ZAC Y.-C. LIU

Building a causal reasoning harness that elevates open models on agentic tasks

About

I’m Zac, co-founder of a stealth AI startup building a causal reasoning harness that enables open-source and mid-tier models to compete with frontier models on complex, agentic, and adaptive tasks—without expensive fine-tuning or reinforcement learning.

The core technology is Causal Reflection, a neuro-symbolic framework I co-developed (NeurIPS 2025). It combines dynamic causal graph construction and updating, structured self-reflection on actions and outcomes, and efficient real-time planning. This lets agents better understand how their environment works, reason about cause and effect, adapt to novel out-of-distribution situations, and make more reliable decisions—directly addressing key limitations of current LLM-based agents.

We are prioritizing validation on public agentic benchmarks (including SWE-bench, TAU-bench, and related suites) that closely reflect enterprise workflows such as software engineering, tool use, and multi-step decision-making. Early results aim to demonstrate that the harness can meaningfully boost mid-tier open models toward frontier-level performance on these tasks.

Previously I was Founding ML Engineer at BrainChain AI (supply-chain disruption detection), Head of AI / Chief Data Scientist leading R&D and product for an AI delivery + MLOps platform (A360 AI / Hypergiant), and Venture Fellow at Laconia Capital. I have built, led, and mentored cross-functional teams taking generative AI, LLMs, data-centric AI, and production systems from ideation to scale. I continue as a Research Associate at Stanford University, mentoring applied ML projects in remote sensing, disease ecology, and climate domains.

Earlier I shipped production ML and tooling at Google Cloud (Kubeflow components), QuantumScape, Driscoll’s, Monterey Bay Aquarium, and contributed image-processing work used by the NASA Cassini mission. I created the open-source AutoDC project (NeurIPS 2021) and have 8+ peer-reviewed applied ML publications with presentations at NeurIPS, ICLR, and related venues.

My previous research life was in geophysics, earthquake dynamics, and planetary sciences. I built numerical models to simulate ice mountain formation on Saturn's moon Titan and climate patterns inferred from sand dunes on Mars, as well as tsunami propagation caused by mega-earthquakes.

Timeline

07/2026-now: Co-Founder @ Stealth AI Startup

05/2025-now: Senior Quantitative Scientist (NLP/LLM) @ Verana Health

11/2017-now: Research Associate – AI/ML @ Stanford University

04/2025-now: Technical Advisor @ Rudiment Medicine

12/2024-06/2026: Co-Founder @ Abide AI

05/2023-05/2025: Venture Fellow @ Laconia Capital

03/2023-11/2024: Founding ML Engineer @ BrainChain AI

05/2020-03/2023: Senior Data Scientist/ Chief Data Scientist @ A360 AI | Hypergiant

06/2019-05/2020: Data Scientist @ Driscoll's

10/2018-06/2019: ML Engineer @ Google Cloud

05/2017-02/2019: Data Scientist @ Monterey Bay Aquarium

08/2013-06/2016: Research Scientist @ Smithsonian | NASA joint contract

08/2010-08/2014: Researcher @ NASA Cassini RADAR Team

Talks

Press Coverage

ML Research

Github and Google Scholar.

Current focus: Advancing and validating a neuro-symbolic causal reasoning harness that sits on top of existing models. Causal Reflection constructs and maintains dynamic causal world models, performs structured self-reflection, and plans efficient action sequences so open-source and mid-tier models can close the gap with frontier models on agentic workloads. Near-term emphasis is strong results on public agentic benchmarks (SWE-bench, TAU-bench, Terminal-bench, and related suites) that map closely to enterprise software-engineering, tool-use, and multi-step decision tasks. Foundational research appears in “Causal Reflection with Language Models” (NeurIPS 2025).

Recent/ Past Works

I built an open source data-centric AI tooling, AutoDC (Automated data-centric processing) that aims to automate dataset improvement process. We tested the framework on image data and AutoDC is estimated to reduce roughly 80% of the manual time for data improvement tasks, at the same time, improve the model accuracy by 10-15% with the fixed ML code. We want to expand to NLP and tabular data, open for contributors; check our NeurIPS paper.

Enterprise product development

ML pipeline/ Kubeflow component contributions

Applied ML | Data science

I led DS teams to build custom ML solutions and data science use cases to support industry partners and clients as well as academic institutions.

Industry

Domain supported: commercial space, CRM, manufacturing sensory and imagery, supply chain operation, oil and gas, sustainable agriculture.

Academia

Domain supported: remote sensing, marine biology and conservation, disease ecology, deforestation/ climate change, and geophysics/ planetary science.

Peer-reviewed Publications

ML (Applied ML, AI Research)

  1. Aryan, A. Liu, Z.Y.-C. Causal Reflection with Language Models. NeurIPS (2025), arXiv:2508.04495.
  2. Chamberlin, A.J., Liu, Z.Y.-C., Cross, C.G.L., Pourtois, J. et al. High-Resolution Canopy Height Mapping in Tropical Rainforests Using Deep Learning and Multi-Source Remote Sensing Data. Remote Sensing (2025), doi: 10.3390/rs17213592.
  3. Jerette, J., Liu, Z.Y.-C., Chimote, P., Hastie, T., Fox, E., and Ferretti, F. Shark detection and classification with machine learning. Ecological Informatics (2022), doi: 10.1016/j.ecoinf.2022.101473.
  4. Liu, Z.Y.-C., Chamberlain, A.J., Tallam, K., Jones, I.J., Lamore, L.L., Bauer, J. et al. Deep Learning Segmentation of Satellite Imagery Identifies Aquatic Vegetation Associated with Snail Intermediate Hosts of Schistosomiasis in Senegal, Africa. Remote Sensing, SI: Remote Sensing and Infectious Diseases (2022), 10.3390/rs14061345b.
  5. Liu, Z.Y.-C., Roychowdhury, S., Tarlow, S., Nair, A., Badhe, S., and Shah, T. AutoDC: Automated data-centric processing. NeurIPS (2021), arXiv:2111.12.
  6. Liu, Z.Y.-C., Tarlow, S., Akbar, M., Donnellan, Q., and Senkow, D. Improved orbital propagator integrated with SGP4 and machine learning. SmallSat 2021, paper link.
  7. Tallam, K., Liu, Z.Y.-C., Chamberlin, A.J., Jones, I.J., Shome, P., Riveau, G., Ndione, R.A. et al. Identification of Snails and Schistosoma of Medical Importance via Convolutional Neural Networks: A Proof-of-Concept Application for Human Schistosomiasis. Frontiers in Public Health (2021): 900, doi: 10.3389/fpubh.2021.642895.
  8. Liu, Z.Y.-C., Moxley, J.H., Kanive, P., Gleiss, A.C., Maughan, M., Bird, L., Jewell, O. et al. Deep learning accurately predicts white shark locomotor activity from depth data. Animal Biotelemetry 7, no. 1 (2019): 1-13, doi: 10.1186/s40317-019-0175-5.

Science (Geophysics, Planetary Science, Climate Change)

  1. Hopkins, S.R., Hazel, A., Pourtois, J.D., Chamberlin, A.J., Gajewski, Z., Harryman, I., Sokolow, S.H., MacDonald, A.J., Nova, N., Ahmad, A., Andiani, J., Emerson, A., Febriani, N., Finley, N.L., Izza, Q., Miller, A., Sartika, I., Webb, K., Burza, S., Siregar, I.Z., Setiawan, Y., Jones, I., Liu, Z.Y.-C., and De Leo, G.A. Pandemic impacts on protected rainforests and rural communities with variable health and livelihood support in West Kalimantan, Indonesia: a mixed-methods approach combining household surveys and remote sensing. Lancet Planet Health (2026), doi: 10.1016/j.lanplh.2026.101455.
  2. Jones, I.J., MacDonald, A.J., Hopkins, S.R., Lund, A.J., Liu, Z.Y.C., Fawzi, N.I., Purba, M.P., Fankhauser, K., Chamberlin, A.J., Nirmala, M. and Blundell, A.G. Improving rural health care reduces illegal logging and conserves carbon in a tropical forest. Proceedings of the National Academy of Sciences (2022), 117(45), pp.28515-28524, doi.org/10.1073/pnas.2009240117.
  3. Liu, Z.Y.-C., Radebaugh J., Harris, R., Christiansen, E.H, and Rupper, S. Role of fluids in the tectonic evolution of Titan. Icarus (2016) (SI: Titan's Surface and Atmosphere), 270, 2-13, doi: 10.1016/j.icarus.201_6.02.016.
  4. Radebaugh, J., Ventra, D., Lorenz, R.D., Farr, T. Kirk, R.L., Hayes, A., Malaska, M., Birch, S., Liu, Z.Y.-C., Lunine, J., Barnes, J., Le Gall, A., Lopes, R.M.C., Stofan, E., Wall, S., Paillou, P., and Wood, C.A. Alluvial and fluvial fans on Saturn’s moon Titan reveal processes, materials and regional geology. In, Ventra, D. & Clarke, L. E. (eds) Geology and Geomorphology of Alluvial and Fluvial Fans: Terrestrial and Planetary Perspectives. Geological Society, London, Special Publications (2016), 440, doi: 10.1144/SP440.6.
  5. Liu, Z.Y.-C., Radebaugh J., Harris, R., Christiansen, E.H, Kirk, R.L., Neish, C.D.,Lorenz, R.D., and the Cassini Radar Team. The tectonics of Titan: Global structural mapping from Cassini radar. Icarus (2016) (SI: Titan's Surface and Atmosphere), 270, 14-29, doi: 10.1016/j.icarus.2015.11.021.
  6. Liu, Z.Y.-C. and Zimbelman, J.R. Recent near-surface wind directions inferred from mapping sand ripples on Martian dunes. Icarus (2015), 261, 169-181, doi: 10.1016/j.icarus.2015.08.022.
  7. Liu, Z.Y.-C. and Hargitai H. Mountain (Titan). (2014) In: A. Kereszturi and H. Hargitai (Eds), Encyclopedia of Planetary Landforms, Springer-Verlag Berlin Heidelberg, Germany, doi: 10.1007/978-1-4614-9213-9_508-1.
  8. Liu, Z.Y.-C. and Harris, R. Discovery of Possible Mega-Thrust Earthquake along the Seram Trough from Records of 1629 Tsunami in Eastern Indonesian Region. Natural Hazards (2013) (SI: Extreme Geohazards), 72(3), 1311-1328, doi: 10.1007/s11069-013-0597-y.

Get In Touch

Building in stealth. Open to conversations with early-stage investors, design partners, and potential customers interested in agentic systems, reasoning infrastructure, and capital-efficient approaches to robust AI. I’m best reached via email (zacqoo at gmail dot com).