Core contributor to Microsoft Agent Framework, ByteDance’s OpenViking, and Tencent’s ncnn.
I build reliability into the systems that agents depend on, from workflow recovery and persistent memory to inference on edge devices.
My research focus is embodied AI and world models. How should an agent change its next move when the world differs from its prediction?
| Project and role | What changed upstream |
|---|---|
| Microsoft Agent Framework · Core contributor | Workflow recovery. I changed thread snapshots to retain a human reply so a workflow can resume with the answer already given. Contribution record · Historical PR #8273 |
| ByteDance’s OpenViking · Core contributor | Agent memory. I coauthored the LangChain fix that lets a restored conversation retain what was said, who said it, and which tools were called. PR #6275, opened by @ehz0ah · Coauthor credit |
| Tencent’s ncnn · Core contributor | Inference reliability. I stabilized LogSumExp on CPU and Vulkan. Both backends now return a finite result for inputs like [1000, 1000]. PR #7055 · Numerical example |
Of the 113 merged PRs across 35 upstream projects in my November 14, 2026 record, 72 are in the three projects above. I opened 99 of the 113 and am credited as a coauthor on the other 14.
- @ehz0ah brought our LangChain fixes together and opened the OpenViking PR. How we preserved conversation history.
- In Tencent ncnn, @nihui and I share credit on three merged PRs. In one, we changed how the pnnx converter stores boolean constants. PR #7006 · Coauthor credit.
View 14 coauthored PRs
- volcengine/OpenViking#7040 — fix(langchain): preserve Unicode archive query terms · PR by @ehz0ah · coauthor credit
- probabilists/zuko#83 — Include broadcast bounds in Truncated batch shape · PR by @francois-rozet · coauthor credit
- volcengine/OpenViking#6479 — fix(langchain): isolate session commits and report reset failures · PR by @ehz0ah · coauthor credit
- volcengine/OpenViking#6404 — fix(langchain): honor nested filters and index field selectors · PR by @ehz0ah · coauthor credit
- volcengine/OpenViking#6342 — fix(langchain): preserve Store reads and namespace discovery · PR by @ehz0ah · coauthor credit
- volcengine/OpenViking#6338 — fix(langchain): preserve recall inputs and context metadata · PR by @ehz0ah · coauthor credit
- volcengine/OpenViking#6324 — fix(langchain): align Store deletion, validation and removal errors · PR by @ehz0ah · coauthor credit
- volcengine/OpenViking#6275 — fix(langchain): preserve conversation text, roles and tool history · PR by @ehz0ah · coauthor credit
- microsoft/agent-framework#8284 — Python: add public TypedDict for AgentExecutor checkpoint state (#8201) · PR by @FOWEPJF255 · coauthor credit
- microsoft/agent-framework#8163 — Python: include checkpoint_id on AG-UI interrupt metadata (#8150) · PR by @FOWEPJF255 · coauthor credit
- Tencent/ncnn#7006 — pnnx: store bool MemoryData attributes as raw int8 · PR by @nihui · coauthor credit
- Tencent/ncnn#7003 — layer param hardening · PR by @nihui · coauthor credit
- Tencent/ncnn#6989 — harden net loading · PR by @nihui · coauthor credit
- volcengine/OpenViking#4692 — fix(pi-extension): replay offline backlog through the batch endpoint · PR by @ZaynJarvis · coauthor credit
A robot can finish a grasp motion while the object stays on the table. That gap is the problem I’m working on in embodied AI. A world model predicts what an action will change. The next observation gives the agent evidence to continue, retry, or choose a different action.
- handoff-pack carries context between models or sessions. It prioritizes instructions, tool errors, and recent messages, then writes a compact Markdown or JSON handoff. Example conversation.
- traj-wash prepares agent trajectories for evaluation and supervised fine-tuning. It keeps the first failed step as evidence and trims the retries that follow. In the seven-step example, four steps remain.
Full contribution record and more project cases
Browse publicly searchable authored PRs.
🤖 Microsoft Agent Framework · Agent workflows
Agent workflows · Recovery · Checkpoints
Agent orchestration, recoverable workflows and checkpoints. In the historical snapshot: 30 authored and 2 coauthored PRs.
- Recoverable human-in-the-loop workflows: Preserved user text in thread snapshots. Microsoft Agent Framework #8273.
- Fixed BreakLoop handling in declarative workflows.
The repository and historical PR links were unavailable when checked on November 18, 2026; the figures are from earlier verified records.
🧠 OpenViking · Agent memory
Agent memory · Context · Integrations
Memory, context and agent integrations. In the historical snapshot: 22 authored and 8 coauthored PRs.
- Fixed CR-delimited MCP event stream parsing.
- Windows CLI discovery: The launcher could miss an existing development executable; the fix selects the Windows filename while preserving the existing lookup order. OpenViking #7251 · Case study.
⚙️ Tencent ncnn · Inference
Inference · Correctness · Portability
Inference correctness, portability and bindings. In the historical snapshot: 7 authored and 3 coauthored PRs.
-
Numerically stable inference: Extreme finite inputs could produce infinities; the fix stabilizes LogSumExp on CPU and Vulkan. ncnn #7055 · Case study.
-
Addressed FP16 GELU overflow on ARM and RISC-V.
-
Tencent ncnn: Improved bool MemoryData handling in pnnx, with @nihui · coauthor credit.
-
Zuko: Included broadcast bounds in truncated batch shapes, with @francois-rozet · coauthor credit.
More selected upstream work
- Agent integrations: Handle CR-delimited MCP event streams and support bracketed IPv6 bind addresses in the OpenWebUI integration.
- Durable workflows: Persist HITL user text in thread snapshots and honor BreakLoop in declarative workflows.
- Inference runtimes: Correct even-sized LRN channel windows and retain Python custom-layer factories for their required lifetime.
- Python infrastructure: Reject readonly buffers before consuming buffered data in fsspec and preserve defaultdict batch conversion in PyTorch Ignite.
- Numerical correctness: Fix complex-weight CP tensor norms in TensorLy — walkthrough and minimal reproduction. Prevent overflow when summing probability-selection fitness in PyGAD.
My contribution loop: reproduce the failure → preserve normal behavior → add a focused regression → close the review loop.
Further merged work includes TensorLy, PydanticAI, Qdrant Client, sklearn-onnx, RapidOCR, pyts, scikit-lego and Patsy.


