const sarwar = {
role: "Senior Software Engineer · AI Engineer",
experience: "5+ years shipping production SaaS",
location: "Karachi, Pakistan 🇵🇰 — remote, UTC+5",
focus: ["Agentic AI", "RAG & LLM orchestration", "Backend APIs", "Data pipelines"],
stack: ["TypeScript", "Node.js", "React", "MongoDB", "PostgreSQL", "Redis", "AWS"],
currently: "Building AI agents that act on systems — not chatbots that talk about them.",
belief: "AI without reliable infrastructure is a demo. I build the infrastructure.",
};The engineers who matter in 2026 aren't just writing code — they're orchestrating intelligence. I build systems where LLMs, tools and backends work together without a human in the loop.
┌──────────────────────────────────────────────────────────────────────────┐ │ AGENTIC SYSTEM ARCHITECTURE │ ├──────────────────┬──────────────────┬──────────────────┬─────────────────┤ │ LLM LAYER │ TOOL LAYER │ ORCHESTRATION │ BACKEND │ ├──────────────────┼──────────────────┼──────────────────┼─────────────────┤ │ OpenAI · Gemini │ Tool calling │ Multi-step loops │ Node.js / Nest │ │ Anthropic │ MCP servers │ State & memory │ BullMQ + Redis │ │ Structured output│ RAG (pgvector, │ Retry & recovery │ MongoDB / PG │ │ Prompt design │ Pinecone) │ Human escalation │ Webhooks · Cron │ └──────────────────┴──────────────────┴──────────────────┴─────────────────┘
What my agents do in production:
- 🔍 Research & enrich — browse, extract, classify and store structured data with no human in the loop
- 🔧 Act on systems — call APIs, write to databases, trigger workflows based on reasoning
- 📊 Report & summarize — turn raw data into structured, scheduled output
- 🔄 Self-recover — retry on failure, log anomalies, escalate when confidence is low
I don't build chatbots. I build systems where AI is a reliable worker.
| Project | What it does | Stack |
|---|---|---|
| 📧 AI Email Sequencer | Multi-step outbound sequencer with LLM personalization, reply classification and branching follow-ups; inbox warm-up with SPF/DKIM/DMARC and bounce monitoring | Node.js · BullMQ · Redis · OpenAI |
| 📡 Signal Intelligence Engine | Turns unstructured web data into buying and competitor intent signals with LLM summarization for qualification | TypeScript · Scraping · pgvector |
| 🧠 Agent Workflow Kit | Agentic loop with typed tool calling, RAG over a knowledge base, memory and escalation paths | TypeScript · OpenAI/Anthropic · Pinecone |
| Domain | What I ship |
|---|---|
| 🤖 Agentic AI pipelines | Multi-step agents with tool integration — stateful, observable, production-grade |
| 🧠 LLM orchestration | Tool calling, structured outputs, context management, RAG — real flows, not toy prompts |
| ⚙️ Backend engineering | Node.js + TypeScript APIs, queue-and-worker systems, OAuth, rate limiting at scale |
| 🚀 Full-stack products | Schema → API → UI with React and Next.js, built end to end |
| 📈 Data pipelines | Enrichment, caching, batching and ETL over tens of millions of records |
| 🔗 Integrations | CRMs (HubSpot, Salesforce), Zapier/Make, payment gateways, webhooks — fail gracefully, never silently |
If it doesn’t hold under real users, real data and real failure — it isn’t done.
- Agents need boundaries — explicit tool scopes, confidence thresholds, human escalation paths
- Structured outputs, always — if downstream code parses it, it must be typed, not hoped
- Observability first — if you can't see what your agent did, you don't have a system, you have a prayer
- Queues over heroics — idempotent jobs, retries and rate limits beat request-time magic
- Business context matters — I've worked both the engineering and the go-to-market side. I know what "done" means from both.
