Deevna
Reddy

PROBLEM

Claims teams drown in unstructured PDFs, scans, and emails containing sensitive policyholder PII. Every claim needed a human just to redact records and figure out what it said before anyone could decide anything.

APPROACH

An agent pipeline (Ingest & Mask PII → Extract → Validate → Route → Queue) sanitizes sensitive policyholder records, pulls 15+ structured fields out of raw documents, then runs an LLM driven validation pass before routing, so low confidence extractions never slip through silently.

OUTCOME

80% of first pass triage now runs without a human touching it, with full PII compliance. The other 20% get flagged with the specific field that failed validation, not just a generic 'needs review' tag.

WHAT I'D CHANGE

The validation step catches malformed fields, not ambiguous ones: a claim that is technically well formed but contextually wrong still passes. I would add a second cross field consistency check before scaling past pilot volume.

PROBLEM

A plain logistic regression baseline flagged obvious fraud fine, but missed behavioral patterns: the same account making small, spaced out, plausible looking transactions that only look wrong in aggregate.

APPROACH

Pattern recognition on transaction features, layered with a behavioral anomaly model that scores sequences rather than single transactions. Served via FastAPI so a flagged transaction shows up on the risk dashboard within a second, not the next morning's batch job.

OUTCOME

Precision recall AUC improved 25% over the logistic regression baseline, with the biggest gains on exactly the slow burn fraud patterns the baseline missed.

WHAT I'D CHANGE

The behavioral model needs a longer transaction history to be confident, since new accounts get scored on thinner evidence. I would add a separate cold start tier instead of forcing new accounts through the same model.

PROBLEM

Lung cancer AI tools typically use one data type and give a bare label. Clinicians cannot act on a black box call, especially on small, ambiguous nodules where a biopsy is invasive and a miss is costly.

APPROACH

A dual branch model: DenseNet121 + Grad CAM reads CT scans (LIDC IDRI, ~53K scans) for visual explanations, while a Random Forest + SHAP branch reads TCGA RNA seq gene expression (~798 samples) for molecular explanations. Late fusion combines both into one classifier predicting LUAD, LUSC, or Normal.

OUTCOME

The fused model hit 98.32% accuracy and 0.99 AUC, beating either modality alone (90.15% imaging only, 95.47% genomics only). SHAP flagged EGFR, TP63, and NKX2.1 as top drivers, consistent with known cancer pathways. First author paper, IEEE ICCDS 2025.

WHAT I'D CHANGE

Validation was retrospective on public datasets (LIDC IDRI, TCGA), not real clinical cases. I would want prospective validation with a hospital partner before trusting this near an actual diagnosis.

PROBLEM

Traditional ASD screening models often underrepresent minority ethnic cohorts, leading to delayed diagnoses and disparate screening outcomes during early childhood development.

APPROACH

Evaluated demographic and behavioral screening indicators across ethnic cohorts using statistical hypothesis testing and supervised classification to measure disparate impact.

OUTCOME

Quantified significant variance in screening markers across ethnic groups, establishing benchmarks for equitable pediatric diagnostic frameworks. Published in Springer Nature.

WHAT I'D CHANGE

I would expand the longitudinal cohort data to track post intervention developmental outcomes across multiple clinical hospital networks over a multi year timeframe.