THE RESEARCH NOTEBOOK

Ask carefully.
Measure honestly.

My research spans efficient neural retrieval and machine learning for financial disclosures. Each question becomes a reproducible experiment, with the methods, measurements, and current findings available to inspect.

Traceable inputsReproducible methodsExplicit limits
STUDY EXPLORER
UNTOUCHED CONFIRMATION23.4%

fewer cross-encoder calls

CROSS-ENCODER CALLS · 1,500 QUERIES

Always rerank1,500
Conditional router1,149

351 fewer calls on untouched confirmation queries.

Quality loss tolerance0.01 nDCG@10
Open the methods & evidence
ToolRet logo
RETRIEVAL / 2026

ToolRet

When does reranking earn its cost?

A learned router chooses when to run an expensive cross-encoder over a 37,292-tool catalog. Development, calibration, and confirmation are separated by relevant-tool components.

Undergraduate research at Cornell Engineering. This is an empirical study, not a peer-reviewed publication.

UNTOUCHED CONFIRMATION23.4%

fewer cross-encoder calls

CROSS-ENCODER CALLS · 1,500 QUERIES

Always rerank1,500
Conditional router1,149

351 fewer calls on untouched confirmation queries.

Predeclared tolerance: 0.01 nDCG@10 loss. Quality superiority was not established.

QUESTION

Is every rerank useful?

Compare always reranking with a conditional decision that uses first-stage retrieval signals.

DESIGN

Freeze before confirmation.

Separate relevant-tool components across development, calibration, and 1,500 untouched confirmation queries.

EVIDENCE

Keep the result inspectable.

Retain configurations, matched-budget random controls across 20 seeds, the report, and an independent audit in the public repository.

ML & QUANTITATIVE RESEARCH / IN PROGRESS

Disclosure Drift

What changes when a company rewrites its risks?

I am studying whether changes in annual SEC 10-K risk disclosures help explain subsequent 20-session stock returns relative to the market. The pipeline pairs filings, extracts text features, and audits the historical identifiers and return labels needed for a valid experiment.

The pipeline spans 22.6M+ daily equity records and a text pilot of 215 SEC 10-K filings, forming 180 annual pairs across 35 issuers. Source fingerprints connect each observation to its inputs.

The earlier October 11 checkpoint retained an unopened 2022–2025 final evaluation. The current pipeline work strengthens alignment and numerical validation; forecasting performance has not yet been established.

COMPLETED NLP EXECUTION1.46M

tokens encoded with MiniLM

A separate semantic analysis processed 155 filings and 135 annual pairs in 5,828 chunks, with zero truncation. Two independent Python processes reproduced all document vectors and pair features exactly.

Inspect the MiniLM execution record
TEXT FEATURES

Measure the change.

TF-IDF/cosine similarity, Jaccard distance, and document-length changes form the primary features. The pinned MiniLM embeddings are a retrospective secondary analysis, kept separate from the primary model.

DATA ENGINEERING

Trace every observation.

Use Python, SQL, DuckDB, and scikit-learn to preserve source fingerprints, original filing acceptance times, and historical identity evidence. Leakage-resistant temporal alignment and chronological splits keep future information out of earlier observations.

REPRODUCIBILITY

Check the whole pipeline.

Independently reconciled 135 return calculations within 10⁻¹². The earlier reproduction record passed 995 tests and the complete synthetic study; historical data review and empirical forecasting evaluation remain in progress.

Read the verification record
THERE’S MORE TO THE STORY

Where to next?