News archive
A dated record of what I have been working on, newest first. Most of it is one project: a critical replication of a published retrieval-augmented mammography report-generation pipeline, and the controlled experiment it turned into.
-
Separation Without Transfer was accepted at the Pacific Symposium on Biocomputing 2027 — for publication in the proceedings, and selected for an oral presentation at the meeting on the Big Island of Hawaii, 3–7 January 2027 — with Dr. Wenjing Yang as co-author. The camera-ready revision is due 1 October 2026 and the manuscript is not public until it is submitted; the proceedings are indexed in PubMed.
-
Wrote an essay for new graduates, “AI Didn’t Take Your Job. It Created One Nobody Is Trained For,” arguing that the slow entry-level market is a skills mismatch rather than a disappearance, and that experimental design and causal inference are the scarce complement to AI work in high-stakes domains. It uses the mammography experiment below as its worked example.
-
Submitted Separation Without Transfer: The Retrieval Encoder's Substrate, Not the VLM Decoder, Decides Whether Contrastive Fine-Tuning Transfers in Mammography Report Generation to the Pacific Symposium on Biocomputing (PSB) 2027, with Dr. Wenjing Yang as co-author. It was accepted the following month — see above.
-
“Will AI Make Human Work Worthless — or Priceless?” won the Graduate Division of the CAUSE Student Data Scrollytelling Contest, run by the Consortium for the Advancement of Undergraduate Statistics Education. It is a nine-act interactive data story in D3.js and Scrollama.js, built around an original nested-CES general-equilibrium model of AI and cognitive labor and written for readers who do not work in economics. I submitted it in June, at the end of a spring quarter that also carried two research positions and three courses.
-
Before submitting, I re-derived every number in the paper directly from the raw prediction files and stored embeddings, and assembled a reproducibility package — checksummed input data, an end-to-end rebuild script, and the executed notebooks — so each figure and table can be regenerated from source.
-
Found the geometric signature behind the encoder effect: the same fine-tuning recipe collapses the held-out effective rank of the three out-of-domain encoders' embeddings (10.5 to 1.1, 9.1 to 2.4, 20.2 to 2.5) while expanding the mammography-native encoder's from 7.2 to 13.3. Collapse versus expansion separates the encoders whose training-set class separation carries over to held-out data from those where it does not.
-
Tested whether the effect was specific to the label the encoder was fine-tuned on by scoring the untrained, finer five-class BI-RADS category instead of the three-class suspicion label. The gain held on both retrieval-sensitive decoders (accuracy +23.2% and +14.6%; quadratic weighted kappa 0.100 to 0.302), though nearly all of it landed on the majority BI-RADS categories rather than the clinically decisive ones.
-
Added a fourth encoder, BioCLIP (organism / tree-of-life images), which shares the other out-of-domain encoders' architecture and fine-tuning schedule but not their pretraining domain. That completed the design: four retrieval-encoder substrates, each evaluated frozen and fine-tuned, crossed with three frozen VLM decoders — 24 evaluation cells. Began drafting the paper in the PSB conference format.
-
Confirmed the headline result. Fine-tuning the mammography-native encoder (Mammo-CLIP) produced the only arm to clear the held-out majority-class retrieval floor of P@1 = 0.487, moving from 0.513 frozen to 0.585 fine-tuned; the same recipe applied to the generic web-image and biomedical-figure encoders inflated training-set separation but left held-out retrieval tied with that floor. The retrieval gain propagated into the generated reports on the two retrieval-sensitive decoders (suspicion accuracy +20.7% and +15.6% relative to the frozen baseline), while MedGemma did not respond to it. Wrote up the full analysis in a 37-page technical report for my advisor, Dr. Wenjing Yang.
-
Audited my evaluation code against the paper I was replicating and found that its published precision, recall and F1 are computed as weighted averages while its text describes them as macro-averaged — the two are not the same metric. My own evaluator follows the paper's stated methodology, so every subsequent comparison is run within my own pipeline rather than against its published numbers directly.
-
Fine-tuned the retrieval encoder for the first time and found a warning sign: its training-set class separation grew far more than its held-out separation. That overfitting signal is what motivated testing the same fine-tuning recipe on encoders pretrained on different image domains.
-
Completed the retrieval-augmented (RAG few-shot) evaluation across all three VLM decoders on the 1,000-study VinDr-Mammo held-out test split, using the retrieval encoder from the paper I was replicating.
-
Built the VinDr-Mammo preprocessing and retrieval pipeline — each of the four views resized to 512×512 and merged into a single composite, then BI-RADS mapped to a three-class suspicion label — matching the replicated paper's stated methodology exactly, including correcting places where its released code did not follow its own description.
-
On a second research position, with Dr. Brian Fischer, designed a 7-table normalized MySQL schema (owl, recording session, neuron, experiment, tuning parameters, spike reliability, ILD conditions tested) for his lab's barn owl auditory-neuroscience recordings — roughly 14,000 raw files in six proprietary formats, 110 neurons across 8 owls, about 261 experiments — and built the 6-phase ETL pipeline that loads it. An analysis loop that previously required 228 individual file loads now runs as a single SQL query.
-
Began a critical replication of a published retrieval-augmented mammography report-generation pipeline: explored the CBIS-DDSM dataset first, then switched to VinDr-Mammo (5,000 studies) to match the paper's own dataset, and set up evaluation of three vision-language model decoders (MedGemma, Qwen2.5-VL, LLaVA-Med) across several prompting conditions using the paper's own frozen retrieval encoder.
-
Joined Dr. Brian Fischer's lab as a research assistant, working on a database for barn owl sound-localization recordings.
-
Started a literature review on retrieval-augmented generation for vision-language systems, the starting point for the mammography replication project.
-
Built and shipped a retrieval-augmented “second brain” system for this site's chatbot.
A note on the clinical labels: VinDr-Mammo carries radiologist BI-RADS assessments, not biopsy-confirmed pathology. The suspicion and BI-RADS results above are agreement with a radiologist's reading, not diagnostic accuracy.