Hi, I'm Musa 👋 Machine Learning Researcher and Software Engineer. BS in Computer Science and Mathematics from NYU Abu Dhabi, with a semester at NYU Courant taking graduate ML courses. I care about why a method works, how it compares to the alternatives, and what the literature actually says: then I go build the thing, ship it, and keep it running. Currently: Machine Learning Research Assistant @ Clinical AI Lab, NYU Abu Dhabi

I ship services end to end: API design, data and retrieval layers, automated testing, CI, and production observability. The part I enjoy most is the unglamorous middle, where a design decision meets a real failure mode.

FIVE Concierge (https://fiveconcierge.vercel.app/) is the clearest example. TypeScript, Next.js, Node.js, Redis, deployed on Vercel:

• Owned it from ingestion through API design, testing, deployment, and monitoring.
• Removed the vector database entirely: truncated embeddings from 3072 to 768 dimensions with L2 renormalization and ran cosine similarity as an in-process dot product. The index fits in 443 KB and the retrieval path benchmarks at 74,000+ operations per second with p99 latency under 18 microseconds, at a measured cost in precision on longer queries.
• Debugged a rate limiter that was silently failing under serverless autoscaling: per-instance in-memory counters multiplied the intended ceiling by the number of live instances. Replaced it with an atomic Redis sliding window shared across instances, and closed an IP-spoofing bypass by deriving client IP from the platform's signed header.
• 67 unit and integration tests in Vitest at 99.6% statement and 91% branch coverage, with regression tests pinning the rate limiter fix. GitHub Actions runs typecheck, lint, and the full suite on every push and pull request, gated by branch protection.
• Structured JSON logging on every request: request IDs, outcome and error classification, and total, retrieval, and generation latency split out separately, with log fields typed as fixed enums so user-submitted text can never reach a log line.

Outside of that I have built interpreters and worked through operating systems, compilers, networks, databases, and programming language semantics.

Adapting LLMs to Arabic medical tasks. Parameter-efficient targeted LoRA adaptation, where the layers to adapt are chosen using activation patching and KL divergence evidence rather than picked by convention. Evaluated against general purpose models, medical models, and standard adaptation baselines. Submitted to EMNLP 2026.

Learned cost models for compilers. Extended the LOOPer and DRL-based PEARL cost models in Tiramisu, comparing GNNs, transformers, and recurrent architectures for the cost prediction component.

Deep RL for robot self-flocking. Extending work on non-holonomic robot flocking from range and bearing data, and the harder range-only case: built the simulation environment and a PPO-based training strategy.

Self-supervised world models. Implemented a JEPA to model agent trajectories in a two-room environment, trained on 2.5M frames with VICReg for collapse prevention. Placed 2nd of 36 teams in the graduate CSCI-GA.2572 Deep Learning course competition.

The research above only works if the engineering underneath it does, so a lot of my time goes to training infrastructure, HPC jobs, experiment tracking, evaluation pipelines, and data tooling.

MCP servers and tool calling to connect models to real systems, multi step workflows with explicit failure handling, and evals that separate retrieval failures from generation failures. Grounding is a systems problem as much as a modelling one: retrieval gates, refusal paths, and citations that the UI can actually trust.

Languages: TypeScript/JavaScript, Python, C, C++, SQL, Bash, OCaml

Backend and data: Node.js, Express, REST APIs, GraphQL, PostgreSQL, SQLite, Redis, Kafka, event-driven design, concurrency

Infrastructure and tooling: Git, GitHub Actions, CI/CD, Docker, Linux, Vercel, AWS, Kubernetes

Testing and reliability: Vitest, pytest, unit and integration testing, coverage reporting, benchmarking, structured logging and observability

ML: PyTorch, scikit-learn, NumPy, pandas, W&B, LoRA/PEFT, retrieval-augmented generation

• Email: musa.khan@nyu.edu
• LinkedIn: https://www.linkedin.com/in/musa-ji