Jul 2019 – Mar 2020 · 8 mos

• Facilitated the transition of 3 orgs’ production infra to kubernetes, collaborated with managers/directors/architects across the company.
• Improved the quality of infra configuration codebase by eliminating hardcoding of org-specific variables, ensuring compliance to DRY.
• Evangelized open-source tech such as Docker/Kubernetes, wrote docs and took multiple sessions.
• Hiring
• Contributed major improvements in SRE hiring process, collaborated across teams for identifying improvements (directors, managers, recruiters).
• Created a question bank for SRE phone screening, used by all interviewers company-wide.
• Created a pre-screen assignment with solutions and scoring rubrics, used company-wide.
• Interviewed ~50 candidates for Senior SWE positions over a period of 7 months.

May 2018 – Jun 2019 · 1 yr 1 mo · Bengaluru, Karnataka, India

• Problem: CodeSearch is the most frequently used tool at LinkedIn. It ingests code deltas and pushes them to Kafka for reindexing later. The crawler fails frequently and has reliability issues.
• Improved CodeSearch’s crawler module stability (error handling, multithreading issues, root-cause analysis, better metrics for troubleshooting).
• Results: Improvements in the operability of crawler, lesser extraneous alerting, lesser outages.
• Problem: Tools ELK cluster used by multiple teams for critical decision making was a single point of failure. Since all the logs are piped via Kafka it could not recover data for more than 4 days old.
• Wrote a module to capture the metrics of the backup process, improved failovers/recovery, wrote a document on how snapshots work in Elasticsearch under the hood.
• Assisted other teams in troubleshooting their Elasticsearch setups.

Jun 2016 – Apr 2018 · 1 yr 10 mos · Mumbai, Maharashtra, India

• Problem: Design a system which Ad analyzers can use for mapping full-text queries to ad keywords (eg: iPhone → {motorola accessories for sale, refurbished samsung phones}), to display contextually relevant ads on publishers such as WebMD and Forbes.
• Experimented with language models (tf-idf vs BM25), stemmers, improved search relevance (word n-grams, precision/recall, manual analysis), quality vs. latency tradeoffs, performed latency/throughput benchmarking.
• Implemented core modules from scratch
• Ingest: crawls all urls from input queues, processes HTML and populates the search index
• Serving: for serving full-text queries (performs dedup, pagination)
• Quality: for comparing search results with other search engines using cosine similarity
• Problem: HTML web pages contain a lot of extraneous text which is typically irrelevant to the page’s context. Multiple projects across orgs need a system which extracts the ‘textual meat’ of the web page.
• Designed and implemented a context extractor system which uses heuristics based on HTML tags to identify web page boilerplate.
• Improved segregation quality by introducing weights for H1, H2 etc. headings and tags (bold, italics).
• Problem: A lot of web pages differ only in boilerplate elements, and some websites copy content from elsewhere, resulting in similar pages appearing in the search results for a given query.
• Evaluated multiple alternatives - Locality-Sensitive Hashing (LSH) techniques - Simhash, minhash among others. Explored research literature for ideas.
• Fine-tuned simhash parameters to balance false positives and storage costs, integrated with search serving in production.