Manvi Sharma

Manvi Sharma

I make AI systems reliable enough to ship — inside industrial plants, healthcare, and regulated enterprises.

A decade of engineering at AWS, Workday and IBM — hardening the mission-critical systems that Fortune 500s can't afford to see fail — now aimed entirely at agentic AI. Most AI pilots don't die because the model is weak; they die at month two, inside a real organization, when no one can prove the system is safe. Turning that proof into something an auditor will sign is the problem I work on.

4 weeks → 15 min · Regulated audit workflow22 agents · Multi-agent orchestration91.6% · MRI classification accuracy0.82 F1 · Clinical entity extraction
Open to Forward Deployed, Applied AI & AI Evaluation roles
Canada / US · TN eligible
View Case Studies→GitHubRead

Featured Projects

4 wk → 15 minEnterprise agentic AI · 22 agents

Agentic Audit Assistant — 22-Agent Orchestration for Regulated Audits

An enterprise agentic system that compressed regulated pre-audit preparation from four weeks to fifteen minutes. A lead orchestrator runs Plan-and-Execute across 22 specialized sub-agents with federated tool and data access, built-in observability, and conversational follow-up — validated in use by the audit team it was built for.

Multi-Agent OrchestrationPlan-and-ExecuteDatabricksAzure AI Search+3
ProprietaryRead Case Study →
200+Sensors · Real-time ML

AI-Driven Anomaly Detection — Industrial Sensor Intelligence

Built an early-warning system for industrial process failures across 200+ sensor channels, combining Isolation Forest, LSTM and Autoencoder ensembles with a Gemini-powered chat interface for operator queries. Shipped a Streamlit monitoring dashboard, a simulated real-time streaming pipeline, and automated alerting workflows, deployed to GCP with a modular Python codebase.

PythonIsolation ForestLSTM AutoencoderTime Series+4
CodeRead Case Study →
91.6%Accuracy · +21% over baseline

Brain Tumor Classification from MRI Scans

Clinical decision-support system achieving 91.6% accuracy on 4-class brain tumor classification - a +21.35% improvement over baseline using transfer learning with EfficientNet-B3 and a custom medical image preprocessing pipeline.

PythonPyTorchComputer VisionMedical Imaging+3
CodeRead Case Study →
View all projects →