What I work on
- Agent pipelines → source, fix, test, ship real code
- ML infra and debugger tooling → inference optimization, CUDA scheduling
- Trustworthy AI → federated learning, zero knowledge proofs, explainability
Merged PRs, last 30 days
| Repo | What it was |
|---|---|
| HKUDS/DeepTutor | Create auth secret and token files with owner-only permissions |
| NVIDIA/TensorRT-Model-Connect | fix(cli): unify kv-cache-size byte parser between CLI and benchmark |
| tenstorrent/whisper | Do not let MXR disable pointer masking in M-mode |
| tenstorrent/tt-installer | fix(installer): disable firmware updates by default in container mode |
| microsoft/apm | fix(audit): allow local Claude-skill deps to omit apm.yml |
| NVIDIA/TensorRT-Model-Connect | Parse family import scanning with ast instead of regex |
| NVIDIA/cudf-spark | Remove inert InMemoryTableScanExec allowances in cache_test.py [databricks] |
| griddynamics/rosetta | Remove dead _validate_topic function |
| DeusData/codebase-memory-mcp | test: cover wrong JSON types in the MCP robustness layer |
| bkuan001/halo-record | Fix chain verification propagating a stale self-declared hash after a break |
| HKUDS/Vibe-Trading | fix(onboard): create .env.partial with owner only permissions |
| yzhao062/awesome-auditable-ai | Standards: add the tamper-evident log and attestation primitives |
| diegosouzapw/OmniRoute | chore(quality): add an RTL layout ratchet |
| usestrix/strix | Create credential files with owner-only permissions |
| DeusData/codebase-memory-mcp | fix(graph-ui): rank Accept-Language by q instead of substring match |
Background
- Backend and automation infrastructure for a data-driven marketing platform → real recurring revenue
- Cerebras Fellowship → paper to deployable product on Cerebras hardware
- Deep Learning Research Intern, IIT Roorkee → SegFormer and CFAN work, 82% mIoU
- AI Intern, Golden Bird Healthcare → multimodal clinical documentation system, 82% Dice score
- Zerobase Startup School (top 20, Korea) · Google for Startups x Antler