Graph engineering
Research and audit · October 2026Task, team, runtime, and knowledge relationships in agent workflows.
A targeted audit, evidence limits, and proposed tests for dependency semantics and graph publication.
Research case studySandeep Barhanpure
I’m an engineering leader and a hands-on builder. At Walmart, I lead work across Health & Benefits. At Terrain AI Labs, I work on software for real operating problems.
I’ve built products and engineering teams from the ground up, and helped established platforms grow. Across Walmart and MTM Health, my work has included researching opportunities, building the business case, securing executive support and budget, hiring engineers, and leading delivery.
I moved from India to the U.S. in 2012. Today I’m based in Northwest Arkansas. I still like getting close to the code, and I make time to mentor people taking on bigger engineering and product roles.
My background
Work
Walmart2024–present
Director, Software Engineering · Health & Benefits Technology
I lead engineering in Health & Benefits, connecting platform strategy, architecture, compliance boundaries, team ownership, and delivery. The work involves improving established systems while creating room for teams to build and grow.
AWS2021–2024
Head of Engineering · Software & Analytics
I built and led a 40+ person software and analytics organization. We worked across contract systems and customer intelligence, reducing the handoffs where negotiated terms and customer context could get lost.
MTM Health2015–2021
Senior Manager · Software Engineering & Analytics
I joined as the first analytics hire, grew the engineering and data-science function to 30+ people, and helped turn internal analytics into a commercial product. The work connected trip data and dispatch decisions with what operations teams and clients needed to do next.
RBL Bank2010–2012
Head of Pricing & Operations · Consumer & Commercial Banking
I built pricing, reporting, and risk frameworks for consumer and commercial products. The challenge was turning product-specific decisions into operating standards that could be used more widely.
Projects
Traverse
Planned · November 2026A workspace for assigning scoped work to coding agents, following progress, and reviewing evidence.
An early local prototype exists. Controlled agent execution is still under development.
Project planHarness
Implementation + historical testsChecks I built in 2026 for developers who coordinated agent work: which files an agent could change, who owned the diff, and whether review approval matched the current commit.
Workflow: define file scope → assign work → check ownership → review the commit.
How the checks workMonty
ExperimentalForecasts pull-request merge times and daily throughput for engineering planning. Produces distributions and intervals, then compares them with simple baselines.
Individual PR timing roughly tied the baseline. Daily throughput was evaluated on only 10 days.
Method and evaluationZeki
Working local CLIA local knowledge tool that helps our agents find operating rules, recent changes, and handoff context, with references to the source.
Workflow: query → retrieve cited records → inspect the context before continuing work.
Project detailsCI policy enforcement
Implemented · 16 test methodsA pinned Ruby checker evaluates workflow triggers, permissions, action references, paths, and container inputs alongside each repository’s own tests.
Policy surfaces and trade-offsJev
Advisory integration · selected evaluationsAn external-model client shared by hooks, MCP tools, and skills. Advisory output can raise risk, but never lower the existing assessment.
Includes the 40-brief label comparison, 80-pair confusion matrix, and recorded latency summaries.
Evaluation and safety boundaryRadioTour
Static publishing implementationA custom generator and content gates for a small publication, with minimal client behavior and self-hosted fonts.
Build workflow and publication stateTechnical notes
- Testing forecasts against simple baselinesMonty
- Turning task boundaries into checksHarness
- When green tests judge the wrong versionAgent infrastructure
- Checking who receives an agent handoffAgent infrastructure
Research themes
The projects and notes above share six research themes in AI-assisted engineering. Each link goes to the page that shows the work and its limits.
- Commit-bound reviewHarness
- Honest forecastingMonty
- Mutation testingTechnical note
- Agent safety boundariesHarness
- Deterministic runsMonty
- Graph engineeringResearch
About
I moved from India to the United States in 2012. My U.S. career began in St. Louis, then took me to Seattle and Northwest Arkansas. Alongside my work at Walmart, I’m the founder of Terrain AI Labs.
I mentor engineers and product leaders working through larger roles, team decisions, and platform problems.