Sina Barazandeh

Safe closed-loop agentic dose-response experiments using MHS

Thu, 27 Aug 2026 00:00:00 GMT

During the limited Model Hardware Standard (MHS) research preview (announced by Anthropic), I led Carnegie Mellon’s work using MHS. The Model Hardware Standard is a new standard for AI agents to safely operate physical equipment in scientific research and advanced manufacturing. The specification shared in the preview is provisional. MHS is model-agnostic; we used Claude as one compatible agent in this work.

We used MHS to let an AI agent operate a liquid handler, plate reader, robotic arm, and cameras across three computers with incompatible interfaces. Serial-dilution dose-response experiments ran about three times faster than our prior workflow. This work is featured in the MHS announcement; CMU SCS also posted on LinkedIn and X.

My focus was the AI orchestration and software integration. I developed the MHS drivers for all three instruments and the layer that let an agent — we used Claude as one compatible agent in this work — run the protocol, reject a saturated curve, and rerun with a lower maximum concentration (), within device-declared safety bounds. Integration took about eight hours, compared with the weeks of per-instrument integration that motivated the work.

Full write-up, photos, and video: Safe closed-loop agentic dose-response experiments using the Model Hardware Standard.

simPCL: An Open-Framework Simulator and Benchmark Generator for Programmable Cloud Laboratories

Mon, 01 Jun 2026 00:00:00 GMT

Problem

Evaluating autonomous agents for experimental orchestration on real laboratory hardware is slow and expensive. Progress on planning, fault recovery, closed-loop scientific discovery, and decision-making needs simulation environments and benchmarks that support controlled offline study.

Approach

I am building simPCL, an open-framework simulator and benchmark generator for programmable cloud laboratories. It supports controlled study of:

long-horizon experimental planning;
recovery from instrument and workflow failures;
reliable execution across multi-step protocols; and
closed-loop decisions in which measurements determine the next experiment.

simPCL complements our work on physical laboratory automation using MHS: real equipment demonstrates what an agent can accomplish, while simulation makes it possible to evaluate many scenarios reproducibly before using scarce laboratory time.

Status

Manuscript in preparation. Related work was presented at the AI Scientist Summer Workshop (Microsoft Research, 2026) in a contributed talk on agentic fault recovery. Code and benchmarks will be released with the paper.

FlashAttention and PagedAttention in MiniTorch

Thu, 15 Jan 2026 00:00:00 GMT

Problem

Efficient attention and memory management are central to training and serving large language models. Coursework in LLM Systems at CMU focused on implementing these ideas in a teaching framework rather than only using them as black-box library calls.

Approach

In a course research project with Instructor Lei Li, I worked on extending FlashAttention and PagedAttention to MiniTorch, implementing core attention and paging concepts in a minimal PyTorch-like system.

Status

Course research project (LLM Systems, CMU). Not a peer-reviewed publication. Code can be linked here when available.

Agentic Orchestration and Fault Recovery

Sat, 01 Mar 2025 00:00:00 GMT

Problem

Autonomous systems that orchestrate multi-step workflows must plan actions, respond to failures, and adapt as conditions change. Fixed scripts are often too rigid for this setting.

Approach

I am developing agentic systems that combine reinforcement learning with large language models for autonomous orchestration and fault recovery. The motivating application is programmable cloud laboratories, but the core problems are agent decision-making, recovery under failure, and reliable execution of long-horizon workflows.

This work asks how an agent should detect that a protocol has deviated from plan, choose a safe recovery action, and resume without repeating completed physical operations. It connects simPCL, which enables controlled offline evaluation, with our MHS physical-laboratory work, which tests closed-loop operation on real equipment.

Status

Work in progress. Contributed talk at the AI Scientist Summer Workshop (Microsoft Research, Cambridge, MA, 2026): Agentic Fault Recovery for Autonomous Orchestration of Programmable Cloud Laboratories.

Evolution-Informed Structural Subgraph Protein Representations

Sat, 01 Feb 2025 00:00:00 GMT

Problem

Foundation models like ESM-2 provide powerful protein embeddings, while explicit structural and evolutionary features may add complementary information for function prediction.

Approach

We developed graph-based protein representations using structural subgraph counts combined with evolutionary features. The method constructs protein graphs from structures and extracts interpretable features that capture local geometry and conservation patterns.

Evaluation

We are evaluating whether structural and evolutionary features improve pathogenicity prediction over ESM-2 and GNN baselines. The representation also provides interpretable features that domain experts can analyze.

Status

Manuscript in preparation. Code includes baseline comparisons and evaluation protocols. Code and data will be released.

Reinforcement Learning for Senescence Biomarker Discovery

Wed, 15 Jan 2025 00:00:00 GMT

Problem

Identifying minimal gene panels that distinguish senescent cells from healthy cells is challenging due to high-dimensional data and complex cellular states in unlabeled data.

Approach

I built a reinforcement learning framework using proximal policy optimization to select informative gene subsets. The reward function balances cluster separation with panel parsimony, and the method handles positive-unlabeled scenarios common in biological data.

Status

Manuscript in preparation. Code will be released with reproducible experiments and ablation studies.