AI Engineer building agents, LLM systems, and developer tools.
I like working on the systems around the model: agent runtimes, tool execution, context engineering, sandboxes, evaluation, persistence, and the infrastructure that makes AI products reliable.
Portfolio · LinkedIn · X · Email
Woopcode — a terminal-native coding agent built from scratch.
- Multi-provider agent runtime for Gemini, OpenAI, and Anthropic
- Tool orchestration, repository-aware context, streaming, and session management
- Fail-closed shell risk classification and approval policies
- Sandbox-aware execution and real-filesystem testing
- Benchmark-driven context engineering and agent evaluation
Nap — an AI app-building platform where you describe an application and the agent builds it while you walk away.
- Agent runtime + tool execution inside isolated E2B sandboxes
- Durable event log with Postgres + WebSocket streaming
- Snapshot/restore for idle projects
- Token, step, sandbox, and per-user quotas
- Authentication, encrypted API-key storage, cancellation, and recovery
- 2,243 tests across 173 files
AI Agents LLM Systems Agent Infrastructure Context Engineering Developer Tools AI Evaluation TypeScript Python
-
Smaller context doesn't necessarily mean cheaper context.
In Woopcode, a compaction strategy reduced peak prompt size by 36–43%, but rewriting the cached prefix destroyed much of the provider's cache reuse. -
Unrecognized shell commands should fail closed.
For an agent with write access to a repository, treating unknown commands as safe is a dangerous default. -
Durability changes what an agent product can be.
Nap persists events before fanout and snapshots idle sandboxes so a user can leave and return without losing the work or paying for an idle machine.
Building AI agents and developer infrastructure, and looking for AI Engineer / Applied AI / AI Infrastructure / Developer Tools opportunities.
I care most about problems where models meet real systems.
🌐 manasr.dev
💼 LinkedIn
𝕏 @Ragu_dev23
✉️ manasr955@gmail.com