Available now · US work authorized (F-1 STEM OPT)

I check the number
I'm hoping not to see first.

Data Scientist and Data Analyst with an M.S. from Boston University. I work on healthcare, quantitative finance and web-scale data — and I'd rather hand you a result that survives a follow-up question than one that looks good on a slide.

Data Scientist Data Analyst BI / Power BI Analyst ML Engineer
M.S. Applied Data Analytics · Boston University · GPA 3.75/4.0 · graduated May 2026
First-authored paper accepted · EAI CSECS 2026 · Springer LNICST
Boston, MA · open to any US location and remote
hover a value →
Every character here is a real number from this work.
How I work

I built a readmission model. Then I checked the AUC: 0.5654.

That's barely better than a coin flip. The easy move was to report the 71% accuracy instead — which sounds respectable until you notice that predicting "nobody gets readmitted" scores higher.

So I stopped tuning the model and asked a different question: what would a hospital actually do with a prediction? That reframed everything. I built a threshold-tunable classifier a care team can dial toward high recall when a missed readmission is costly, added SHAP so clinicians can see the reasoning, and modelled the cost-benefit at each operating point.

The dataset had weak signal. The decision framework was the real deliverable. I'd rather tell you that than quote a number that falls apart in the interview.

records30,000
class imbalance7.17 : 1
algorithms benchmarked10
accuracy (the flattering one)71%
AUC (the honest one)0.5654
what shippeddecision framework
projected saving / hospital$1.77M
Projected from a cost-benefit model at the chosen operating point — not realized savings. The distinction matters.
60M
revision events processed on GCP Dataproc
1.16M
editors in the WikiFlow cohort
4.5M+
transactional records cleaned in industry
3+1
papers published or accepted, plus 1 under review
3rd
of 40+ teams, BU MedAI Hackathon
Selected work

Every project starts with a question worth answering

Eleven projects. Real datasets, stated methods, and results reported the way I'd defend them in an interview — including the ones that did not work.

Accepted · EAI CSECS 2026
Q: Can you predict who stops contributing to Wikipedia — from their first ten edits alone?

WikiFlow — distributed editor-retention pipeline

A fully cloud-native pipeline over 60M revision events (1,158,248 editors, 25 GB) on a 1+4 node GCP Dataproc cluster. 13 leakage-free behavioural features, selected from 40+ candidates under strict no-leakage constraints. From-scratch RDD logistic regression against an Elephas-distributed Keras DNN. Runs end to end in under two hours with no local data movement.

AUC 0.909 DNNvs 0.845 baselineFirst author PySparkGCP DataprocElephas/KerasBigQuery
Q: Does an intraday edge on the dollar index survive real transaction costs?

DXY & FX intraday strategy research

24 sub-period strategies across the U.S. Dollar Index and its six constituent FX pairs over 3,932 trading days, decomposing each day into a close-to-open leg and an open-to-close leg. The two move in opposite directions across the entire sample. I spent most of the project trying to break the result, because a return that large usually means the test is wrong. It broke. Newey–West passed at t = 3.68, but the Deflated Sharpe Ratio — which corrects for having searched 24 trials — came back at 0.046, and White's Reality Check at p = 0.062. What survived is the decomposition itself, and the overnight leg alone, which clears the same correction at 0.405. The paper is built around that, not the headline.

$100 → $318.80vs $121.72 B&HSharpe 1.06DSR 0.046 — failedbreak-even 1.56 bp PythonstatsmodelsNewey–West HACblock bootstrap

All figures from backtesting, not live trading. First-authored manuscript under review.

Q: What does a hospital do with a prediction that's barely better than chance?

Hospital readmission — decision framework

30,000 discharge records at a 7.17:1 imbalance, ten algorithms benchmarked with 5-fold stratified CV. The honest AUC came out at 0.5654, so the deliverable became the decision logic: a threshold a care team can tune toward recall, SHAP explanations clinicians can read, and cost-benefit modelling at each operating point.

AUC 0.5654 — reported anyway$1.77M projected10 algorithms scikit-learnSHAPSMOTE
Q: Can a star schema and a semantic model turn 100K raw orders into something a business can act on?

RetailPulse — e-commerce analytics

A 9-table, 100K-order dataset modelled into a PostgreSQL star schema, with analytical SQL underneath (window functions, CTEs, cohort retention, RFM segmentation) and a Power BI semantic model with DAX measures on top. The finding that mattered to operations: late deliveries drag review scores from 4.29 to 2.57 stars.

4.29 → 2.57 stars9-table star schema100K orders PostgreSQLPower BIDAXSQL
3rd of 40+ teams
Q: Can a 3D CNN quantify amyloid burden across four different PET tracers?

Amyloid PET Centiloid prediction

Volumetric PET regression for Alzheimer's early screening. A MedicalNet ResNet-34 backbone with FiLM tracer conditioning after each stage, trained with mixed precision, AdamW and Huber loss. I report the test figure, not the friendlier validation one.

Test MAE 12.67 CLr = 0.968val MAE 7.31 PyTorchMedicalNetFiLM
Q: Which health and socioeconomic factors actually predict learning difficulty in children?

Learning difficulty screening — NHIS 2024

7,439 children and 353 features from the National Health Interview Survey, run through 36 models across 9 algorithms and 4 class-balancing strategies. XGBoost with downsampling came out ahead; the feature importances pointed at the factors worth acting on.

36 models9 algorithms7,439 children RcaretXGBoost
Q: How early can you see a jet engine beginning to fail?

Predictive maintenance — NASA C-MAPSS

Run-to-failure turbofan sensor data. I isolated the genuinely degrading sensors with t-tests and ANOVA before modelling, then built regularised classifiers for failure detection and regression for remaining useful life. Only 237 of 12,640 test cycles are actually near-failure, so I quote AUC rather than the 96.5% accuracy that imbalance buys you for free.

AUC 0.94596.5% acc — imbalanced, AUC is the honest oneR² 0.50 RUL RLassoElasticNet
Q: Does Boston's subway split into communities that match how people actually ride it?

MBTA network community detection

Graph analytics on the MBTA subway network built from GTFS data, comparing Louvain, METIS and spectral clustering to find operationally meaningful communities.

Modularity 0.8113265 stations10 communities PythonNetworkX
Q: What actually drives V2X throughput — signal quality, or how the scenario is configured?

V2X wireless performance analysis

13,208 observations (10,252 V2V across 7 scenarios, 2,956 V2I across 3) analysed with t-tests, one-way ANOVA, chi-square with Cramér's V and Tukey HSD, then multiple regression with a polynomial term to adjusted R² 0.8689. Scenario configuration dominated signal quality — the opposite of the intuitive answer.

adj. R² 0.868913,208 obs10,252 V2V + 2,956 V2I RANOVA
Q: Does a road-segmentation model trained on Western benchmarks work on Indian satellite imagery?

Road segmentation for ISRO/SAC

U-Net and DeepLabV3+ (ResNet-50 encoder, Dice loss, IoU) trained in PyTorch on 6,723 aerial images from the Massachusetts Roads and DeepGlobe benchmarks. Roads are a hard segmentation target — they share almost every visual feature with rivers and rail, and they are a thin minority class. The answer to the question was no: the models did not transfer, because every sensor writes its own raster metadata and a model trained at one resolution cannot read another.

Transfer failed — reported6,723 imagesU-Net vs DeepLabV3+ PyTorchQGISRemote sensing

Supervised by Dr. Jai G. Singla, Space Applications Centre, ISRO. The cross-dataset generalization gap has since become a published research line in that group.

Q: Can a model learn to undo a degradation it has never seen?

Meta-learning for blind super-resolution

B.Tech thesis. Standard super-resolution assumes you know how the image was degraded; real low-resolution images do not come with that information. Rather than training one model to reverse one known kernel, I trained initial parameters that adapt fast to whatever degradation a new image turns out to have — first-order MAML to keep second-derivative cost out of the meta-optimization, and MZSR for zero-shot adaptation at inference from a single image.

PSNR 33.513 on SET5SSIM 0.9324non-bicubic kernel PyTorchFOMAMLMZSRGANs

Applying SHARP-MAML to MRDA did not produce considerable gains. I reported that rather than dropping the branch.

Experience

Where the work happened

Research Assistant — Quantitative FX & DXY Strategy Research
Sep 2025 – May 2026
Boston University · Prof. Eugene Pinsky · Boston, MA
  • Sourced, cleaned and validated 3,932 trading days of OHLC data (Dec 2010 – Dec 2025) across the DXY futures contract and its six constituent currency pairs, building the reusable Python modules every result in the paper ran on.
  • Evaluated the complete pre-specified grid — 24 strategies from five elementary position rules, across 7 instruments, at 3 transaction-cost regimes. The best, (Short, Long), turned $100 into $318.80 against $121.72 for buy-and-hold: 7.71% CAGR, Sharpe 1.06, max drawdown −13.0%.
  • Proved the 24-strategy set is not 24 independent experiments. Every rule has an exact negation inside the set, so it partitions into 12 mirror pairs — which halves the effective multiple-testing burden. Then benchmarked against the harder of two admissible conventions ($121.72 rather than $104.76) rather than the flattering one.
  • Built the validation layer in statsmodels: Newey–West HAC (9 lags, Bartlett kernel), Probabilistic and Deflated Sharpe Ratios, stationary block bootstrap at 5,000 resamples, permutation tests and White's Reality Check. The headline result failed at DSR 0.046 and Reality Check p = 0.062. I isolated the one component that did clear correction — the overnight leg alone, at DSR 0.405 — and made that the paper's conclusion.
  • Documented a second, independent finding: the winning strategy changes with the cost assumption (Strategy 7 at 0 bps, Strategy 20 at 1 bp, buy-and-hold at 2 bps), and ranking candidates by gross Sharpe systematically selects the ones least able to absorb their own execution costs.
  • First-authored the resulting manuscript, now under review at an MDPI journal.
Teaching Assistant — Advanced ML & Data Science with Python
Jan 2026 – May 2026
Boston University, Metropolitan College · Boston, MA
  • Supported 60+ graduate students across MET CS 767 and MET CS 577 with model implementation, debugging and evaluation.
  • If I can't explain why a model is broken to someone who's stuck on it, I don't really understand it either. Two courses of that sharpened how I talk to non-technical stakeholders.
Data Analyst Intern
Oct 2024 – Jan 2025
Unified Mentor · Remote
  • Gathered reporting requirements from product and operations stakeholders, then wrote and tuned SQL over 4.5M+ transactional records, resolving nulls, duplicates and schema inconsistencies.
  • Owned the recurring reporting cadence and automated it, cutting manual reporting effort by an estimated 40%.
  • Built self-service Tableau and Power BI dashboards across revenue, retention and supplier performance, analysing $577K in revenue across 4.9K orders.
  • Went past reporting into modelling: a fraud classifier at 0.96 precision / 0.80 recall over 284K transactions, and a demand forecaster at 8.4% MAPE.
Machine Learning Research Intern
Jun – Jul 2022
Space Applications Centre (SAC), ISRO · under Dr. Jai G. Singla
  • Trained U-Net and DeepLabV3+ road-segmentation models in PyTorch (ResNet-50 encoder, Dice loss, IoU) on 6,723 aerial images from the Massachusetts Roads and DeepGlobe benchmarks, with QGIS in the geospatial loop.
  • Swept optimizers and architectures — the gap between Adam and RMSProp on the same architecture was as large as the gap between architectures.
  • The result that mattered was the failure: benchmark-trained models did not transfer to SAC's Indian satellite imagery, because every sensor writes its own raster metadata and a model trained at one resolution cannot read another. I wrote that up as the finding rather than burying it.
Research

Publications

Two published, one accepted, one under review. I keep those categories separate on purpose.

WikiFlow: Predicting Wikipedia Editor Dropout at Scale Using Distributed Machine Learning on Google Cloud

Accepted
Meena, A., Patel, K., Alizadeh-Shabdiz, F. · EAI CSECS 2026 · Springer LNICST · Scopus-indexed
First author. The first fully cloud-native pipeline for Wikipedia editor-dropout prediction, scaling to 1,158,248 editors — more than 10× the cohort size of prior single-machine studies — and the first application of synchronous distributed deep learning to this problem.

Accelerated CNN Training with Genetic Algorithm Optimization

Published
Meena, A., et al. · IEEE IATMSI · 2024
Integrated evolutionary optimisation with CNN training, cutting training time by over 50%, and proposed DoubledMNIST as a harder benchmark for generalisation testing.

Enhanced Scene Text Recognition for Accessibility

Published
Meena, A., et al. · ICADIE · Springer · 2024
Multi-class SVM over HOG features for character detection, aspect-ratio re-scoring before non-maximum suppression, then word assembly as a Boolean quadratic program relaxed to an SDP. We score 52.2% against the baseline's 56% — but the baseline decodes against a 50-word lexicon while we use only a bigram model built from the full 427-word one, a substantially harder setting.

Where Does the Dollar's Return Come From? An Overnight and Daytime Decomposition of the U.S. Dollar Index

Under review
Meena, A., Pinsky, E. · MDPI · 2026
First author. Fifteen years of DXY split into close-to-open and open-to-close legs, which move in opposite directions across the entire sample: $100 held only in the daytime session grows to $196.01, held only overnight it falls to $61.02, against $121.72 for buy-and-hold. We evaluate the complete 24-strategy set, then report that our own best result — 218.8% at a Sharpe of 1.06 — fails the Deflated Sharpe Ratio at 0.046 and White's Reality Check at p = 0.062, and that the edge vanishes at 1.56 basis points of execution cost. We treat that as the central finding, not a footnote.
Toolkit

What I actually use

Everything here appears in a project or a role on this page — nothing listed that I couldn't be questioned on.

Languages & query

PythonSQLRPostgreSQLBigQuery

Analytics & statistics

EDAHypothesis testingA/B testingANOVARegressionTime seriesCohort & RFM

BI & visualisation

Power BIDAXTableauStar-schema modellingPlotlyMatplotlib

Machine learning

scikit-learnXGBoostPyTorchRandom ForestSHAPSMOTEFeature engineering

Big data & cloud

PySparkApache SparkGCP DataprocGCSBigQuery

Tooling

GitDockerQGISLinuxJupyterLaTeX
Contact

Tell me what you're trying to find out

I'm looking for full-time Data Scientist, Data Analyst and BI Analyst roles — healthcare and fintech especially. If you have a messy dataset and a decision riding on it, that's the conversation I want.

Boston, MA · open to any US location and remote · available immediately
authorized to work in the US on F-1 STEM OPT