# Portfolio Work Source This source supports recruiter-chat answers about Atharva Fulay's professional work and should stay aligned with the public Work + Education page. It adds retrieval-friendly project detail from his current resume and accomplishment bank. ## Professional Focus Atharva works at the intersection of applied AI, analytics engineering, machine learning, enterprise data products, and technical strategy. A recurring pattern in his work is converting specialist knowledge or manual analytical services into reusable systems that other teams can operate directly. ## LinkedIn ### Role and Context Atharva has been a Senior Analytics Engineer at LinkedIn since January 2025. His work spans AI reporting architecture, analytics products, data foundations, performance engineering, platform migrations, and cross-functional operating models for LinkedIn Marketing Solutions. ### AI Reporting Agent and Reusable Skill Architecture LinkedIn's reporting expertise was distributed across more than 60 reports, presentation logic, analyst knowledge, documentation, and separate technical teams. A report-by-report chatbot would have reproduced those silos. Atharva leads the insights architecture for an AI reporting agent. He: - designed and coded the first reusable reporting-skill framework and Share of Feed reference implementation; - proposed decomposing reports into reusable insight modules instead of mapping one report to one skill; - defined the MVP reporting portfolio and competitive-benchmarking tools; - shaped orchestrator-worker design, output strategy, narrative generation, and compliance considerations; and - led working sessions that clarified responsibilities across Insights, Analytics Engineering, and Platform teams. This architecture creates a path for teams to encode analytical reasoning as reusable capabilities with defined interfaces and ownership. The agent is being tested by approximately 60 power users ahead of a broader launch. The current status is active development and testing, not a claim of completed general availability. ### Holistic Signals Reporting Platform A manual analyst workflow for conversion analysis took roughly two to three hours per report, was difficult to reproduce at global scale, and initially focused too narrowly on CAPI. Atharva transformed it into the Holistic Signals reporting platform spanning CAPI, Insight Tag, and Offline Conversions. He owned or drove the product requirements, delivery epic, seller discovery, PowerPoint and Excel output requirements, sample validation, metric definitions and limitations, launch enablement, training, FAQs, and communications. He aligned Product, Engineering, Product Marketing, Analytics, and Sales around the product and its data constraints. The resulting self-service workflow lets sellers generate client-ready insights in minutes rather than hours. It reduces dependence on scarce analyst capacity and creates a foundation for broader conversion reporting. Claims about exact current run counts or adoption beyond availability should be verified before use. ### Holistic Signals Data Foundation The reporting platform exposed a deeper data problem: signal and attribution information was spread across datasets with inconsistent schemas, identifiers, ownership, attribution stages, and deduplication behavior. Atharva documented the dataset gaps, use cases, schema and performance requirements, and the need for a deduplicated source. When the initial engineering path was deprioritized, he redirected the work toward Analytics Engineering and co-defined conversion logic and implementation diagnostics across Insight Tag, CAPI, and offline conversion data. This work aligned teams around a shared technical definition and turned an unresolved dependency into a defined engineering initiative. The data foundation was under validation at the time captured by the accomplishment bank, so it should not be described as fully launched without newer evidence. ### Sponsored Messaging Reporting Migration A measurement-methodology change replaced impressions and clicks with sends and opens. Many reporting tools depended on impression-based logic, creating a portfolio-wide migration with uncertain platform dependencies. Atharva personally developed updates for LinkedIn's eight highest-usage reporting tools, adapting filters, calculations, and outputs across SQL, presentation output, and downstream reporting. He then turned that implementation into a distributed migration standard and led updates to more than 15 additional tools by: - authoring the product requirements and technical background; - creating the project structure, tracker, and per-tool implementation rules; - pre-analyzing impacted metrics and assigning owners; - coordinating dependencies, reviews, and field communications; - supervising intern delivery on four migrations; and - deciding to move the second phase forward instead of waiting indefinitely for a delayed platform dependency. The work combines hands-on engineering with technical program leadership: Atharva established the tested pattern, then scaled it across a broader reporting portfolio. ### Performance Optimization of High-Usage Reporting Tools Several heavily used tools had long runtimes, high CPU consumption, recurring failures, and significant support burden. The worst could take more than 80 minutes to run. Atharva profiled expensive code paths and failure points, removed redundant transformations and scans, improved query structure and exception handling, and partnered with engineering on dataset and compute improvements. He also created a reusable Python and SQL performance-tuning framework for other developers. Across the three most resource-intensive tools: - runtime fell by more than 80%, in some cases from more than 80 minutes to under 15 minutes; - CPU usage fell by as much as 95%; and - reliability improved for more than 1,000 global sales users across roughly 3,000 monthly runs supporting more than $500 million in revenue. The accomplishment bank recommends describing the revenue as supported or associated with the reporting rather than claiming direct revenue causation. ### Share of Feed Platform Consolidation LinkedIn maintained three overlapping Share of Feed products with duplicated maintenance, inconsistent capabilities, date restrictions, and gaps identified by field users. As business owner, Atharva gathered and prioritized feedback, defined enhancements, coordinated development, maintained the change tracker, wrote training and FAQ material, added video metrics and automated takeaways, and handled escalated production and data-quality issues. Major enhancements launched, including paid and organic filters, broader inputs, longer trend windows, rank and audience-penetration outputs, video metrics, and automated takeaways. Full legacy deprecation was still dependent on final execution work, so this should be described as platform consolidation in progress rather than a fully completed deprecation. ### Product Governance, Reliability, and Operating Models Additional LinkedIn work includes: - driving legal and information-security approval for a flexible-date Competitor Analysis enhancement; - validating replacement functionality and coordinating deprecation of Standard Metrics Lite; - taking ownership of Performance ROI tools, correcting account-to-representative mappings, testing through local and Airflow workflows, and stabilizing inherited systems; - translating field feedback and support patterns into roadmaps, OKRs, trackers, and ownership structures; and - resolving escalated production issues across reporting products and feeding recurring problems back into documentation and product priorities. These are supporting examples of product ownership, governance, lifecycle management, and reliability rather than Atharva's primary headline accomplishments. ### LinkedIn Tools and Themes Python, SQL, Spark, Databricks, Trino, Airflow, AWS, analytics products, agentic workflows, skill architecture, performance optimization, data modeling, product requirements, legal and compliance partnership, operating-model design, and cross-functional delivery. ## Nike ### Role and Context Atharva was a Senior Data Scientist at Nike from June 2021 through May 2024. He built production machine-learning and self-service systems for personalization, audience strategy, product discovery, replenishment, and planning. ### Affinity V2 Personalization System Nike's legacy affinity model was slow, brittle, limited in the signals it could use, and able to score audiences only weekly. Atharva redesigned the system end to end. He separated ingestion, feature construction, training, scoring, and audience delivery; rebuilt the pipeline in PySpark and Databricks; introduced richer behavioral and transactional features; evaluated multiple model families and cluster configurations; added data-quality guardrails; and partnered with Marketing on backtesting and campaign validation. Outcomes included: - runtime reduced from more than eight hours to under 30 minutes; - daily audience scoring instead of weekly runs; - predictive accuracy improved from below 60% to approximately 85%; - audience size reduced by roughly 40%; - unsubscribe rates reduced by roughly half; and - revenue per send increased by nearly 50%. The exact accuracy metric is not specified in the source material and should not be inferred. ### Self-Service Ad-Hoc Affinity Platform After Affinity V2 launched, Marketing repeatedly needed custom audiences for emerging categories, campaign ideas, and product trends. Each request previously required bespoke data-science work. Atharva built a self-service audience-generation platform on top of the validated affinity framework. Business users could name a custom affinity and select relevant product attributes while reusing production scoring logic and standards. The focused self-service workflow created audiences in under 10 minutes, compared with up to roughly 30 minutes for a broader full-audience training process. It was adopted across multiple global marketing teams and contributed an estimated $14 million in annual incremental revenue through more responsive personalized targeting. ### Modular, Metadata-Driven Affinity Architecture Affinity definitions had been tightly coupled to legacy product-code logic, so changing categories could require weeks of coordinated work. Atharva moved affinity definitions into metadata, separated concepts from underlying product codes, removed outdated taxonomy logic, and onboarded approximately 40 new affinities. This reduced turnaround for adding or modifying affinities from roughly two to four weeks to under three days and became the standard approach for scaling affinity-based targeting. ### Replenishment Decision System Nike needed a repeatable way to predict when members were likely to repurchase consumable or repeat-use products under a compressed holiday timeline. Atharva set model guardrails with Marketing, built the pipeline from four years of purchase, engagement, and product data, engineered recency, frequency, seasonality, product, and interval features, backtested performance, and productionalized the model. The broader system also supported substitute recommendations when products were unavailable. The system launched for the holiday season, achieved approximately 20-day root mean squared error, produced click-through rates near 5%, delivered roughly four times average revenue per send, and maintained unsubscribe rates near baseline. ### Databricks Runtime Migration A mandatory Databricks runtime upgrade affected a portfolio of production machine-learning pipelines on a compressed schedule. Atharva negotiated testing runway, inventoried affected pipelines, created the migration plan and upgrade guide, piloted early migrations to identify breaking changes, standardized validation, and ran cross-team coordination. All affected pipelines were migrated in approximately one month with zero downtime or material performance regression, and other data-science teams reused the process. Source materials differ on the exact pipeline count, so the count should be verified before making it a headline claim. ### Affinity Growth Forecasting Atharva also served as product lead for an affinity-growth forecasting capability. He interviewed stakeholders, translated planning questions into requirements, defined granularity and cadence, partnered with data scientists on signals and validation, led the roadmap and rollout, and improved forecast error by more than fivefold based on mean absolute percentage error. The forecasts became a recurring input into planning discussions. ### Nike Tools and Themes Python, SQL, PySpark, Spark, Databricks, AWS, Hadoop, gradient boosting, feature engineering, predictive modeling, personalization, experimentation, forecasting, production pipelines, metadata-driven design, and self-service analytics products. ## Black Mountain Systems ### Role and Context Atharva was a Project Manager at Black Mountain Systems from July 2017 through June 2019. He combined client discovery, product scoping, implementation, SQL and workflow analysis, quality assurance, training, and delivery for financial-services software. ### MiFID II Compliance Platform New MiFID II rules required European clients to transmit trade messages within 15 minutes of execution and again at the end of the day. Atharva coordinated regulatory, Product, Engineering, and client stakeholders; defined data and message schemas; implemented formatting, publishing, and parsing logic; built a QA suite; and created customer-facing configuration guidance. The functionality launched on time, his customers went live without material issues on the first day, and his guide became an internal onboarding standard. ### High-Volume Trade Entry Workflow Traders were losing time entering high volumes of trades through a full workflow page with unnecessary navigation and premature data entry. Atharva mapped the desk's workflow, designed and built a lightweight batch-entry widget in the panel traders already used, validated it with users, enforced downstream data requirements, and tested edge cases. It became the default high-volume workflow and reduced manual effort by approximately 10 hours per week across the desk. ### Other Client Systems Work Atharva also reverse-engineered legacy SQL procedures for a complex cap-table import workflow and reset scope and roadmap expectations with stakeholders for a major client implementation. These examples demonstrate legacy-system analysis, testing, stakeholder alignment, and expectation management. ## Education - Master of Science in Applied Data Science, University of Southern California, May 2021. - Bachelor of Science in Applied Mathematics, University of California, San Diego, June 2017. ## Useful Recruiter Questions - What applied-AI systems has Atharva designed? - How has Atharva scaled analyst expertise through reusable products? - What did Atharva personally build versus lead at LinkedIn? - What performance-engineering outcomes has Atharva delivered? - How did Atharva turn Nike's affinity models into a self-service platform? - What examples show technical program leadership or cross-functional influence? - What is Atharva's experience with production ML, migrations, and regulated systems?