Devyanshi Singh profile photo 1 on Tal

Devyanshi Singh

Data & Applied Scientist 2, Microsoft AI • ex-IBM • IITK’22

Bengaluru

4+ years of work experience

Microsoft
Mar 20251 yr 6 mosPresent

Vertical Ads Team, Microsoft AI — Working on ad selection, ranking, and monetization systems for Microsoft's Vertical Ads platform. - Architected a unified offline evaluation pipeline for Filtered-DiskANN ad selection, spanning query-ad location relevance, precision, recall, hitrate, and multi-intent analysis — used to calibrate the segment-agnostic VA Selection model - Oversaw quality analysis and production integration of SLM-generated AdProfile attributes (page summaries, SEO keywords, dynamic callouts) for ad URL enrichment across the ad stack - Drove end-to-end experimentation for AdProfile keywords in IRMatch & DiskANN Selection — from feature engineering and index flighting to defect tuning and prod code changes, delivering significant RPM uplift - Pioneered LLM-generated contextual keywords for Professional Services vertical selection, driving measurable CTR-driven RPM gains - Launched an image quality assessment pipeline for Hotel Ads landing pages, unlocking a new revenue stream from previously unservable ad copies
Microsoft
Jul 20222 yrs 9 mosMar 2025

Vertical Ads Team, Microsoft AI — Owned the pClick prediction stack and LLM-powered quality pipelines for Microsoft's global Vertical Ads platform. - Engineered an end-to-end global pClick prediction framework: training data pipeline, automated model refresh, online-offline parity checks, and Treatment vs. Control cross-evaluation for any new feature - Drove consistent monthly CTR, GoodClickYield, and offline AUC gains through automated model refresh cycles - Iterated on the global pClick model incorporating SLM-generated query-profile signals, page position features, and seller bias corrections - Built an LLM-powered pipeline to detect and auto-correct invalid advertiser feed entries, expanding ad coverage for Professional Service Ads - Designed a GPT-based Defect Labelling system for query intent, query-ad relevance, and ad-image relevance classification (Hotel Ads) - Stood up two cold-started verticals end-to-end — Doctors & Clinics and Lawyer & Legal Services — including multilingual intent triggering, allow/disallow lists, and pClick model.
Microsoft
May 20213 mosJul 2021

Bengaluru, Karnataka, India

• Scrutinized Cross-lingual embedding algorithms to transfer NER Ad-Offer Tagger to low-resource languages in a zero-shot way • Modified MUSE & RCSLS algorithms on AEther, educed ‘pseudo parallel’ data using SCOPE, & visualized improved alignments • Obtained remarkable Precision up to 0.96 & F1 Scores up to 0.82 for FR & DE attributes, enhancing relevancy of Bing-Ads Filters
IBM
Jan 20214 mosApr 2021

Bengaluru

- Optimized Apache Spark's workload scheduling using reinforcement learning - Enhanced Spark's hashmap with Locality-Sensitive Hashing (HashAlign, Tree LSH, SimHash) and deployed a DDQN-based RL model for dynamic scheduling - Achieved ~27% throughput improvement over baseline; recommended for production by IBM senior partners in collaboration with IBM Research, Zurich
GitHub
@dev-yan-shi

Devyanshi’s AI Stack

claude code
claude code
claude
claude
codex
codex
github copilot
github copilot
grok
grok
hugging face
hugging face

Education

Indian Institute of Technology, Kanpur

Bachelor of Science - BS, Mathematics and Scientific Computing

Indian Institute of Technology, Kanpur

20184 yrs2022

Higher Secondary Education

Delhi Public School, Eldeco, Lucknow

20089 yrs2017

Join Devyanshi and 10k+ other professionals on

Tal Boss

Tal is the app where bosses meet engineers & builders directly, without middlemen

Download Tal Boss

Devyanshi Singh on Tal

Headline

Data & Applied Scientist 2, Microsoft AI • ex-IBM • IITK’22

About

I build the ML systems behind the ads you click - and the ones you don't (because relevance matters). Over the past 4 years at Microsoft Ads, I've navigated the full advertising ML stack - from triggering & selection to relevance, ranking, allocation, and pricing - with deep hands-on expertise in intent triggering, ad selection, and pClick prediction. I build the systems that keep the ad platform's trifecta in harmony: user experience, advertiser conversions, and the publisher's revenue humming quietly in the background :) Currently exploring: Agentic AI workflows, multi-model orchestration, and how the next generation of AI tools reshapes the data scientist's craft. Always happy to connect with people solving interesting problems!

Location

Bengaluru

AI stack

claude code

claude

codex

github copilot

grok

hugging face

Work experience

Data and Applied Scientist 2 at Microsoft, Mar 2025 - Present · 1 yr 6 mos, Vertical Ads Team, Microsoft AI — Working on ad selection, ranking, and monetization systems for Microsoft's Vertical Ads platform. - Architected a unified offline evaluation pipeline for Filtered-DiskANN ad selection, spanning query-ad location relevance, precision, recall, hitrate, and multi-intent analysis — used to calibrate the segment-agnostic VA Selection model - Oversaw quality analysis and production integration of SLM-generated AdProfile attributes (page summaries, SEO keywords, dynamic callouts) for ad URL enrichment across the ad stack - Drove end-to-end experimentation for AdProfile keywords in IRMatch & DiskANN Selection — from feature engineering and index flighting to defect tuning and prod code changes, delivering significant RPM uplift - Pioneered LLM-generated contextual keywords for Professional Services vertical selection, driving measurable CTR-driven RPM gains - Launched an image quality assessment pipeline for Hotel Ads landing pages, unlocking a new revenue stream from previously unservable ad copies

Data and Applied Scientist at Microsoft, Jul 2022 - Mar 2025 · 2 yrs 9 mos, Vertical Ads Team, Microsoft AI — Owned the pClick prediction stack and LLM-powered quality pipelines for Microsoft's global Vertical Ads platform. - Engineered an end-to-end global pClick prediction framework: training data pipeline, automated model refresh, online-offline parity checks, and Treatment vs. Control cross-evaluation for any new feature - Drove consistent monthly CTR, GoodClickYield, and offline AUC gains through automated model refresh cycles - Iterated on the global pClick model incorporating SLM-generated query-profile signals, page position features, and seller bias corrections - Built an LLM-powered pipeline to detect and auto-correct invalid advertiser feed entries, expanding ad coverage for Professional Service Ads - Designed a GPT-based Defect Labelling system for query intent, query-ad relevance, and ad-image relevance classification (Hotel Ads) - Stood up two cold-started verticals end-to-end — Doctors & Clinics and Lawyer & Legal Services — including multilingual intent triggering, allow/disallow lists, and pClick model.

Data and Applied Scientist Intern at Microsoft, May 2021 - Jul 2021 · 3 mos, Bengaluru, Karnataka, India, • Scrutinized Cross-lingual embedding algorithms to transfer NER Ad-Offer Tagger to low-resource languages in a zero-shot way • Modified MUSE & RCSLS algorithms on AEther, educed ‘pseudo parallel’ data using SCOPE, & visualized improved alignments • Obtained remarkable Precision up to 0.96 & F1 Scores up to 0.82 for FR & DE attributes, enhancing relevancy of Bing-Ads Filters

Software Development Intern, Area631 at IBM, Jan 2021 - Apr 2021 · 4 mos, Bengaluru, - Optimized Apache Spark's workload scheduling using reinforcement learning - Enhanced Spark's hashmap with Locality-Sensitive Hashing (HashAlign, Tree LSH, SimHash) and deployed a DDQN-based RL model for dynamic scheduling - Achieved ~27% throughput improvement over baseline; recommended for production by IBM senior partners in collaboration with IBM Research, Zurich

Education

Bachelor of Science - BS, Mathematics and Scientific Computing, Indian Institute of Technology, Kanpur

Higher Secondary Education, Delhi Public School, Eldeco, Lucknow

Public links

github: https://github.com/dev-yan-shi