Vaishnavi Kallakuri — Product Analyst

01 — My Story

Engineer by training,
analyst by calling

I started out studying Electrical & Electronics Engineering in New Delhi — drawn to how systems work, where they break, and how to make them better. That problem-solving instinct followed me into industry.

At Tata Consultancy Services, I joined as a team member supporting Lloyd's Banking Group and Citi Bank — building dashboards, writing SQL pipelines, and learning the cadence of fast-moving fintech operations. Over three years I grew into a team lead role, owning 15–25 stakeholder KPI dashboards, presenting findings to senior leadership, and coordinating requirements across regulatory, engineering, and product teams.

Those years taught me how to translate a data anomaly into a boardroom insight — and how to earn the trust of people who need data to make decisions, not just consume it.

In 2024 I moved to Buffalo, NY to pursue my M.S. in Industrial Engineering (Data Analytics) at University at Buffalo, where I worked as a Graduate Research Assistant applying clustering, SQL, and Tableau to real procurement problems. I graduated in January 2026 and I'm now actively seeking my next role — open to relocating anywhere in the US.

Core Competencies

SQL

Data Analytics & Visualization

SQL · Python · Tableau · Power BI · BigQuery · GA4 · A/B Testing · Funnel Analysis

PM

Product & Delivery

Agile · Sprint Planning · KPI Dashboards · Stakeholder Mgmt · Requirements Gathering · SDLC

ML

ML & Optimization

scikit-learn · Random Forest · MILP · K-means · SHAP · statsmodels · ETL · Snowflake

6σ

Process & Quality

Six Sigma Green Belt · DMAIC · FMEA · SAP/ERP · Lean · Change Management

Career Timeline

B.E.

2016 – 2020

B.Tech · Electrical & Electronics Engineering

Guru Gobind Singh Indraprastha University, New Delhi

Systems thinking Engineering fundamentals

TCS

2021 – 2022 Team Member

Analyst · Tata Consultancy Services

Contributor on Lloyd's Banking Group & Citi Bank — dashboards, SQL pipelines, agile squads

SQL Tableau Agile

LEAD

2022 – 2024 Team Lead

Senior Analyst · TCS (Lead)

Led KPI dashboards for 15–25 stakeholders, drove executive presentations, shaped sprint backlogs

Python ERP/SAP RCA Executive reporting Sprint planning

M.S.

2024 – Jan 2026

M.S. Industrial Engineering · University at Buffalo

Graduate Research Assistant — supply chain analytics, clustering, Tableau, SQL

MILP Optimization K-means BigQuery · GA4

NOW

2026 — Now Open to work

Seeking Product Analyst roles · Open to US-wide relocation

F-1 STEM OPT · Six Sigma Green Belt · Agentic AI Certified

Product Analytics Data Analytics Business Analytics

03 — Experience

Where I've Made an Impact

Graduate Research Assistant — Product Analytics

Center for Supply Chain Analytics, University at Buffalo · Jul 2024 – Jan 2026

Academic · Applied Research

Problem Space

The supply chain center needed to streamline supplier qualification across 20+ accounts in aerospace, semiconductor, and industrial sectors — while giving procurement teams real-time visibility into product KPIs across 8–10 commodity categories.

Impact

20% faster supplier eval

30% more planning confidence

What I Did

→

Standardized product data collection for 20+ accounts, extracting ISO/ITAR/CMMC certifications into MS-SQL to accelerate sales team qualification

→

Built multi-tab Tableau dashboard tracking KPIs across 8–10 commodity categories for procurement & commercial teams

→

Gathered customer feedback, identified supply base gaps, presented FMEA-aligned risk findings to leadership

→

Applied Python K-means clustering + SQL to segment suppliers by capability, capacity, and geography

Research Outcomes Snapshot

Supplier eval cycle

–20% cycle time

Planning confidence

+30% stakeholder confidence

Accounts standardized

20+ accounts

Analyst — Product Operations & Data Enablement

Tata Consultancy Services · Gurgaon, India · Mar 2021 – Jul 2024

3+ years · Team member → Team lead

Problem Space

Lloyd's Banking Group and Citi Bank needed reliable, automated analytics to support product operations — but reporting was manual, ERP incidents were slow to resolve, and leadership lacked real-time product performance visibility.

Impact

45% less manual reporting

25% faster MTTR

30% better decision quality

What I Did

→

Automated Tableau KPI dashboards for 15–25 stakeholders tracking order fulfillment, inventory turnover, and ops efficiency

→

Diagnosed 20+ high-priority ERP/SAP incidents using Python and SQL root cause analysis, restoring data pipeline integrity

→

Partnered with product, engineering, and regulatory teams to translate business requirements into analytics solutions

→

Applied statistical modeling on Bloomberg & ICE data; delivered executive presentations to senior leadership

→

Led sprint planning and backlog prioritization across multiple workstreams as responsibilities grew

Before vs After — Efficiency Improvements at TCS

Reporting effort

–45%

MTTR (incidents)

–25%

Decision quality

+30%

04 — Projects

Selected Work

Click any project title to expand the full dashboard, charts, and analysis.

A/B Testing · Python · SQL · statsmodels

Checkout Discount UI — Conversion Lift Study

n = 350 usersTwo-proportion Z-testp = 0.00083 membership segments

▼

E-commerce platforms consistently lose users at checkout. This study investigated whether presenting a discount prompt at checkout (Treatment) increases purchase conversion compared to a standard checkout experience (Control). 350 users were split equally and analyzed across membership type, age segment, and spend tier. A user was defined as "converted" if they met both a spend and items purchased threshold — filtering for meaningful, high-intent purchases. The two-proportion Z-test was used for statistical validation at α = 0.05.

Dashboard · Experiment Results

350

Total users · 175 each group

51.4%

Control conversion · 90/175

33.7%

Treatment conversion · −17.7pp

+30.4pp

Lift for Age 31–40 · p<0.001

Overall Conversion by Group

Control

51.4%

90 users

Treatment

33.7%

59 users

Key insight: The overall negative lift masks strong segment-level signals. Averages hide the story — segmentation reveals it.

Conversion by Membership Type

TypeControlTreatment

Gold

100%

Silver

54.2%

0% ⚠️

Bronze

Gold converts 100% both ways — discount adds zero value, pure margin cost

Conversion by Age Segment — The Key Finding

Control

Treatment

Lift

18–30

98.3%

42.4%

−55.9pp

31–40

28.2%

58.6%

+30.4pp

Age 31–40 — Responds strongly to perceived savings. Discount triggers fence-sitters into buying.

Age 18–30 ⚠️ — May perceive discount as quality signal or face UX friction. Needs investigation.

Two-proportion Z-test

Test method

Z = −3.3512

Z-statistic

p = 0.0008

p-value · α = 0.05

H₀ Rejected ✅

Statistically significant

Conclusion & Recommendation

Do NOT roll out universally

Overall conversion dropped −17.7pp — statistically significant
Age 18–30 showed a large −55.9pp negative effect
Gold members at 100% either way — discount is pure margin loss
Silver Treatment anomaly (54% → 0%) needs investigation

DO target Age 31–40 + Mid-spend

+30.4pp lift for Age 31–40 is significant (p < 0.001)
Mid-spend tier shows positive treatment response
Discount triggers fence-sitters — perceived value signal
Run targeted follow-up on ~120 users in this segment

Tools:Python · statsmodelsSQLMatplotlib

+30.4pp for 31–40 cohort−17.7pp overall

Funnel Analysis · BigQuery · GA4 · Python · SQL

Google Analytics Funnel Analysis — Google Merchandise Store

38 unique usersGA4 BigQuery export6-stage funnel82% drop-off

▼

This project analyzes user behavior across the Google Merchandise Store using real GA4 BigQuery event data. The goal: map the complete e-commerce funnel from session to purchase, identify the largest conversion bottlenecks, and deliver device, traffic source, geographic, and product-level segmentation to support prioritized UX recommendations. Advanced SQL with UNNEST handled BigQuery's nested event schema, and Python (Pandas, Matplotlib) drove the segmentation and visualization pipeline.

Dashboard · Funnel & Segmentation Results

Unique users · Jan 2021

Funnel stages · Session→Purchase

82%

Biggest drop-off · Product→Cart

2.6%

Overall conversion rate

Mobile product views

User Funnel — Session → Purchase

All Sessions

38 users · 100%

Page View

11 · 29%

−71%

Product View

2 · 5%

−82%

Add to Cart

1 · 3%

−65%

Checkout

1 · 3%

−45%

Purchase

1 · 2.6%

−60%

Product View → Cart is the single highest-priority optimization target (82% loss)

Drop-off Rate at Each Funnel Stage (%)

71%

Session
→ Page

82%

Page
→ Product

65%

Product
→ Cart

45%

Cart
→ Checkout

60%

Checkout
→ Purchase

Biggest bottleneck: Product view → Cart at 82% — above-fold CTA redesign is the #1 priority.

Users & Product Views by Device

Users

Desktop

24 · 63%

Mobile

14 · 37%

Mobile dead zone: 37% of users, 0% product views — critical UX failure.

Users by Traffic Source

Google

16 · 42%

Other

11 · 29%

Direct

9 · 24%

GMS Ref.

Traffic quality gap: Google = 42% of users but 0 product views. Audit landing pages + UTMs.

Geographic Distribution

🇺🇸US

26 · 68%

🇶🇦Qatar

12 · 32%

Qatar = 32% of traffic — untapped MENA opportunity with localization potential

Key Insights & Recommendations

HIGH

Redesign Product Page CTA

Add urgency signals, social proof, and move add-to-cart button above the fold. ~65% of product viewers don't convert to cart.

Expected: +15–25% view→cart rate

HIGH

Mobile Checkout Optimization

Mobile = 37% of users, 0% product views. Audit page load speed (<2s target), sticky add-to-cart, one-tap checkout for returning users.

Expected: +20% mobile conversion

MED

UTM Parameter Discipline

Google drives 42% of traffic but 0 product views. Tag all campaigns with UTMs to reveal which channels actually convert.

Expected: Better attribution accuracy

MED

Qatar Geo-targeted Testing

Qatar = 32% of traffic with unknown conversion behavior. Test localized landing pages, QAR currency, MENA shipping info.

Expected: New revenue stream

Tools:BigQuery SQLPython · PandasMatplotlibGA4 Export

82% drop-off at product→cart0% mobile product views

MILP Optimization · Python · scipy / HiGHS · Supply Chain

EV Battery Supply Chain Optimization — MILP Model

Mixed-Integer Linear Programscipy.optimize.milp (HiGHS)5-Stage Supply Chain$66.37M Optimal Cost

▼

This project builds a production-grade mathematical optimization model for designing the optimal EV battery supply chain network. Given 5 mines, 4 refineries, 3 gigafactories, 3 assembly plants, and 5 regional EV markets — the model finds the minimum-cost configuration that satisfies all demand while respecting every capacity constraint. Binary decision variables force each facility to be either open or closed — which is why a simple heuristic fails and a MILP (HiGHS solver) is required. Branch-and-bound guarantees a provably optimal solution.

Dashboard · Optimization Results

$66.37M

Total optimal cost · proven minimum

8 / 10

Facilities opened · 2 closed

35,000

Battery packs delivered

52.4%

Cost is raw material procurement

5 / 5

Demand regions satisfied exactly

5-Stage Supply Chain Network — Optimal Configuration

Mines

Chile

Australia

Indonesia

Canada

Philippines

→

Refineries

S.Korea 83%

Germany 100%

China 100%

USA ✗

→

Gigafactories

CATL 100%

LG 100%

Panasonic ✗

→

Assembly

Shanghai 100%

Munich 92%

Michigan 64%

→

Markets

R1 ✓

E.Asia

R2 ✓

Europe

R3 ✓

N.America

R4 ✓

S.America

R5 ✓

SEAsia

Green = OPEN · Faded = CLOSED · % = utilization. All 5 demand regions satisfied exactly.

Cost Breakdown — Total $66.37M

Raw Material

$34.80M

52.4%

Processing Var.

$10.10M

15.2%

Cell Mfg Costs

$6.75M

10.2%

Pack Assembly

$3.50M

5.3%

Transport (all)

~$5M

~10%

Fixed (all fac.)

$4.25M

6.4%

Raw material = 52.4% of cost. Supplier selection is the primary lever — not logistics optimization.

Open Facility Utilization Rates

Processors

P1 S.Korea

83%

P2 Germany

100% full

100%

P3 China ⚡

100% bottleneck

100%

Gigafactories

C1 CATL ⚡

100% bottleneck

100%

C2 LG Energy

100% full

100%

Assembly Plants

A1 Shanghai

100% full

100%

A2 Munich

91.7%

92%

A3 Michigan

64.3%

64%

Facility Open / Close Decisions

Facility	Status	Throughput	Util.
P1 S.Korea Refinery	OPEN	10,000 t	83%
P2 Germany Refinery	OPEN	10,000 t	100%
P3 China Refinery	OPEN	15,000 t	100%
P4 USA Refinery	CLOSED	—	—
C1 CATL Gigafactory	OPEN	20,000 u	100%
C2 LG Energy Plant	OPEN	15,000 u	100%
C3 Panasonic Nevada	CLOSED	—	—
A1 Shanghai Assembly	OPEN	15,000 pk	100%
A2 Munich Assembly	OPEN	11,000 pk	92%
A3 Michigan Assembly	OPEN	9,000 pk	64%

P4 USA closed: Highest fixed cost ($700K) + $420/ton variable — Asia-Pacific dominates.

C3 Panasonic closed: C1+C2 cover all 35K demand. Opening C3 adds $1M fixed with no saving.

Demand Satisfaction — All 5 Regions ✓

10,000

E. Asia

8,000

Europe

9,000

N. America

4,700

S. America

3,300

S.E. Asia

Key Findings & Business Implications

Raw material = the primary cost leverAt 52.4% of total cost, procurement strategy dominates. Better supplier contracts yield far more savings than logistics optimization.

Transport costs are only ~10% combinedThe network is geographically well-positioned. Facility selection matters more than shipping route optimization.

P3 China + C1 CATL are critical bottlenecksBoth at 100% utilization. Any demand growth hits these nodes first — expand here for maximum ROI per dollar of capex.

Provably optimal — not just "good"HiGHS branch-and-bound guarantees the global minimum. A nearest-neighbor heuristic gives a feasible but suboptimal answer with no proof of optimality.

Tools:Python 3scipy.optimize.milpHiGHS SolverPandas · Matplotlib

$66.37M optimal8/10 facilities open35,000 packs · 5 regions

ML Pipeline · scikit-learn · Random Forest · Gradient Boosting

Hepatitis C Early Detection — ML Classification Pipeline

AUC 0.98Random Forest + GBMFeature EngineeringCross-Validation

▼

Built an end-to-end machine learning pipeline for early Hepatitis C detection using Random Forest and Gradient Boosting. The pipeline covers feature engineering, cross-validation, model interpretability, and threshold optimization — designed for clinical screening triage use cases where false negatives carry high cost. Achieved AUC of 0.98 with well-calibrated precision-recall tradeoffs across both ensemble models.

Dashboard · Model Performance & Clinical Insights

0.98

Best AUC · RF & GBM

97.3%

Best accuracy · Random Forest

615

Patient records · UCI dataset

Biochemical markers analyzed

ML models compared

Model Performance Comparison — AUC & Accuracy

Model
AUC
Accuracy
AUC

0.98

97.3%

0.98

GBM

0.98

97.1%

0.98

SVM RBF

0.97

97.8%

0.97

0.96

96.5%

0.96

GAM

0.95

92.0%

0.95

GLMNET

0.94

91.2%

0.94

CART

0.90

89.5%

0.90

RF + GBM best10-fold CV · 5 repeats

Top Biomarkers — Point-Biserial Correlation with HCV Status

AST

r = 0.61

p<.001

GGT

r = 0.58

p<.001

BIL

r = 0.52

p<.001

CHOL

r = 0.38

p<.001

ALB

r = 0.31

p<.001

Mean Marker Levels — Donor vs HCV Patient

AST (U/L)

22 vs 45

GGT (U/L)

28 vs 61

BIL (mg/dL)

0.7 vs 1.2

CHOL (mg)

190 vs 160

ALB (g/dL)

4.2 vs 3.5

All differences p<0.001

End-to-End ML Pipeline

1. Data Preprocessing

615 records · 11 markers · Median imputation (<2% missing) · 5-NN imputation (2–8% missing) · Winsorization at 99th pct for outliers

2. Feature Engineering

Log transforms for right-skewed markers (AST, ALT, GGT) · PCA — first 3 PCs explain 72% variance · Point-biserial correlation ranking

3. Modeling

7 models: GLM, GLMNET, GAM, MARS, CART, RF, GBM, SVM, NN · Nested 5-fold CV within 10-fold outer · Hyperparameter grid search

4. Interpretability (SHAP)

SHAP values confirm AST & GGT as primary drivers · Clinical alignment — high AST SHAP values correlate with disease severity · Enables clinical trust

Sensitivity · Specificity · Kappa — Best Models

Random Forest AUC 0.98

96%

Sensitivity

98%

Specificity

0.94

Kappa

Gradient Boosting AUC 0.98

95%

Sensitivity

98%

Specificity

0.93

Kappa

SVM (RBF) AUC 0.97

96%

Sensitivity

98%

Specificity

0.95

Kappa

Conclusions & Clinical Impact

RF & GBM are the optimal modelsBoth achieve AUC 0.98, 97%+ accuracy, sensitivity 0.95–0.96, specificity 0.98 — robust across all 5 repeats of 10-fold CV with minimal variance.

AST & GGT are the primary clinical driversPoint-biserial r = 0.61 and 0.58 respectively. SHAP analysis confirms these are the top predictors — matching known liver damage physiology.

Non-invasive screening at scaleRoutine blood tests (no invasive biopsy needed) fed into this pipeline can flag HCV-positive patients early — critical when disease is asymptomatic until advanced stages.

SHAP interpretability enables clinical trustBlack-box models are augmented with SHAP value analysis — clinicians can see which biomarkers drove each prediction, making the model deployable in decision support systems.

Multiclass extension: Expand binary HCV/donor classification to stage-level (fibrosis, cirrhosis) using ordered multinomial models.

Longitudinal tracking: Develop time-series models to monitor patient markers over visits — detecting deterioration trends before clinical threshold is crossed.

Clinical deployment: Integrate into EHR decision support systems with SHAP-powered explanations, ensuring regulatory compliance and physician trust.

Tools: Python · scikit-learn Random Forest · GBM SVM · NN SHAP · Cross-validation

AUC 0.98 97.3% accuracy 615 patients · 11 markers

From circuits to dashboards.
Turning data into decisions.

Engineer by training,
analyst by calling

Technical Arsenal

Where I've Made an Impact

Selected Work

Credentials & Learning

Let's Connect

From circuits to dashboards.Turning data into decisions.

Engineer by training,analyst by calling

Technical Arsenal

Where I've Made an Impact

Selected Work

Credentials & Learning

Let's Connect

From circuits to dashboards.
Turning data into decisions.

Engineer by training,
analyst by calling