Headline

Applied Scientist 3 at Glance | IIT Guwahati

About

Applied Science professional with a Master’s degree in Machine Learning and 5+ years of experience across industry and research. Experienced in developing and deploying ML and deep learning systems across recommendations, ad monetization, speech processing, NLP, computer vision, and reinforcement learning. Proven track record of taking ML solutions from research and experimentation to production-scale systems, including driving a SaaS product from zero to adoption across 40+ enterprise clients. Strong focus on scientific problem-solving, experimentation, and translating ML research into practical solutions for real-world business use cases.

Location

Greater Bengaluru Area

AI stack

claude code

elevenlabs

gemini

DSPy

Guardrails AI

Kimi Code

Work experience

Applied Scientist 3 at Glance, Jan 2026 - Present · 9 mos, Glance is a leading publisher platform that delivers personalised news content directly to smartphone lock screens across millions of devices. I am currently working on ML-driven yield optimisation for Glance's ad inventory by building dynamic floor-pricing models that predict optimal reserve prices per user × ad-unit context, influencing bid behavior to maximize revenue per user across multiple ad touchpoints. Key Contributions: 1. Built end-to-end ML infrastructure on Databricks spanning feature engineering, model training and inference, A/B experimentation, and low-latency online serving to enable rapid production rollouts. 2. Conceptualised a value function approximator approach for bid-price prediction, leveraging user and ad-unit features for robust, distribution-aware estimation. 3. Unified fill probability, bid revenue, and prediction uncertainty objectives into a single policy learning task for one-shot inference, reducing serving latency across large user × ad-unit combinations. 4. Launched LinUCB-based contextual bandits to dynamically adjust floor values based on real-time bid trends, balancing exploration and exploitation for long-term revenue optimisation. Scale: 100M+ daily ad requests · $600K+ daily revenue

Applied Scientist 2 at Glance, Mar 2025 - Dec 2025 · 10 mos, In my previous role, contributed across the personalisation and content serving stack for Glance’s L0 mobile lock-screen surface. I was involved in building large-scale, fan-out-based recommendation systems — internally referred to as “planners” — serving millions of users. Key Contributions: 1. Architected a Kafka-based content ingestion pipeline on Kubernetes processing 20K+ content items/day, with automated feature extraction, entity extraction, quality scoring, and locality-based targeting. 2. Built a large-scale candidate generation pipeline using ScaNN-based ANN retrieval on AlloyDB, combining exploration and diversity strategies across popularity, recency, and content semantics. 3. Engineered user embedding pipelines and feature stores on GCP Bigtable to support downstream ranking and re-ranking workloads. 4. Developed and iterated on heuristic, neural listwise/pointwise, and XGBoost ranking models using user-content affinity signals, deployed as batch Spark workloads. Scale: 4M DAU · 2.5M DEU · 6M Clicks

Data Scientist at NoBroker.com, Sep 2024 - Feb 2025 · 6 mos, NoBroker.com, India’s first proptech unicorn, is a B2C platform that eliminates brokerage by directly connecting tenants, buyers, and owners. Its core business focuses on property discovery and tenant–owner matchmaking, complemented by value-added services across rentals, relocation, home services, and financial products. Key Contributions: 1. Optimised tenant–property matching by integrating skip-gram embeddings and fine-tuning ranking models, achieving a 15% lift in conversion rates while reducing manual recommendation overhead. 2. Developed lead identification and user segmentation models to target premium properties and high-ticket customers for post-paid plans, improving conversion rates from 1.5% to 2% (~35% relative gain). 3. Built BERT-based intent classification models for the NoBrokerHood community forums, enabling personalised nudges and targeted campaigns. 4. Contributed to the development of NB-Estimate, a real estate analytics product that leverages ML to extract key information from transactional sales and rental data, enriching the database to boost website traction and drive redirection to other NoBroker services. https://nobroker.in/nb-estimate

AI/ML Engineer at ConvoZen.AI, Aug 2023 - Feb 2025 · 1 yr 7 mos, Bengaluru, Karnataka, India, Convozen.ai (https://convozen.ai/) is an in-house SaaS product developed at NoBroker, leveraging Deep Learning and Generative AI to derive actionable insights from call center telephony recordings. During my tenure, the product scaled from zero to adoption by 40+ enterprise clients across banking, education, and healthcare. Along with AI/ML focussed contributions, I was also involved in client onboarding, product demonstrations, and driving adoption by collaborating closely with cross-functional product and engineering teams. Key Contributions: 1. Fine-tuned Conformer STT models for 10+ languages with different decoding strategies, achieving WER < 20 on Indian/indic ASR benchmark. 2. Improved Whisper transcription accuracy in multilingual code-switching by integrating custom anti-hallucination logic, significantly reducing WER. 3. Enhanced multilingual TTS by fine-tuning XTTS V2, improving pronunciation and tonal accuracy for Hindi-English code-mixed inputs. 4. Fine-tuned NVIDIA NeMo Diarizer for multi-speaker segmentation, achieving a DER of 36. 5. Architected a real-time speech intelligence framework using Kafka pub/sub to process call center recordings, extracting speech emotion, audio quality, language ID, music endpointing, and speaker demographics via Trillsson and YamNet encoder models. 6. Pretrained XLM-RoBERTa-large with MLM + Contrastive Learning across 10+ languages, creating a joint embedding space, later fine-tuned for retrieval and semantic tagging. 7. Pretrained and fine-tuned T5 models for grammar correction and text standardization tasks. 8. Built a prompt tuning & benchmarking framework using a Leader-Follower LLM strategy with few-shot, CoT, and self-consistency prompting, enabling rapid evaluation aligned with evolving LLM releases.

Associate Data Scientist at NoBroker.com, Jun 2022 - Aug 2024 · 2 yrs 3 mos, Key Contributions: 1. Led UX-driven ML initiatives to boost CTR by developing property nudges such as Premium, Easy Commute, Good Nightlife, and Green Locality tags, alongside automated property video generation pipelines using FFmpeg. 2. Enhanced LightGBM-based Rent & Sale price prediction models for real estate properties to support tenant/owner journey assistance and serving ~20K API requests/day. 3. Maintained and optimised a high-throughput ML-based image processing pipeline 'IRIS' built on microservices architecture, integrating Celery and RabbitMQ for distributed task management. Delivered critical functionalities including spam detection, room categorisation, and object recognition, processing ~40K images/day with <3s per-image latency. 4 .Fine-tuned ViT-base-patch16-224-in21k for multi-class image classification into 20+ room-type categories, achieving >80% precision and recall across all classes. 5. Fine-tuned SSD MobileNetV2 (Large) for watermark detection and competitor logo flagging via object detection, predicting bounding boxes and class probabilities across multiple scales.

Graduate Teaching Assistant at Indian Institute of Technology, Guwahati, Jan 2021 - May 2022 · 1 yr 5 mos, Guwahati, Assam, India, Teaching assistant for EE322M (Fall 2021) and EE102 (Winter 2021) courses offered by EEE department.

Education

M.Tech, Machine Learning and Signal Processing, Indian Institute of Technology, Guwahati

B.Tech, Electrical and Electronics Engineering, Vellore Institute of Technology