AI agent developer · Memory systems · Developer relations
I build agents that use tools, remember context, and ask before they act. My work spans native macOS apps, MCP and WebMCP interfaces, memory governance, durable workflows, and backend engineering. I also turn the failures and design decisions into technical guides developers can use.
I publish as Himanshu Kumar and himanshu748.
Engineering portfolio · Build case studies · DevRel portfolio · Writing · LinkedIn · Resume · Email
OmniDev — a native agent for your Mac
Search files and screenshots with citations, edit code, and run tests through scoped tools and approval prompts. A shared local knowledge engine serves both the SwiftUI app and external MCP clients. Ollama mode runs inference locally; cloud inference is optional.
SwiftUI · Python · FastAPI · SQLite · Ollama · MCP
Engineering case study · Product and real app recordings · Architecture
| System | The engineering problem | Start here |
|---|---|---|
| Graft | Turn semantic HTML into reviewed WebMCP contracts; inspect evidence, export an adapter, and verify the deployed tools. | Case study · Workbench |
| ContextFirewall | Check recalled memory for stale facts, contradictions, secrets and missing evidence before assembling an agent's context. | Case study and evidence |
| VOX | Voice-to-plan support workflows with explicit approvals, persisted results and a spoken briefing. Built with Mastra, Temporal, Gemma and ElevenLabs. | Case study · Judge demo — fictional tickets; customer sends are simulated |
- EmbodiWeave: my agent read its own failure and repeated it — graph-enforced memory, recorded decisions and fixed promotion gates in a robot simulation.
- ToolTrap: my support agent handed a customer a fake phone number — untrusted tool results, held-out cases, legitimate-detail retention and scorer corrections.
- We instrumented an AI agent swarm with SigNoz — what telemetry changed about routing, model budgets and review quality.
- A trust firewall for my AI agent's memory — Cognee's memory lifecycle, provenance and the failures around recall.
- The smolagents bug that retried valid code three times — a reproducible debugging story, recognized as a DEV Summer Bug Smash runner-up.
I help AI and developer-tool teams turn product capabilities into working integrations, technical tutorials, API documentation, sample apps, and inspectable demos. I work across Python, TypeScript, SwiftUI, FastAPI, React, SQL, MCP, WebMCP and OpenTelemetry.
Open to AI agent engineering, developer relations, and technical writing contracts. Tell me what you're building →
- Merged contributions to Kestra.
- Flatpack, a narrated-video-to-instructions prototype, placed 2nd in the Agent Builder track at the Livepeer Agent Hackathon 2026.
- AMD DevMaster 2026 Excellent Award, listed under team himanshu748.
GitHub Sponsors helps support more open-source tools and practical guides. Thank you to Cap, my first GitHub sponsor.