Muhammad Ibrahim Khalid
Final-year BEng Electronic Engineering, HKUST
I work on efficient ML systems: getting large models to run under tight memory and latency budgets, from vision-language models on an NVIDIA Jetson to LLM inference that no longer fits in GPU memory.
My final-year thesis with Prof. Zili Meng studies weight and KV-cache placement for LLM inference on wireless GPU infrastructure. I'm also a research intern at Gray Scrubs AI, working on generating rare, safety-critical events. I'm looking for research positions and master's programmes starting in 2027.
News
- Started as a research intern at Gray Scrubs AI, working on generating rare, safety-critical events.
- Submitted my thesis proposal on weight and KV-cache placement for LLM inference, with HK$22,500 of compute funding on the HKUST SuperPOD.
- Released SignalProof, a backtest auditor that catches look-ahead leakage.
- Joined LWK + Partners to automate building-code compliance checks in Revit.
- Finished the Jetson VQA project: 2.2× lower on-device latency from token pooling alone.
Research
-
Hot/Cold Weight Swapping and KV-Cache Placement for LLM Inference on Wireless GPU Infrastructurein progress
WiCi lets a mobile client run CUDA on a nearby GPU server over Wi-Fi. I'm designing the server-side memory policy for when the model and its KV cache exceed one H800's 80 GB: hot layers and live KV pages stay in VRAM, cold layers are prefetched from DRAM or NVMe while the current layer runs, and only tokens cross the wireless link.
Target: decode at ≥ 80% of a fully VRAM-resident baseline, compared against FlexGen-style offload, measured as TTFT, p95 inter-token latency and tokens/s.
-
Generating Rare, Safety-Critical Eventsongoing
Rare, safety-critical events are hard to capture, which makes them hard to learn from. I'm part of the research team at Gray Scrubs AI working on generating them. More detail once the work is public.
Selected projects
-
Real-Time Visual Question Answering for Low-Vision Navigation on a Jetson Orin NX
A camera, a Jetson and a small vision-language model (SigLIP + Qwen2.5-0.5B) that answer crossing, stairs and obstacle questions out loud. I built the vision encoder, token pooling, the evaluation framework and all on-device benchmarking, plus CUDA/PyBind11 kernels for preprocessing.
Pooling image tokens 576 → 36 cut mean on-device latency from 747 to 346 ms (2.2×) while accuracy rose from 25% to 50%.
-
SignalProof: Auditing Equity Backtests for Look-Ahead Leakage
A Hong Kong equity research lab that injects look-ahead leaks on purpose and audits every run with purged walk-forward validation, the Deflated Sharpe Ratio and the Probability of Backtest Overfitting. A local LLM writes the research memo from the metrics alone and never sees prices.
Both injected leaks are flagged FAIL (PBO 0.90 and 0.66); the clean pipeline gets WARN rather than a false pass.
-
Automating Air-Ventilation Compliance (PNAP APP-152) in Autodesk Revit
A pyRevit extension (about 17,000 lines of Python) that runs Hong Kong's street-canyon and air-permeability checks on the live building model, including wind corridors that may bend by up to 15°, and draws the PASS/FAIL evidence into dedicated audit views.
A typical run went from 10–30+ minutes to 30–90 seconds after I split the plane search into a cheap sweep and a full check on the best candidates.
proprietary code -
Retrieval-Augmented Generation on an 8 GB Laptop GPU
A RAG pipeline sized for an RTX 4060: embeddings run on the CPU so the full 8 GB goes to the language model, with one switch between a local 4-bit model and NVIDIA's hosted NIM behind the same response schema.
Over 10 runs: 14.94 tokens/s locally vs 33.66 tokens/s on NIM (0.771 s to first token).
-
Signed Area: Linear Algebra as a Scroll-Driven Story
One Canvas 2D scene in which the same particles move through orientation tests, convex hulls, SVD, PCA and softmax attention as you scroll. All of the math is implemented by hand (no Three.js) and unit-tested with Vitest.
-
Smart Recycling Station on an STM32
A voice-activated sorting station: real-time audio capture with DMA, a live volume display on the LCD, and servo-driven sorting, all running on an STM32.
Experience
- Gray Scrubs AI · Research InternSep 2026 – presentResearch on generating rare, safety-critical events.
- LWK + Partners · Workflow Automation InternJun – Aug 2026PNAP APP-152 compliance automation in Revit, and a proof of concept connecting an AI agent to Revit 2024 over the Model Context Protocol.
- Gray Scrubs AI · Hardware Engineer InternMar – May 2026Maintained the NVIDIA Jetson fleet used for multi-country pilots and paper experiments.
- SUSTAINALPHA · Core EngineerFeb – May 2026Monte Carlo and mean-variance risk forecasting, and LLM news summarisation, for an HKSTP-backed risk platform.
- HKUST · Undergraduate Research (UROP)2026Added Omi Glass support to the xg-glass-sdk (merged upstream), then surveyed GPU remoting and LLM offloading, which became my thesis topic.
Other projects
- AI Document Reviewer: FastAPI, React and SQLite app for LLM-assisted document review, deployed on AWS EC2.
- Stock analytics dashboard: incremental yFinance ETL into SQLite with a Flask API and live charts.
- Bluetooth robot car: Arduino and an L298N motor driver, controlled from a React Native app.
- C++ hash table: templated, and grows by linking larger tables instead of rehashing.