Senior Software Engineer with 7+ years shipping production web apps (Angular, React, Vue, Node.js, ASP.NET Core), now going deep on TypeScript RAG systems, agentic pipelines, and LLM evaluation. Based in Lahore, Pakistan (UTC+5), working remotely.

I read library source before I trust it. That habit has produced 52 merged pull requests across 29 open-source projects (counted November 2026), including BeeAI, Pinecone, Weaviate, USearch, Immer and Google's ADK and A2A SDKs. Full list with links.

• i-am-bee/beeai-framework: #1852 keep the category and link data and encode the query in ArXivTool; #1854 round the retry delay after scaling it to milliseconds.
• open-multi-agent: #652 keep parallel streamed tool calls apart in the Azure and Copilot adapters.
• pinecone-io/pinecone-ts-client: #446 parse the final streamed event when it has no trailing newline; #434 retry transient connection failures.
• a2aproject/a2a-js: #701 forward pageSize and pageToken for push-notification config listing.
• unum-cloud/USearch: #813 validate nested vector batches in the JavaScript bindings.
• weaviate/typescript-client: #471 map the Boost decay field to decayValue and remove a debug log.
• google/adk-js: #1135 send array query parameters as repeated keys in RestApiTool; #1136 require a line of exactly --- as the SKILL.md frontmatter delimiter.
• milvus-io/milvus-sdk-node: #618 f32ArrayToInt8Bytes sent the whole backing buffer of an Int8Array view.
• immerjs/immer: #1314 draft relocated base objects at previously assigned indices in the array-methods plugin.
• samanhappy/mcphub: #1447 handle nullable and union tool parameters.

Currently in review: VoltAgent #1510 (tiktoken offsets drifted on multi-byte text, so chunks overran maxTokens) · LangChain.js #12131 (TokenTextSplitter cut multi-byte characters in half) · cron-parser #453 (addHour stuck across a sub-hour DST fall-back)

All TypeScript, with real test suites and live demos. No framework hiding the mechanics.

• MELAI-Engineering-lab: an interactive lab to observe, compare, evaluate and debug AI systems, with a CI gate that fails on a regression eval.
• agent-loopguard: a loop and budget guard for AI agents, published on npm as @osama.ansar/agent-loopguard.
• refract-rag: hybrid BM25 + HNSW retrieval fused with Reciprocal Rank Fusion, multi-hop agentic decomposition, and a UI that visualizes every retrieval score live. → demo
• rag-eval-harness: an LLM-judge eval harness for RAG pipelines (faithfulness, relevance, context precision/recall) with a CI regression gate. → live report
• agentic-rag: multi-hop agentic RAG that decomposes questions, retrieves per sub-question with from-scratch BM25, checks sufficiency, and synthesizes a cited answer.
• hnsw-vector-index: HNSW approximate-nearest-neighbor search implemented from scratch, benchmarked against brute-force (recall@k: 66% → 100% as efSearch increases).
• doc-chat-citations: a docs chat assistant where every claim cites and jumps to its exact source passage.
• generative-ui-chat: a chat assistant that streams back real interactive React components (charts, tables, booking cards) instead of plain text.

TypeScript Node.js Angular React Next.js Vue ASP.NET Core PostgreSQL AWS Docker Kubernetes Terraform Vercel AI SDK LangChain.js Vector Databases (Qdrant, Weaviate, Pinecone, Supabase) Vitest

Portfolio · LinkedIn · osama.ansar.pk@gmail.com