Manthan Gupta is an AI research engineer who writes about inference, agents and distributed systems. He created Water, a workflow harness around agent tasks, and Auto-Inference-Optimiser, an experiment loop for studying language-model inference on Apple Silicon.[1][3][4]
Start with Auto-Inference-Optimiser →
X Background
Gupta describes his approach as building systems to understand them and writing to clarify what he learns. His current interests include GPU kernels, inference engineering and distributed training. Alongside technical essays, he lists projects such as Water, the memory-layer tool BYOM and CricLang, a programming-language experiment. That combination makes his writing a route into the implementation choices beneath an AI application, rather than a catalog of model announcements.[1]
Contributions
A harness around agent work
Water separates a workflow into tasks, flows and an execution engine. A task has typed inputs and outputs and an execution function; a flow determines whether tasks run in sequence, in parallel or through branches and loops. The engine handles ordering, retries, checkpoints and events. The project also exposes approval and sandboxing features. This lets a developer reason about orchestration and recovery independently from whichever model or agent performs a task.[3]
Optimization that keeps quality visible
Auto-Inference-Optimiser fixes an evaluation harness while allowing a coding agent to change the inference pipeline and record experiments through Git. Its report includes changes that were rejected as well as retained. One instructive result is that an apparently attractive numerical metric can hide damaged output: a cache-quantization experiment passed one quality measure while failing simple task checks. The project’s value is the inspectable search process and its gates, rather than a speedup promised for every model or Mac.[4][5]
Tools, writing & talks
Tools
Start here
Auto-Inference-Optimiser
github.comAn experimental loop for changing an MLX inference pipeline while holding evaluation and quality checks fixed.
Start with the experiment log to see accepted changes, reversions and quality failures. Reproducing a run requires Apple Silicon, Python and the documented MLX model; reading the results requires only basic inference concepts.
Repository owner[4]Original source- Watermanthanguptaa.in
A workflow harness with typed tasks, flow composition, retries, checkpoints, approval gates and execution events.
Creator and author[3]Original source
Writing, research & talks
An account of inference experiments, failed approaches and the constraints behind measured improvements.
Author[5]Original source