Hey, I'm
$ git log --oneline --graph
Five years of turning messy data into statistical evidence set the direction. Every stop since has been about building models and systems that hold up under real scrutiny, not just look good in a notebook.
feat: SQL-based data validation framework, -60% discrepancies
First real lesson in production data: a 4.5M-customer billing base doesn't forgive weak validation. Designed row-level checks and reconciliation logic that cut discrepancies 60% and shortened month-end close by 10%.
feat: anomaly detection at scale, 600K+ subscribers
Moved from fixing issues after the fact to designing the checks that catch them first — anomaly detection and reconciliation holding 98%+ accuracy within SLA. Earned "Best Upcoming Talent," FY23 Q3.
feat: resume-matching model + compliance engine, 96% accuracy
Went back to build the theory behind the practice: a TF-IDF → SBERT → neural refinement pipeline for candidate screening, and a deterministic compliance engine auditing AI fall-monitoring technology against 5 CMS-mapped thresholds.
feat: OLS spend forecasting + AI cost-attribution, 5.5x growth surfaced
Where the two threads met: an OLS-regression model projecting $53K-$57K in annual spend, and a 211-check LLM-assisted rule pipeline flagging 11k records across 23 tables.
$ cat PRINCIPLES.md
Everything I build is grounded in a production mindset.
"Understanding existing systems deeply before changing them, designing for the next person who has to maintain my work, and communicating technical findings in a way non-technical stakeholders can act on." That’s the standard I actually build to.
AI Sentinel pairs its local-LLM explanation layer with a deterministic rule engine and a 100%-deterministic fallback, so a hallucinated inference never becomes an unchecked compliance decision. Every finding is SHA-256 audit-logged, nothing is a black box.
→ see it in AI Sentinel 02In the osteoarthritis market-share forecasting project, every model had to beat "same as last month" under rules written before any model ran. For the national number none of three did, so the simple rule stayed. The Resume–JD pipeline follows the same habit: TF-IDF first, then SBERT, each stage measured against the one before.
→ see it in the OA project 03SBERT's embeddings accelerate semantic matching far beyond what keyword search alone could do, but every embedding-based match is still validated by the supervised neural layer against labeled outcomes before it becomes a probability score anyone should trust.
→ see it in Resume–JD Matching 04The sentiment-analysis pipeline's location-normalization step cut unresolved geography from 72.8% to 43.85% before a single sentiment score was trusted. Cleaning happens before modeling, not after a bad result shows up.
→ see it across Experience$ cat experience.log
Barton Malow Featured
Pennsylvania State University
Oracle Corporation
Accenture Solutions Pvt Ltd
$ ls ./projects
Data science flagship project. Can anyone forecast a branded osteoarthritis injection's share of doctor visits, and can the data be trusted? IQVIA visit data plus 171 public files in a Bronze/Silver/Gold warehouse, 7 leakage-tested feature groups, and models held to simple baselines under rules written before any model ran.
Autonomous compliance agents detecting AI-driven regulatory risk in real time. Top 10 of 60+ teams, Penn State Nittany AI Challenge 2026. Deterministic rule engine derives 5 CMS-mapped thresholds via Gemini Pro 3.1, explained in plain English through a local Llama 3 layer.
3-stage semantic ranking pipeline built for RA candidate screening at Penn State: TF-IDF baseline → SBERT semantic embeddings → supervised neural refinement, reaching 96% accuracy across 120 resumes and 2,277 projects.
End-to-end NLP + geospatial pipeline tracking global sentiment toward LLMs across 305K+ tweets. Location-normalization cut unresolved geography from 72.8% to 43.85%.
View repo →Deep Dive · Data Science Flagship Project, 2026
In plain English: a drug company's sales were growing, but the market data said doctors were prescribing its osteoarthritis injection far less. I tested whether anyone can forecast that trend, and found out first that the data itself could not be fully trusted.
A brand manager selling an osteoarthritis injection (Zilretta) needs an early signal when competing steroid and NSAID treatments take doctor-visit share. The one market dataset (IQVIA visit counts) disagreed with the company's own sales.
A sales target or forecast built on a number that may be wrong is worse than no forecast. Before predicting anything, the work had to show whether the data and the models could be trusted.
A Bronze/Silver/Gold SQLite warehouse over 72 months of IQVIA visits and 171 public CMS, CDC, SEC and FDA files, 7 leakage-tested feature groups, then models held to simple rules in walk-forward testing (about 46 test months). The pass/fail rules were written before any model ran.
Sales rose 5% and 6% while visits fell 14% and 36%. A target built on that decline could be wrong by a large factor until IQVIA confirms panel coverage.
They are 64% of the net fall of 11,256 visits, so that is where a field or marketing review pays off soonest.
Michigan, Arizona, Florida, Minnesota and Indiana have the most room to grow, on a ranking that held steady between 2022 and 2024 (0.92).
Last month's value beat all three trained models, so plan on it with its range. A model would add cost and no accuracy.
Status: analysis and modelling are complete. The stakeholder website and question box are planned, and nothing is deployed. These are decisions the analysis supports, not measured revenue impact.
This is where principle 02 (baseline before complexity) isn't an abstraction: every model had to beat "same as last month" under rules written before it ran. For the national number none did, and reporting that plainly is the result.
Deep Dive · Penn State Nittany AI Challenge, 2026
Fall-monitoring AI systems can quietly restrict resident movement in ways that violate federal nursing-home regulations, and nobody was watching for it in real time.
A missed compliance risk isn't just a fine, it's a resident's autonomy on the line. And a black-box LLM verdict isn't good enough when the finding has to hold up to an auditor.
A hybrid architecture: a deterministic rule engine derives 5 CMS-mapped thresholds via Gemini Pro 3.1, a local Llama 3 layer explains each finding in plain language, and a 100%-deterministic fallback means the LLM never gets the final say.
This is where principle 01 (a model never gets the final say) isn't an abstraction: every LLM explanation is checked against the deterministic rule engine before it's trusted, and every finding is SHA-256 audit-logged.
In the news: AI Sentinel was named among the 15 Prototype Phase winners advancing to the MVP Phase of the Nittany AI Challenge. Read the Penn State News story →
Deep Dive · Penn State Research Assistant, 2026
Screening RA candidates by keyword matching misses people who describe their skills differently than a project listing does, and over-ranks resumes that just repeat the right buzzwords.
A ranking system that can't tell semantic fit from keyword-stuffing produces bad shortlists on both ends: real candidates filtered out, weak ones let through.
A 3-stage pipeline that only adds complexity when it earns its place: a TF-IDF baseline first, then SBERT semantic embeddings, then a supervised neural layer that outputs a calibrated match probability instead of a raw similarity score, reaching 96% accuracy across 120 resumes and 2,277 projects.
This is where principles 02 (baseline before complexity) and 03 (AI accelerates the analysis, it doesn't replace the validation) aren't abstractions: TF-IDF proved the floor, SBERT proved the ceiling was higher, and every embedding-based match still had to earn its place against a supervised, labeled baseline before it became a trusted probability.
$ cat research.md
Accepted and presented at SEET 2026 (2nd International Conference on Software Engineering of Emerging Technologies), Penn State Behrend, PhD Research Track. Listed on page 76 of the conference abstract book. Now published in the official conference proceedings by Springer Nature (CCIS vol. 3126).
Read on Springer → View conference →Won the Prototype Phase and MVP Phase, advanced through funded phases to a Top 10 finish at the Pitch Contest, Hintz Family Alumni Center, University Park. Named among the 15 Prototype Phase winners in Penn State News (Feb 4, 2026).
Read the Penn State News story →Recognized for exceptional contributions and high-impact performance on the Telia Company billing analytics engagement.
Representing students in discussions with university leadership; championing AI literacy and equitable AI access across Penn State's Commonwealth Caucus.
$ cat skills.yaml
Regression · Classification · Clustering · NLP · Time Series Forecasting · Feature Engineering · Model Evaluation (AUC-ROC, F1)
Python (Pandas, NumPy, Matplotlib, Seaborn, scikit-learn, TensorFlow, PyTorch, Keras) · SQL · R · Java
ChatGPT · Julius AI · Cursor IDE · Perplexity AI · Claude Code · GitHub MCP · Claude Skills
Databricks · PySpark · Tableau · Advanced Excel · AWS · GitHub
Git · GitHub Actions · Jupyter · Jira
$ cat education.log
Master of Data Analytics · GPA 3.96/4
Statistical Analysis, Data-Driven Decision Making, Predictive Analytics, Deep Learning, Natural Language Processing.
B.Tech, Electronics & Instrumentation Engineering · GPA 3.3/4
Control Systems, Digital Electronics, OOP, Data Structures & Algorithms, Web Technology, Artificial Intelligence.
$ mail tushar
Open to full-time Data Scientist / Machine Learning Engineer roles starting Dec 2026.