I'm Maahir Garg, and I build careful systems for unruly data.

Built an internal agentic CLI agent and a sandboxed Chainlit assistant over proprietary models, with 8 custom tools, multi-agent orchestration, and skill-routing, giving an 11-person team AI tooling where classified-data constraints rule out Copilot. Engineered step-level guardrails via custom hooks that audit and validate every agent action, and designed an agentic audit pipeline that queries Snowflake and SharePoint as tools, runs ETL, and generates audit-ready reports from plain-English prompts, automating two previously manual data sources from hours to near-instant.

Co-inventor of a multimodal tracking framework (patent application in progress) for hand-eye coordination assessment in stroke rehabilitation, unifying Apple Vision Pro and iPhone into one tracking surface at 89% cross-device accuracy, 30 fps, and sub-millisecond latency (Swift, RealityKit). Ran controlled user studies with healthcare practitioners and iterated system design from empirical findings. Co-author on a research paper in preparation.

Three semesters across CS3230 (Design & Analysis of Algorithms), CS2040S (Data Structures & Algorithms), and CS1231S (Discrete Structures). Coached students through dynamic programming, graph algorithms, and complexity analysis. 4.8/5.0 teaching effectiveness rating vs 4.2 faculty average; named to the NUS Honour List of Student Tutors.

Effectively a TA for a full semester-long software-engineering project: advised 12 student teams across the whole build: architecture, implementation planning, code review, and release management. All 12 projects passed at or above their targeted grade level.