Projects & Reports

Synthetic digital-banking analytics from lifecycle metrics and causal estimates to churn ranking and contact economics.

• Built a deterministic 100,000-customer evidence layer in DuckDB with explicit activation, retention, campaign, churn, monitoring, and economics contracts.
• Measured a +9.16 percentage-point randomized activation effect, then showed that every simulated strategy that contacted customers lost value under the fictional cost assumptions.

Local-first Markdown reading and editing for the browser and macOS, with explicit privacy and file safety.

• Shipped a browser release with rich Markdown rendering, local drafts, save fallbacks, and production-verified privacy controls.
• Extended the shared core into an Apple Silicon macOS beta with native file handling and unsaved-change protection.

Offline MEA spike detection, burst analysis, waveform metrics, and electrode comparisons.

• Built a local Streamlit workflow for continuous recordings and pre-sorted NeuroExplorer spike data, covering spike, waveform, ISI, firing-rate, and burst analysis.
• Added deterministic regression fixtures for burst algorithms, filtered detection, waveform boundaries, and EDF calibration; these checks validate implementation behaviour, not biological validity.

Citation-grounded research across financial documents using hybrid retrieval and reranking.

• Combined BM25 and vector retrieval with reciprocal-rank fusion, company-balanced retrieval, cross-encoder reranking, and page-resolved citations.
• Recorded 19/19 evaluation checks on three real SEC filings and 290 passing backend tests, while documenting live-use failures and remaining retrieval limitations.

Multimodal VQA fusion benchmarking with ViT, BERT, co-attention, and LoRA.

• Compared eight frozen-CLIP fusion heads; the 657K-parameter GeometryFusion model led the track at 63.3% validation accuracy.
• Trained an end-to-end ViT-B/16 and BERT co-attention model to a 67.6% VQA soft score after fixing soft-label, masking, pooling, and tokenizer issues.

Venue-calibrated ML paper evaluation built from historical OpenReview evidence.

• Built a leakage-safe, forum-level evaluation workflow with closed model inputs, private labels, disjoint calibration sets, and hash validation.
• Across descriptive ICLR pilots, calibration reduced rating MAE by 21.8%, increased decision accuracy from 62.9% to 77.1%, and reduced false accepts from 9 to 2.