Open to Work · 2026
“Every data has a story it isn't telling yet. I enjoy the process of finding it.”
“The real value of data is not in the dashboard or the number, it is in the decision it improves.”
Focus
Data Analyst · Causal Inference · LLMsDegree
M.S. Computer Science · 3.97 GPAI started out doing data analysis in places that had no data analyst. At Sree Nirman, 50,000+ construction records lived across scattered files and inconsistent formats. Pricing decisions were made on gut feel because nobody trusted the numbers enough to use them. At Avanthi High School, it was the same problem in a different shape: 12,000+ financial records and 50,000+ expense records with no standardized schema, and a scholarship process nobody could fully explain. In both places, my first job was earning the right to be trusted with the data at all, building the SQL and Python pipelines, validation checks, and Tableau dashboards before I could say anything interesting about what the numbers meant.
Once that foundation existed, the real work started. At Sree Nirman, I built a tender estimation model to flag pricing risk before bids went out. At Avanthi, I replaced the scholarship process with a system combining regression and a fine-tuned language model, then ran a two-channel A/B test to confirm a new admissions initiative was actually working. It was.
The project that changed how I think about analysis was FanHouse, a study I built with synthetic retail data. Members spent 22.3% more after launch, but eventual members already spent 40.6% more before it. I used matching followed by difference-in-differences to make the comparison more useful. Estimated product value increased before discounts, while higher net payments were not established. I kept the pretrend warning and sensitivity results visible. That was when I started asking not just whether a number looked convincing, but what evidence would let me defend it.
That's the thread through everything I do now: SQL and Python to make data trustworthy, Tableau to make it legible, and causal inference to make it actionable. I'm currently looking for my next place to do that, as a Data Analyst, Business Analyst, BI Analyst, Analytics Engineer, or in Applied AI.
Questions I explored, things I built, and what the evidence taught me.
Can modern machine learning beat a classical forecast on five U.S. loan-delinquency series?
ARIMA beat both ML models after data leakage was fixed.
I followed 40,000 synthetic customers to separate who joins a membership from what the program changes.
Net-payment 95% interval: −$0.0232 to +$0.0027 per customer/day. Higher payments were not established.
Turn scattered daily job postings into searchable, explainable market intelligence.
An automated pipeline connects live postings to grounded AI summaries.
Adapt a small language model to write fantasy, romance, and science-fiction stories.
Lower held-out perplexity after QLoRA adaptation.
Combine clinical records and environmental context to forecast 30-day hospital readmission.
Risk predictions paired with contextual explanations and follow-up actions.
Take an insurance eligibility model from an experiment to a reproducible prediction service.
Versioned data, validation, and CI/CD support the deployed API.
Help movie lovers discover their next watch through personalized recommendations.
Recommendations, watchlists, ratings, and AI review summaries in one app.
Use academic and enrollment data to classify dropout, enrolled, and graduate outcomes.
The strongest tested model on accuracy and F1-score.
Narrative-style technical writing on data, machine learning, LLMs, and applied AI.
A personal journey into tokenization, probability, and next-token prediction.
I wrote this as a learning trail through LLMs, starting from the first question that bothered me: how does a model even see text? The article walks through tokenization, Byte Pair Encoding, next-token probabilities, inference, and how a base model behaves like a compressed memory of internet text.
Read on LinkedIn →A story-style introduction to how neural networks read text.
I used a bedtime-story format to explain why neural networks cannot directly read words and why text must become numbers first. The article introduces tokens, vocabularies, encoding, and the strange little bridge between human language and machine-readable input.
Read on LinkedIn →The story of Byte Pair Encoding.
This article explains Byte Pair Encoding through a story where characters, fragments, and frequent pairs slowly become a vocabulary. I wanted to show how BPE helps models handle unfamiliar words by breaking language into reusable pieces instead of memorizing every possible word.
Read on LinkedIn →How embeddings and position help models understand context.
I built this article around a simple misunderstanding between Arjun and Meera: “I need space.” From there, I explain input-target pairs, token embeddings, positional embeddings, and why the same word can mean different things depending on where it stands and what surrounds it.
Read on LinkedIn →A bedtime story about how machines learned to remember.
I used a story between a father and a little girl to explain why attention became such an important idea in AI. The article starts with old encoder-decoder models, memory, vectors, and the problem that eventually made attention feel less like a feature and more like a rescue mission.
Read on LinkedIn →Drag sideways or use the buttons →
Where it started is at the bottom. Everything since has grown up from there.
Aug 2025 – May 2026
University of Missouri-Kansas City
AssistantshipTechnical Support • Lab Operations • Student Data Support
Owned technical support for student-facing computer labs, resolving 75+ tickets per semester across hardware, printer, login, and workstation issues, alongside Python, SQL, R, and notebook-based coding support.
Drove analytics coaching across 20+ student academic and research projects in Python and R, strengthening EDA rigor, statistical validation, and assumption-checking in peer analysis.
May 2023 – Jun 2024
Sree Nirman, Hyderabad, India
Internship → Full TimeConstruction Analytics • Machine Learning • Growth Analytics
Owned data quality across a 50K+ record construction portfolio spanning cost, labor, budget, and sales, standardizing fragmented schemas to lift reporting reliability by 30%.
Designed SQL logic using joins, CTEs, CASE expressions, and window functions to standardize cost, labor, and sales calculations enterprise-wide across variance, productivity, and progress-tracking metrics.
Addressed a limited sample of 150–200 real project records by generating synthetic training records sampled from real feature distributions, scaling the dataset to a size viable for reliable model benchmarking.
Engineered 20+ predictive features across 7+ project dimensions, benchmarking Ridge and Lasso regression against a baseline linear model using R², MSE, and MAE to estimate tender pricing within 10–20% of held-out results.
Delivered Tableau dashboards, Excel scorecards, and executive reporting packs surfacing ranked cost drivers and variance signals, presenting bi-weekly to stakeholders and driving a 15% gain in operational efficiency.
Configured a scheduled Tableau Server extract refresh, automating weekly dashboard updates and eliminating manual republishing.
Tracked and compared model experiments using MLflow, logging runs and metrics across candidate models and registering the best-performing version for reuse.
Supported the co-founder’s YouTube content strategy on construction education and pricing transparency, contributing to channel growth from roughly 100 to 60,000 subscribers, 12 house-construction deals, and 80+ consultation engagements.
Apr 2022 – Jan 2023
Avanthi High School
InternshipFinancial Analytics • Education Data • ML Decision Support
Built the school’s analytics foundation from scratch, structuring 12K+ student financial records and 50K+ expense records into validated, reporting-ready datasets in SQL, Python, and Tableau.
Delivered recurring Tableau dashboards and reporting packs on fee collections and budget variance, giving leadership a consistent, standardized view of financial performance.
Analyzed 50K+ expense transactions across dining, hostel, and academics, identifying 10–15% in cost-saving opportunities that supported the institution’s first break-even cycle within four months.
Replaced an inconsistent scholarship-allocation process with an XGBoost model predicting final marks at 0.8 MAE, engineering per-grade snapshot features for partial-history students.
Fit quantile models attaching a 95% confidence interval to each prediction, consolidating a four-person review panel into a single reviewer.
Scored admission-test responses through an OCR-to-GPT-2 pipeline, measuring agreement against teacher grading before rollout.
Paired the redesigned allocation formula with a new admission-test fee stream, together driving a 50% rise in school revenue.
Designed a two-channel A/B comparison isolating the admission test’s marketing impact specifically, accounting for 15 of those 50 percentage points and driving an 18% rise in admissions.
Tools and the things I used.
Tools I use to turn messy business data into decision-ready reporting.
Cleaning, validating, reconciling, and preparing reliable datasets.
Building predictive models for risk, pricing, operations, and outcomes.
Testing whether an observed effect is real before trusting it.
Fine-tuning, evaluating, and building AI-assisted decision systems.
Moving models from notebooks into reproducible and deployable systems.
Building full-stack data products, apps, and cloud-backed ML systems.
Always learning.
For data roles, applied AI conversations, collaborations, referrals, or project discussions, these are the easiest ways to reach me.