Work and Education

Ask Atharva

ask a real question...

Press Enter to send.

View selected work ↓

~/Atharva Fulay/work_and_edu % ls ./selected-work/

[01] NEON LinkedIn / AI reporting [02] Holistic Signals LinkedIn / reporting platform [03] Nike Affinity Platform Nike / personalization [04] Platform Stability LinkedIn / operational AI [05] Sponsored Messaging LinkedIn / migration

~/Atharva Fulay/work_and_edu % cat ./selected-work.md

Selected Work

I bring analytics rigor, platform thinking, and end-to-end technical ownership to building applied AI and data products. I’m most energized by turning ambiguous, expert-led workflows into reliable, reusable systems that people can trust, adopt, and scale.

[01]

LinkedIn / AI reporting

NEON

I lead the insights architecture for a LinkedIn AI reporting agent, designing reusable skills that turn reporting expertise into composable capabilities.

role
Technical lead for insights architecture, reusable skill design, and the cross-team operating model.
context
Reporting expertise lived across many reports, presentation rules, documentation, and teams.
evidence
Reusable reporting-skill framework and Share of Feed reference implementation; testing with approximately 60 power users ahead of broader launch.
Learn more about NEON
constraints
A report-by-report chatbot would have reproduced the same knowledge silos and ownership ambiguity.
built
I designed and coded the first reusable reporting-skill framework, defined the MVP reporting portfolio, proposed reusable insight modules, and led working sessions that established responsibilities across Insights, Analytics Engineering, and Platform teams. I also shaped orchestrator-worker and output strategy.
impact
The architecture creates a path for analytical reasoning to become reusable capabilities with defined interfaces and ownership. The current public status is active testing, not completed general availability.
tools
Python, SQL, reporting systems, skill architecture, orchestration, evaluation, and cross-functional operating-model design.
[02]

LinkedIn / reporting platform

Holistic Signals

I transformed a manual conversion-analysis workflow into a self-service reporting platform, including the Offline Signals foundation.

role
Product and delivery owner for requirements, data dependencies, seller discovery, launch enablement, and cross-functional alignment.
context
The earlier analyst-led workflow took roughly two to three hours and focused too narrowly on CAPI.
evidence
Client-ready insights in minutes rather than hours across CAPI, Insight Tag, and Offline Conversions.
Learn more about Holistic Signals
constraints
Signal and attribution data was spread across schemas, identifiers, owners, attribution stages, and deduplication behavior.
built
I owned or drove the product requirements, delivery, seller discovery, presentation outputs, sample validation, metric definitions, launch enablement, training, FAQs, and communications. I aligned Product, Engineering, Product Marketing, Analytics, and Sales around delivery and data constraints, then redirected the Offline Signals foundation into a defined engineering initiative when the original path was deprioritized.
impact
The platform reduced dependence on scarce analyst capacity and created a foundation for broader conversion reporting. The Offline Signals data foundation remains described as an engineering initiative under validation where the source material requires that qualifier.
tools
SQL, Python, reporting outputs, metric design, data contracts, product requirements, and stakeholder alignment.
[03]

Nike / personalization ML

Nike Affinity Platform

I redesigned Nike’s affinity system end to end, then expanded it into a self-service, metadata-driven personalization platform.

role
End-to-end technical owner across model architecture, self-service platform delivery, validation, and campaign integration.
context
The legacy system was slow, brittle, limited in signals, and able to score audiences only weekly.
evidence
Runtime fell from more than eight hours to under 30 minutes; daily scoring, under-10-minute custom audiences, and an estimated $14M annual incremental revenue contribution.
Learn more about Nike Affinity
constraints
Model quality, audience delivery, product taxonomy, and campaign validation had to work together across production data and marketing workflows.
built
I redesigned the system end to end: separated ingestion, feature construction, training, scoring, and audience delivery; rebuilt the pipeline in PySpark and Databricks; added data-quality guardrails; built self-service audience generation; and moved affinity definitions into metadata.
impact
Predictive accuracy improved from below 60% to approximately 85% in the source wording, revenue per send increased by nearly 50%, unsubscribe rates fell by roughly half, and category changes moved from roughly two to four weeks to under three days.
tools
Python, SQL, PySpark, Databricks, AWS, gradient boosting, feature engineering, backtesting, experimentation, and production pipelines.
[04]

LinkedIn / platform operations

Platform Stability & Operational AI

I re-architected high-usage reporting tools, stabilized inherited systems, and turned recurring support patterns into reusable standards, roadmaps, and operational AI.

role
Technical and product ownership across performance engineering, reliability, platform consolidation, and operational AI.
context
High-usage tools combined long runtimes, recurring failures, duplicate product surfaces, and recurring support burden.
evidence
Across three resource-intensive tools, runtime fell from more than 80 minutes to under 15 minutes, CPU usage fell by up to 95%, and support improved for 1,000+ global sales users across roughly 3,000 monthly runs supporting $500M+ in revenue context.
Learn more about Platform Stability
constraints
The work spanned inherited systems, changing measurement standards, field requests, production incidents, data-quality issues, and teams with different ownership boundaries.
built
I re-architected expensive code paths, removed redundant transformations and scans, improved query structure and exception handling, and created a reusable Python and SQL tuning framework. As business owner for Share of Feed, I prioritized enhancements, coordinated delivery, and handled escalated production and data-quality issues. I also translated recurring support patterns into standards, metrics, roadmaps, and an internal support-ticket agent workflow.
impact
Reporting became faster and more reliable while platform decisions became easier to operate. Revenue is described as supported or associated with reporting, not as direct causal attribution.
tools
Python, SQL, Spark, Databricks, Trino, Airflow, AWS, performance profiling, metrics, roadmaps, and operational AI.
supporting work index

LinkedIn

Owned Share of Feed consolidation and Performance ROI stabilization; drove Competitor Analysis governance, the Standard Metrics Lite transition, support-pattern roadmaps, and production issue resolution.

Nike

Built replenishment decision systems, led affinity-growth forecasting, and ran a zero-downtime Databricks runtime migration with cross-team validation.

Black Mountain Systems

Designed and delivered MiFID II compliance messaging and a high-volume trade-entry workflow that reduced manual effort by approximately 10 hours per week.

~/Atharva Fulay/work_and_edu % cat ./education.md

Applied math and data science foundations.
[01]

USC — MS Applied Data Science

University of Southern California · 2021. Data systems, machine learning, NLP, algorithms, data mining, and UI/UX.

[02]

UCSD — BS Applied Mathematics

University of California, San Diego · 2017. Data inference, statistics, probability, linear algebra, and Java.