Learning, Building, Researching

Research on LLM agents in resource-constrained environments, distributing capabilities across fine-tuning, retrieval, read-only tools, and gateway policies. Integrated QLoRA / SimPO training and vLLM serving with a 1,420-chunk project knowledge base built using Qwen3 Embedding and Reranker, six read-only tools for runtime inspection, and an authenticated OpenAI-compatible gateway supporting SSE, tool calling, and safety checks before generation.

Evaluated four ablation settings on a 250-question benchmark under matched configurations. RAG with tools improved ARS from 0.3720 to 0.6845, with tool selection and execution accuracy both reaching 1.00; the gateway reduced hard failures from 9 to 0.

Reproduced the NeuroAdapter pipeline for NSD subject 1, including data conversion, Schaefer parcel labeling, brain-region selection, conditional training, and Stable Diffusion image decoding. Developed single-GPU and 2/4-GPU DDP training, resume, and checkpoint-management scripts, and completed the 50k-to-100k-step training stage on four A40 GPUs. Implemented multi-candidate generation and selection using brain-response correlations from whole_brain_encoder.

Adapted outputs to the authors' evaluation format and evaluated PixCorr, SSIM, AlexNet, Inception, and CLIP metrics. With a fixed seed, the 100k-step checkpoint achieved 89.22% CLIP and 85.14% Inception two-way identification. Investigated resume instability through data-alignment checks, learning-rate and global-batch comparisons, and optimizer-state ablations; identified a mismatch between the public ROI mapping and the paper's 50/53/103 functional region groups as a reproduction limitation.

Fine-tuned Qwen2.5-Coder-7B with QLoRA on approximately 10,000 OpenCodeInstruct and CommitPackFT examples, applying license filtering, deduplication, dataset splitting, and indentation-preserving cleaning. Compared standard cleaning, indentation preservation, interface-contract augmentation, and unit-test feedback with a fixed base model, QLoRA configuration, and effective batch size. Built a deterministic generation and isolated testing pipeline that attributes failures to syntax, structure, interface contracts, semantics, and boundary conditions.

Across 73 valid evaluation tasks, indentation-preserving cleaning raised the pass rate from 20.55% to 84.93% and eliminated 55 structural errors; contract augmentation further improved it to 87.67%. Retained the negative ablation result: feedback augmentation reduced the pass rate to 83.56%, with three interface-contract regressions.

Audited the public artifacts of the NeurIPS 2024 SWE-agent paper and conducted budget-constrained Agent-Computer Interface (ACI) experiments with a modern model. Completed the audit of 54 paper outputs and all 828 preregistered experimental runs across 23 SWE-bench Lite development instances, nine ACI configurations, and four repetitions. Used paired exact McNemar tests, Holm correction, and 10,000 bootstrap resamples clustered by instance.

Of the 828 runs, 285 fully resolved their tasks (34.4%); the default ACI resolved 34 of 92 runs (37.0%). None of the eight variants differed significantly from the default after Holm correction at alpha = 0.05. The study covers public-artifact auditing and a bounded modern-model reproduction, rather than an exact rerun with the paper's retired models. With only 23 independent instances, nonsignificant results do not establish equivalence between ACI designs.

A cloud-first autonomous agent with a persistent core, shared persona and memory, QQ and Telegram adapters, a management dashboard, and a human-approved GitHub self-improvement workflow.

My current interests span deep learning, computer vision, LLM-based agents, and the engineering infrastructure needed to turn research ideas into reliable systems.

GitHub, X, Telegram, Instagram, and email are linked in the icon row above. The RSS feed is linked in the footer.