Skip to content
View Usama1002's full-sized avatar
4 followers · 11 following

Achievements

Achievement: Arctic Code Vault Contributor

Achievements

Achievement: Arctic Code Vault Contributor

Highlights

  • Pro

Block or report Usama1002

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Usama1002/README.md

Muhammad Usama

From robotic navigation to signal integrity to 70B-parameter LLMs. Nine years of trying to ship research, and research what I ship.

usama1002.github.io  ·  Google Scholar  ·  Hugging Face  ·  LinkedIn  ·  usama@kaist.ac.kr

About

I did my Ph.D. at KAIST, closing a long arc through robotics, reinforcement learning, and foundation-model efficiency. My day splits three ways:

  • Research. Publishing with a thesis that actually fits on one line: representations teach more than rewards, adaptive rank beats fixed rank, and custom ASR beats a big generic model when the phoneme distribution shifts far enough.
  • Industry partnerships. Collaborations where the ML has to survive a semiconductor engineer reading the plots, or a doctor reading the confusion matrix.
  • Production. Voice agents, multi-agent systems, and medical pipelines running on AWS, GCP, and Azure for real end users every day.

The pieces talk to each other more than they look like they would. A representation that cleans up a DRAM eye diagram is cousin to one that steadies a long-context LLM. A custom ASR trained for a rare clinical accent is cousin to a robust vision model that transfers across fundus cameras without retraining. One idea, five surfaces.

What I work on

Pillar Sketch
Foundation-model efficiency LoRA, ARD-LoRA, DoRA, memory-augmented long context, multi-agent orchestration on 70B-class LLMs.
Voice AI & speech Custom ASR from scratch for low-resource and pathological speech, Whisper fine-tuning, Fish Audio cloning on low-GPU, Deepgram real-time STT, Aura TTS, Gemini Live speech-to-speech with VAD and barge-in.
RL & signal integrity PPO, DDPG, distributional and multi-agent RL, information-bottleneck representations, latent anomaly detection for high-speed signals; GA, PSO, and neural-surrogate DFE optimization.
Computer vision & medical AI Cross-domain fundus screening, time-aware transformers for longitudinal disease progression, frequency-aware CNNs for deepfake detection, wearable fNIRS signal processing.
Production & full-stack FastAPI, LangGraph, Docker microservices, CI/CD for ML, on-prem low-GPU deployment; React, Next.js, TypeScript, Node.js, GraphQL for AI-facing UIs.

By the numbers

Metric Value
AI/ML research 9 years
Peer-reviewed papers 17+, first author on 14
Filed patents 2 (RL equalizer optimization; multi-biosignal monitor)
ARD-LoRA on LLAMA-3.1-70B 99.3% of full-fine-tune at 0.32% trainable params
Multimodal adaptation memory −41% vs. DoRA baseline
Longitudinal medical prognosis −45.6% prediction error
KAIST academic ranking Top 3.7%
Korean proficiency TOPIK Level 2 and climbing

Now

A window into what I'm actively doing. This section gets edited, not appended.

  • Finishing the KAIST thesis on latent representation learning for DRAM anomaly detection and signal-integrity enhancement.
  • Five first-author IEEE Transactions manuscripts in parallel review (T-CAD, T-AI, T-Affective Computing) and one ICML 2026 submission on frequency-aware deepfake detection.
  • Shipping real-time medical voice pipelines with a custom Korean ASR trained from scratch, a domain where off-the-shelf models quietly fail.
  • Building multi-agent research and coaching bots on LangGraph + Gemini ADK, with cross-channel memory so a voice call continues seamlessly over SMS.
  • Reading Korean technical papers at a still-embarrassing pace, and loving it.

Selected writing

A curated slice. The full list lives on Google Scholar.

  • ARD-LoRA. Dynamic rank allocation for parameter-efficient fine-tuning. IEEE T-AI, 2025.
  • DistRL-IB. Distributional RL with information bottleneck for DRAM equalization. IEEE T-CPMT, 2026.
  • DRL DFE. Deep-RL DFE structure optimization for high-speed DRAM. IEEE T-CPMT, 2025.
  • Fair Representations. Learning flexible and fair data representations. IEEE Access, 2022.
  • Cross-domain fundus. Desktop-to-handheld diabetic-retinopathy transfer, 92.3% sensitivity / 94.2% specificity. Ophthalmology & Therapy, 2026.
  • SR-Retina. Dual-transformer residual super-resolution for retinal fundus images. Frontiers in Medicine, 2025.

Stack

Languages        Python · C/C++ · TypeScript · JavaScript · MATLAB · LabVIEW · LaTeX
ML & DL          PyTorch · TensorFlow · HuggingFace Transformers · scikit-learn
LLM / Agents     LangGraph · LangChain · Gemini ADK · MCP · RAG · NL2SQL
                 OpenAI GPT-4o · Anthropic Claude · Gemini 2.5 Pro/Flash · Qwen · Mistral
Voice AI         Deepgram · Aura TTS · Gemini Live · Whisper · Fish Audio · custom ASR
Vision & medical EfficientNet-B4 · ConvNeXt · time-aware transformers · fNIRS signal DSP
Full-stack web   React · Next.js · FastAPI · Flask · Django · Node.js · GraphQL
                 Gradio · Streamlit · Jupyter
Data & vectors   PostgreSQL · MongoDB · Redis · SQLite · Qdrant · Pinecone
Cloud & ops      AWS · GCP · Azure · Docker · Linux · CI/CD for ML · on-prem GPU
Hardware         NVIDIA Jetson · Raspberry Pi · PLC / HMI / SCADA · LabVIEW (Certified)
Human languages  English (fluent) · Urdu (native) · Korean (TOPIK L2)

Principles I try to write by

"Research that doesn't ship, shipping that isn't research. I try to avoid both."

Representations teach you more than rewards.

A well-picked baseline is worth three new architectures.

If the off-the-shelf model fails on your distribution, train your own. The fix is closer than you think.

Ship research. Research what ships. Keep closing the loop.

Say hi

If any of this rhymes with what you're building (signal-integrity ML, parameter-efficient training, voice agents that have to work in a clinic, medical vision that has to transfer across devices, or multi-agent systems that have to survive a real user), I'd enjoy talking.

Popular repositories Loading

  1. EBE-Exploration EBE-Exploration Public

    Code for our paper 'Learning-Driven Exploration for Reinforcement Learning'.

    Python 3 1

  2. MMD-fair-representations MMD-fair-representations Public

    Python 1

  3. pytorch-rl pytorch-rl Public

    Forked from navneet-nmk/pytorch-rl

    This repository contains model-free deep reinforcement learning algorithms implemented in Pytorch

    Python

  4. FlightGoggles FlightGoggles Public

    Forked from mit-aera/FlightGoggles

    A framework for photorealistic hardware-in-the-loop agile flight simulation using Unity3D and ROS. Developed by MIT FAST Lab.

    C++

  5. Counterfactual-Fairness---An-implementation-in-python Counterfactual-Fairness---An-implementation-in-python Public

    Here is the python implementation of paper Counterfactual Fairness by Matt Kusner et al. (https://arxiv.org/pdf/1703.06856.pdf)

  6. pytorch-template pytorch-template Public template

    Forked from victoresque/pytorch-template

    PyTorch deep learning projects made easy.

    Python