Hi, I'm Bharat

Apple (Contract)

Building LLM and agent infrastructure across three production systems: an MCP-served analysis agent, a retrieval-backed engineering assistant, and an internal developer platform.

• Built and productionized an agentic system automating hardware reliability analysis end to end — test data in, statistical strength and stress models fitted (Weibull/lognormal, censored data), predicted defect rates in DPPM, finished report out — architected against a shared processor interface, cutting container build times from 30+ to 15 minutes
• Shipped it as an MCP server with six composable agent skills, then a skill-authoring tool letting reliability engineers across multiple product lines contribute failure models without platform-team involvement — users became contributors
• Established the LLM evaluation practice for a production failure-mode-analysis assistant: Langfuse tracing (per-request cost, latency, prompt versions), LLM-as-judge metrics, reference-free RAGAS scoring, and A/B tests with and without a root-cause knowledge base — measuring what retrieval contributed instead of assuming it helped
• Maintained that assistant's RAG pipeline (Milvus ingestion, document upload, embeddings) through an upstream embeddings-provider migration, and closed three prompt defects — including a model-version regression duplicating four column values on every generated row
• Replaced dynamic code execution in the reliability agent with a declarative rule engine, reading numeric values at >90% accuracy against engineers' hand-computed values; still live 8 months later
• Co-delivered automated PII/PHI scanning and LLM-egress redaction on an internal app-hosting platform, owning the source-code path as two independently feature-flagged capabilities, with an A/B comparison against Presidio to measure false-positive rates
• Designed and co-built static application security testing on every app publish, running asynchronously so no publish waits on a multi-minute external call
• Redesigned the platform file-storage SDK with typed errors, structure-preserving folder uploads, and per-user tokens brokered by a secrets sidecar, keeping credentials out of application containers

Georgia Institute of Technology

Benchmarking AI developer productivity tools across SDLC phases.

• Benchmarked GitHub Copilot, Codeium, and Continue (GPT-4o) across SDLC phases on a Spring Boot codebase, measuring a 20% developer-productivity gain
• Evaluated tools across System Architecture, Feature Development, and Testing phases

Mad Street Den

Intelligent document processing pipelines and recommendation systems.

• Improved question-recommendation accuracy 2-3x through data-centric changes to an A/B testing pipeline on Amazon Redshift, Elasticsearch, and S3
• Raised field-level confidence-score coverage 50%+ by adding Google Cloud Vision OCR and Amazon Textract to a document processing pipeline deployed on Kubernetes

ServiceNow

Full-stack product engineering across Field Service Management, performance, and in-product ML.

• Cut an enterprise scheduling calendar's load time 6x (30s → 5s) for a leading European travel company — timing every business rule and Script Include with the Transaction Monitor to find a per-day loop re-fetching the entire week's events, then removing the cleanup pass that existed only to hide it
• Enabled agent-to-agent parts transfer in Field Service Management by modeling each agent's vehicle as a personal stockroom, so Asset Management's existing Transfer Order workflow applied unchanged — cutting tracked travel duration to customer site 20%
• Trained and shipped a Predictive Intelligence similarity model surfacing related knowledge articles to field agents in-app, matching on work-order descriptions and tuned via the solution definition's word corpus — raising dwell time 5-10%
• Won an internal hackathon against 12 teams with a hands-free field-agent prototype on Dialogflow that advanced Work Order states from speech; the capability shipped into the internal Virtual Agent
• Made field-service ETAs traffic-aware by threading live-traffic durations from the Google Maps Distance Matrix API through a Java service to its JavaScript client — correcting estimates ~20% against the traffic-blind averages used before — behind a toggle, since traffic-aware calls cost more
• Brought PDF invoice previews to the field-service mobile app by adapting HR Service Delivery's Document Templates engine to FSM with its owning team, merging timesheet, transport, and parts charges into one translatable document — and showed the invoice before signature capture, where mobile had shipped a signature pad only
• Rebuilt a Performance Analytics dashboard interface from Angular to React, shipped to a 10-customer beta program
• Migrated customer and consumer portal experiences to meet WCAG 2.1 level AA, standardizing page structure and keyboard/screen-reader behavior across both surfaces

ConcertAI

Clinical trials search engines using NLP techniques.

• Improved retrieval precision by 40% extracting biomarkers, stage, and ECOG from clinical trial records
• Built a Clinical Trials Database search engine

Georgia Institute of Technology

AI Specialization

GPA: 4.0/4.0

BITS Pilani, Hyderabad

GPA: 9.13/10.0