LLM systems | evaluation | post-training | retrieval
I build reproducible infrastructure for model behavior: benchmark runners, verifier-guided inference, retrieval diagnostics, traceable agent workflows, and upstream model-system fixes.
SFT/DPO-style flows, verifier checks, and executable benchmark exports.
Query planning, source trust, citation grounding, and regression tests.
Tool traces, validators, retry paths, reward labels, and escalation boundaries.
Single-GPU training, quantization, runtime/cache correctness, and reproducible runs.
Selected systems
The strongest projects expose the full path from setup to artifact: commands, benchmark inputs, result files, known boundaries, and the exact claim each run supports.
Upstream OSS | compiler/runtime
Merged PRMerged upstream runtime cache-group integrity fix that treats incomplete cache groups as misses, with unit coverage and full Triton integration CI across NVIDIA, AMD, and macOS runners.
Post-training | verifier-guided inference
RepositorySingle-L20 post-training and verifier-guided inference stack for executable code benchmarks, with reproduction scripts, artifact hashes, and explicit claim boundaries.
Fine-tuning | evaluation
RepositoryText-to-SQL fine-tuning and multi-path inference with Qwen2.5-Coder-7B.
Reranking | FinanceMTEB
RepositoryChinese finance reranking snapshot with Qwen3-Reranker-8B and public comparison context.
Retrieval | agent systems
RepositorySearch and retrieval workbench with query planning, source-trust tiers, and citation checks.
AI4S infra | scientific agents
RepositoryTrace, validation, and reward infrastructure for scientific agents.
Retrieval benchmark | provenance
RepositoryCoREB retrieval benchmark snapshot with CI-backed artifacts and result provenance.
Evidence model
The page is optimized for reviewers who scan quickly: what problem is solved, what can be reproduced, what was measured, and where the claim stops.
No invisible benchmark setup. Inputs should be inspectable before running anything.
Scripts, Make targets, and CI paths make evaluation repeatable after code changes.
Reported numbers should point to hardware notes, data versions, and known failure modes.
Tool calls, citations, validators, retries, and escalation paths are stored as system data.
Stack map
PyTorch, Transformers, LoRA/QLoRA, vLLM, lm-eval, Triton runtime, C++/CUDA
Golden sets, regression tests, semantic judges, citation checks, benchmark reports
RAG, text-to-SQL, tool use, structured generation, scientific workflows
FastAPI, TypeScript, Docker, GitHub Actions, SQLite, PostGIS, Redis, Kafka
Repository matrix
Contact