most energy-efficient answer for the current network.
Read the details →An LLM request router: on-device, edge or cloud model, chosen by cost, latency and answer quality
Send each user question to the cheapest model that can answer it well — a tiny local model for easy questions, a large cloud model only when needed.
Read the details →Scheduling LLM inference requests on a shared edge GPU server with KV-cache awareness
Decide in which order and in which batches to serve many users of one GPU so that nobody waits too long and the GPU memory (the KV cache) is not wasted.
Read the details →An offline campus assistant on a quantized small language model running on a Raspberry Pi / Jetson
A question-answering assistant about university rules, courses and services that runs entirely on a small device, without internet and without sending data out.
Read the details →Prompt compression for LLM offloading over low-bandwidth networks
Shorten long prompts on the device before sending them to the edge or the cloud, so that answers come back faster and cheaper with the same quality.
Read the details →Speculative decoding between a small on-device model and an edge LLM
A tiny model on the phone guesses the next words quickly; a larger model on the edge server only checks them — answers keep the large model's quality but arrive much faster.
Read the details →Edge and cloud scheduling
Budget- and deadline-aware autoscaling of Kubernetes microservices with reinforcement learning
Learn when to add or remove copies of each microservice so that response times stay under target while the cloud bill stays within budget.
Read the details →Carbon-aware scheduling of data-pipeline workflows across cloud regions
Run each step of a data pipeline at the time and in the cloud region where electricity is cleanest, without missing the deadline.
Read the details →Spot-instance-aware workflow scheduling to cut cloud costs for data pipelines
Use cheap but interruptible cloud machines (spot instances) for a workflow, and recover safely when one is taken away.
Read the details →Predicting serverless cold starts and pre-warming functions with machine learning
Predict when a serverless function will be called next and start it just before, so users do not wait for a "cold start" — without paying for idle machines.
Read the details →A digital twin of a multi-access edge computing network for testing task-offloading policies (open simulator and benchmark)
An open, easy-to-use simulator in which students and researchers can compare task-offloading algorithms on the same realistic scenarios.
Read the details →Scheduling deep-learning training jobs on a shared university GPU server
Many students and researchers share a few GPUs: decide who runs when, so that jobs finish sooner, nobody waits forever and the GPUs never sit idle.
Read the details →Federated learning and IoT
Communication-efficient federated learning for human activity recognition on wearables
Train an activity-recognition model (walking, sitting, falling…) across many smartwatches without collecting the users' sensor data, while sending as little as possible over the network.
Read the details →Federated anomaly detection for smart-building and factory IoT sensors
Detect faulty machines or abnormal energy use from sensor data in several buildings or factories, without moving their raw data to one place.
Read the details →Federated LoRA fine-tuning of small LLMs across organisations without sharing data
Several organisations (hospitals, schools, company branches) adapt the same small language model to their domain together, exchanging only small LoRA adapters, never their documents.
Read the details →Detecting slow and malicious clients in federated learning on edge devices
Keep federated training fast and correct when some devices are very slow (stragglers) and some send wrong or poisoned updates.
Read the details →UAVs and smart city
UAV-assisted edge computing for disaster response: drone trajectory and task offloading with deep reinforcement learning
Drones carrying small computers fly over a disaster area, collect data from ground sensors and rescuers, and process it on board; learn where to fly and what to compute.
Read the details →Drone inspection of solar panels and power lines: on-board vs edge inference trade-offs
Detect defects (cracked panels, damaged insulators) in drone images, and decide what to compute on the drone and what to send to a ground edge server.
Read the details →Privacy-preserving campus traffic and parking sensing with edge AI cameras
Count cars and free parking spots from cameras, processing the video on a small device next to the camera so that no image of a person or plate ever leaves it.
Read the details →Short-term energy forecasting for smart buildings with transfer learning at the edge
Predict tomorrow's electricity use of each campus building from its meters and the weather, with a small model running on site, even for new buildings with little history.
Read the details →Security-aware computing
Security-aware task offloading: choosing the encryption level versus latency for IoT data
When an IoT device sends work to an edge server, choose how strongly to protect the data (none, light, strong encryption) according to its sensitivity, without missing deadlines.
Read the details →LLM-based log anomaly and intrusion detection for small-company networks
Read server and network logs automatically, flag unusual events, and explain them in plain words to an administrator who is not a security expert.
Read the details →Lightweight federated intrusion detection for IoT gateways
Home and office IoT gateways learn together to recognise attacks in network traffic, each keeping its own traffic private.
Read the details →Phishing e-mail detection with a small language model running on the mail gateway
Catch phishing e-mails before they reach staff and students, with a compact model that runs on the organisation's own server and explains why a message is suspicious.
Read the details →NLP and data quality
Few-shot relation extraction to build knowledge graphs from company documents (contracts, maintenance reports)
Extract facts such as "machine X — has_fault — bearing" or "party A — pays — party B" from documents with only a handful of labelled examples, and store them as a searchable knowledge graph.
Read the details →A retrieval-augmented (RAG) assistant over enterprise technical documents, with hallucination checking
A chat assistant that answers engineers' questions from the company's manuals and cites the exact passage — and warns when its answer is not supported by the documents.
Read the details →Quality control of crowdsourced and LLM-generated labels: aggregating unreliable annotators
Combine labels from several imperfect sources — human crowd workers and LLMs — into one reliable label, by learning how good each source is.
Read the details →Automatic, trustworthy feedback on student programming assignments with LLMs
Give students instant hints on their code — what is wrong and where to look — without giving away the solution, and check that the hints are correct.
Read the details →Benchmark
An open benchmark of edge AI hardware (Jetson, Raspberry Pi, phone NPU) on real workloads
Measure, in the same way, how fast and how energy-hungry common AI tasks are on popular small devices, and publish the results so companies can choose hardware.
Read the details →A reproducible benchmark of federated learning frameworks on real edge devices
Run the same federated training tasks with popular frameworks (Flower, FedML) on real phones and Raspberry Pis, and publish how they compare in time, energy and accuracy.
Read the details →Interested in a topic?
Write to me with the topic title and the courses you have taken. Every capstone is an individual project. Topics can be adapted to your interests, and companies are welcome to propose a real problem.
dhamza@kean.edu