Mayank Sahu / AI systems engineering
Agent infrastructure, local AI memory, backend prototypes, and applied research systems with explicit constraints.
Infrastructure, memory, and the interfaces between people and intelligent systems.
Many agents.
One place to keep hold of them.
A cross-platform command center for parallel coding agents. Separate worktrees, real terminals, and live sessions you can inspect.
Explore the code
Git worktrees keep concurrent changes separate. The supervisor owns the process lifecycle; the repository stays inspectable.
$ fleetorch spawn codex
refactor-auth
"Refactor AuthMiddleware"
--repo .
Agent definitions are TOML files. PTY-backed execution preserves interactive command-line behavior on Unix and Windows.
$ fleetorch list
$ fleetorch attach refactor-auth
$ fleetorch logs refactor-authAttach to a running session, inspect its output, or detach without ending the agent. Multiple clients can share the same terminal stream.
The conversation ends.
The useful context stays.
Capture context across AI tools, distill it into typed memories, and retrieve it through local embeddings and hybrid search.
Inside the project
Active investigation across desktop automation, semantic graphs, and event vision. Working prototypes and benchmarked models.
Demonstrate a workflow.
Compile the execution graph.
Compiles demonstrated Linux desktop workflows into parameterized execution graphs. Supported paths replay natively through accessibility interfaces, with local recovery and agent escalation.
Inspect technical manuscript
Coordinate-free resolution.
Preserve semantic identity.
A semantic workbook graph for finding tables and other artifacts by meaning, tracking identity through structural sheet alterations, and tracing formula dependencies across workbooks.
Inspect technical manuscript
Tropical algebra.
A non-linear view of motion.
A high-throughput Rust time-surface ingestion core with Python C-bindings and synthetic test coverage. The ongoing research explores max-plus tropical algebra for sensor motion estimation.
Live research notebook
Latency, local execution, cost, reliability. The constraint is part of the problem, not a footnote.
Readable interfaces, observable behavior, and evidence close to the implementation.
Build a reference. Test the assumption. Follow the result even when it changes the direction.
Behind the decisionsThe Lab