education

msc advanced computer science (ai)

university of leeds | september 2025 – september 2026

b.tech ai & data science

viit, pune | january 2022 – may 2025

experience

ai engineer · hopps

september 2024 – june 2025 | pune, india

developed ai recommendation system using ::b::gemini api:: and ::p::gcp cloud functions::, delivering real-time dining suggestions for ::y::70+ restaurant listings:: across beta testing phase.
implemented ::y::objectbox local caching strategy::, reducing firestore read operations by ::y::98.6%:: (from 70 to ~1 read per user) and optimizing api costs against a 1,000 mau baseline.
built serverless backend using ::b::firebase cloud functions:: to handle ai content generation pipeline for both consumer and partner-facing applications.
prototyped ::p::embedding-based retrieval system:: to llm based recsys, achieving ::y::60-80% api latency reduction:: (5s to 1-2s) and enabling potential on-device ml inference for cost optimization.
engineered api safeguards including ::y::rate limiting:: (3 requests/user/day), quota management, api key rotation, and graceful degradation patterns to maintain service reliability under free-tier constraints.
coordinated cross-functional delivery across ::y::8-member team:: spanning engineering, operations, and business development, translating ai capabilities into business value for restaurant partners.
mentored 2 interns for 3 months on ::b::llm-to-embedding migration strategies:: and synthetic data pipeline design using gemini through regular briefings.
resolved malformed llm output issues through ::p::iterative system prompt refinement::.
implemented comprehensive error handling and debug logging across all cloud functions, enabling systematic troubleshooting and monitoring via ::y::gcp ai playground dashboards::.

research intern · iim udaipur

september 2023 – april 2024 | remote

built automated classification system for ::y::250,000+ nft records:: using hugging face to differentiate ai-generated vs. human content.
engineered asynchronous scraping infrastructure (::b::beautifulsoup4::) processing ::y::113,000+ web pages::, optimizing pipeline throughput.
explored distributed scraping architecture on ::p::gcp cloud functions:: for ip rotation, demonstrating initiative despite resource constraints.
contributed to ::y::two research manuscripts:: by developing automated data preprocessing workflows and applying nlp techniques for market analysis.

ai intern · konverge.ai

august 2023 – october 2023 | nagpur, india

applied statistical modeling for customer segmentation, purchase prediction, and profitability analysis.
built ::b::sentiment classification model:: using nlp to extract actionable insights from user review data.

projects

vocabscapes · multi-agent puzzle generation system live demo

architected production ::p::5-agent langgraph pipeline:: (writer, critic, improver, finalizer, explainer) on ::b::google cloud run:: with jwt authentication, serving ::y::12+ restful endpoints:: and pre-generating 140+ puzzle variants.
implemented ::y::8-point validation system:: ensuring lexical integrity and contextual quality with 2-cycle average validation convergence across puzzle generation workflow.
achieved ::y::95% reduction in api overhead:: for 1,000 mau benchmark by architecting a ::b::postgresql/supabase:: threshold-based caching layer that decoupled operational costs from user growth.
engineered database architecture with ::y::row level security (rls):: and upsert patterns for data isolation and conflict resolution.
developed a scalable relational backend with ::b::fastapi and supabase::, optimizing for asynchronous performance and relational data integrity to support multi-agent generation workflows.

maira · context-aware ai companion with memory architecture live demo

engineered ::p::rag-based memory system:: using ::b::sqlite embedding storage:: with ::y::200-500ms retrieval performance::, enabling context-aware conversations through semantic similarity search.
designed parallel memory pipeline using ::y::google gemini api::: main thread for responses, background thread for memory extraction with ::b::importance scoring and redundancy detection::.
built full-stack mobile application with firebase authentication, tracking ::y::emotional states and life events:: for realistic, judgment-free companionship.
implemented ::p::imessage-style ux:: focusing on long-term context persistence and judgment-free companionship.
implemented ::y::privacy-first architecture:: storing all memories locally on-device rather than cloud storage, ensuring user data remains private.

privacy-preserving fraud detection via federated learning overview

architected ::p::federated learning system:: for privacy-preserving training across ::y::3 distributed clients over 10 rounds::.
engineered pipeline for ::y::90%+ class imbalance:: (ethereum transactions) using ::b::smote:: and hypothesis-driven feature engineering.
achieved ::y::99.7% auc-roc and 96.6% recall:: using a custom fraud detection classifier.
built ::b::flask-based dashboard:: for real-time monitoring of model performance metrics and fraud patterns.

engineered tiktok-style recommendation engine using ::b::faiss indexivf:: for semantic video retrieval.
implemented ::p::approximate nearest neighbor search:: for ::y::3072-dimensional embeddings:: across 500+ objects.
designed ::y::sliding window model:: (last 5 interactions) with ::y::70-30 exploitation-exploration split:: for content discovery.
built multi-signal scoring system tracking ::b::6 interaction types:: (view time, shares, rewatches) to capture nuanced user preferences.

brewstories · llm fine-tuning for creative writing

fine-tuned ::b::tinyllama (1.1b):: on a ::y::10,000-story dataset:: using pytorch and hugging face, completing training in 8 hours.
developed end-to-end pipeline: tokenization, dataset preparation, fine-tuning, and evaluation for narrative coherence.
achieved ::p::runner-up recognition:: at visionary techfest 24 for stylistic consistency in creative text generation.

undp climate survey analysis

developed ::b::random forest classifier:: achieving ::y::89.76% auc-roc and 84.51% accuracy:: predicting high climate commitment support (≥85% threshold) across ::y::73,000+ respondents from 73 countries::.
built ::p::linear regression model:: achieving ::y::r² 68.23% and rmse 8.47:: for continuous support prediction, comparing interpretability vs accuracy trade-offs across model types.
engineered feature extraction pipeline for ::y::20,000+ null values:: using standardscaler and onehotencoding.
segmented countries into ::b::6 typologies:: using k-means clustering and pca dimensionality reduction.
created ::y::tableau visualizations:: to communicate global age-gap trends and demographic predictors.

research

“attention span and its correlation with mental health”

richa shah, sarthak rak, ayush patne, dr. laxmi bewoor

4th asian conference on innovation in technology (asiancon), august 2024

technical skills

languages: python (pandas, numpy, scikit-learn), dart (flutter), sql
ai & ml: pytorch, tensorflow, hugging face, langchain/langgraph, rag, generative ai, llms, federated learning, fine-tuning (lora/peft)
web & backend: fastapi, flask, firebase (auth, firestore, hosting), supabase (postgresql)
cloud & devops: gcp (cloud functions, cloud run), aws, docker, git, ci/cd, tableau

honors & certifications

runner-up, ai/ml domain – visionary techfest 24 hackathon (binghamton university & visionary club)
machine learning specialization – stanford online & deeplearning.ai
generative ai with large language models – deeplearning.ai & amazon web services
aws cloud quest – cloud practitioner & genai (in progress)

leadership & extracurricular

hackathon lead organizer – viz-a-thon, viit
video & photography team head – ai student association, viit & indo-japan hyper-k workshop
media team head – tedxviit
class representative – academic year 2021-22