Ama Senevirathne

AI / agent engineer · Full-stack software engineer

Discuss an engineering role · LinkedIn · DEV.to · X · GitHub

I build and test the boundaries around AI systems: provider integrations, tool calling, security tooling, distributed backends, and the interfaces people use. My source includes PolyAI .NET, CredScan, and mcp-breakbench.

C# multi-provider SDK with streaming, structured output, and runtime tool discovery. Tool schemas follow runtime types and reject unsupported parameter shapes. README · ToolRegistry · Provider routing

Python scanning for source, staged changes, and agent configuration. It inspects MCP environment values, headers, and arguments, and reads Git index blobs so staged content is scanned as staged. Agent artifacts · Git integration

Python MCP interoperability and regression lab that launches configured servers over stdio, snapshots advertised tool contracts, and runs explicit allowlisted cases with deterministic JSON receipts. README · Runner · Contract diff

Python experiments for coding-agent guardrails: syntax checks, forbidden-phrase checks, and compact execution logs. Implementation

A .NET / Aspire reference backend with an Angular storefront. Checkout state explicitly models stock and payment success or failure; activities and consumers isolate side effects; payment events and a projection marker share one MongoDB transaction. Checkout state machine · Payment event store · Build and test workflow

Angular storefront with signals and HttpOnly-cookie sessions. Auth boundary

C# duplicate-file tooling with reversible quarantine and restore.

Accessibility monitoring with Playwright and axe-core. README

• AI integrations and tools — C# provider boundaries, JSON Schema, streaming, and structured output. Evidence
• Agent security — Python inspection of MCP configuration and staged Git content. Evidence
• MCP interoperability — real stdio probes, contract snapshots, explicit allowlists, and typed receipts. Evidence
• Distributed systems — MassTransit state machines and transactional event storage. Evidence
• Data and recovery — MongoDB event storage and projection markers in one transaction. Evidence
• Frontend — Angular, TypeScript, signals, and a deliberately scoped auth boundary. README · Auth boundary
• Delivery and accessibility — reusable restore, build, and test workflow; Playwright and axe-core monitoring described in the project README. Workflow · README

20 verified merged pull requests across 17 repositories. Live GitHub search

Repositories: apache/lucenenet · apmantza/pi-lens · Arize-ai/phoenix · atretyak1985/swarmery · BerriAI/litellm · calibrain/shelfmark · emdash-cms/emdash · felladrin/MiniSearch · fleetdm/fleet · floci-io/floci-ui · MihaelaAghirculesei/n8n-probe · plur-ai/plur · schubydoo/clauster · theagentplane/tokenops · tphakala/go-remedy · vllm-project/vllm · wemake-services/django-modern-rest

• BerriAI/litellm — Made budget resets invalidate the affected end-user spend counter and cache.
• fleetdm/fleet — Prevented Windows client-certificate validity from being truncated to one year.
• Arize-ai/phoenix — Removed a return from a finally block in playground_users.get_user to resolve a PEP 765 diagnostic.

Sources: public GitHub repository and pull-request records.