Hi, I'm Shiv

• Revamped research agent architecture, slashing API calls by 67% (from 3 to 1 per query), shrinking operational costs by 90% (to $0.003 per request), and accelerating response time by 75%.

• Designed and deployed three specialized research pipelines for distinct query categories, improving retrieval quality, response relevance, and overall research coverage across diverse user requests.

• Built a comprehensive LLM evaluation framework featuring a two-stage "LLM-as-Judge" architecture with Perplexity Sonar Pro for real-time web fact verification and GPT-5 for structured assessment of hallucination, factual accuracy, clarity, and FDA-style content quality metrics. Enabled parallel A/B testing across up to 8 model variants.

• Contributed to data analysis and model benchmarking that guided the selection of an improved AMD prediction model, achieving sub-100 ms inference latency with reliable predictive performance.

• Engineered cross-service credit deduction pipeline across three microservices using checkpoint-based aggregation and idempotent upserts, eliminating double-charge drift with balance checks refreshed within 60 seconds.

• Collaborated and delivered 15+ production enhancements to React.js interfaces for job recommendations and worker dashboards, reducing First Contentful Paint from 1.4s to 0.9s and Time to Interactive from 2.4s to 1.6s, achieving Google Lighthouse performance score of 92/100.

• Integrated FastAPI-based AI services with frontend components to enable intelligent task matching, boosting automated assignment accuracy by 18% and reducing manual intervention overhead by 30 hours/week, in collaboration with cross-functional teams.

• Engineered cross-service credit deduction pipeline across three microservices with a 60-second cache-backed balance check, using checkpoint-based aggregation to unify 5 enrichment and 8 research credit types under one consistent ledger.

• Spearheaded development of Siamese Neural Network in PyTorch for patient verification system, achieving accuracy rate of 98.6% through rigorous testing and validation of core ML components across 5,000+ patient records.

• Engineered backend services with Node.js and Express.js, designing hospital system integration architecture that enabled 400+ daily patient transactions while maintaining 99.5% system uptime.