Ian Greenberg
I build agent systems that run real business operations on local models — and contribute to the inference stack they run on.
Work
Systems in production, not demos.
vLLM — PR #46218
A dtype bug in the AWQ/GPTQ path for MLA attention. Inference engines are where a single wrong assumption costs everyone throughput, so the fix is small and the reading was not.
Director of AI & Automation
Built the company's AI stack: custom MCP servers over CRM, calendar and document storage; agent memory; a company knowledge wiki; a Telegram agent interface. Used daily by leadership, with every consequential action behind an approval gate. Also shipped customer-reactivation agents that recovered meaningful revenue in their first two months.
Automated course creation
Co-built software that turns source material into structured university courses. Deployed at roughly twenty universities through Z-School's partner network.
Founder & AI Engineer
The practice the Z-School and Cortex work was built through. Also where the internal tooling lives — automated financial reporting, marketing systems, and the harnesses the projects below came out of.
SMU — Center for Creative Computation
Affiliated researcher. Cluster and DGX SuperPOD allocation through 2027, used for inference work on large models — speculative decoding, quantisation, and what actually holds up at long context.
Before this I spent three years running a company I co-founded — one food truck in Lubbock that became three concurrent locations and 30+ employees. Payroll depending on you is why I care whether an agent's action is reversible.
Projects
Each links to a technical write-up.
Coding agents get animated bodies on an island you can watch. Underneath: a real orchestration pipeline in isolated git worktrees, local Whisper and Kokoro for voice, persistent memory across sessions.
Visit → realconnectorC++ · Qt6Remote-desktop control built around Moonlight, Sunshine and Tailscale — an operator console for machines, not screen sharing.
Read → localnpc-engineRustA harness for running local models on real tasks: tools, replayable traces, evals, and a gate that verifies before it lets a run finish.
Read → localbrowserPython · DockerSelf-hosted web search and page reading for local agents. No API keys, no per-query bill, and a quarantined model between the open web and the one holding your tools.
Read → npcagentPythonTurns a sentence into a durable run in a local repo — by driving the coding CLI you already have, inside tmux, and handing it tools back over MCP.
Read → daedalusTypeScriptCompiles a goal into a contract, freezes the evaluators behind a hash, and scores every attempt against them. Proof before progress.
Read → assetfarmerPythonOne keyed generation becomes registered sprite frames. An agent writes the choreography; deterministic code does the keying, slicing and sub-pixel registration.
Read → codesearchPythonTree-sitter chunking, embeddings and lexical search fused by reciprocal rank, then reranked. Built to answer questions about a codebase, not to grep it.
Read → ian-operatorJavaScript · MCPAn operator kernel where the model proposes and code disposes. Every consequential action is graded against a policy lattice before anything executes; defaults are refusals.
Read → ian-brainTypeScript · MCPA folder of markdown read as a graph. Vector, keyword and neighbour retrieval behind one router, so an agent pulls the relevant subgraph instead of the whole wiki.
Read → alwaysPython · RustA self-hosted voice assistant that listens continuously — capture, transcription, barge-in and turn-taking, built so interruption feels like conversation rather than a race.
Read → FutureMarketPython · ReactAn equities research system that logs every prediction with an entry price, reconciles it against what actually happened, and grades its own strategies on the result.
Visit →