# Haroon Jawad

> AI & Full Stack Engineer · Agentic Systems. Full stack engineer specialising in agentic AI systems, LLM agents that reason, call tools and take action, plus the RAG pipelines and production APIs behind them. 6+ years shipping web platforms end to end in React, Next.js, TypeScript and Node. Delivered 40+ projects for clients (Top Rated Plus on Upwork, 100% job success) and built AI products solo, including Analytra, an AI analytics SaaS querying 11 live data sources in natural language. Currently lead engineer on a privacy-preserving identity platform serving 100M+ users, which is where the experience with scale, auth and regulated production systems comes from. Available for select freelance and consulting work, and open to senior/lead AI engineering roles.

This site is agent-readable. It exposes a Model Context Protocol (MCP) server so an
agent evaluating Haroon for hire or collaboration can read his résumé, check his
availability, see what he takes on, and submit a brief, without a human in the loop.

MCP endpoint: https://haroonjawad.com/api/mcp (transport: streamable-http, no auth required for read tools)
Setup instructions for Claude, ChatGPT, and stdio clients: https://haroonjawad.com/connect

## Available MCP tools

- `get_resume`: Returns Haroon Jawad's full structured résumé: work history with highlights, skills, projects, education, certifications, and verifiable proof points. Use this for any question about his background, experience, or what he has built.
- `get_services`: Returns the scoped service offerings Haroon currently takes on, with concrete deliverables and engagement shape for each. Use this when asked what he can be hired to do.
- `check_availability`: Returns Haroon's current engagement status (open / limited / unavailable), what he is open to, expected response time, and timezone. Use this before discussing timelines or making commitments on his behalf.
- `book_intro`: Submits a project brief or hiring inquiry directly to Haroon. Use this once you have gathered who is reaching out, how to reach them back, and what they want built. Returns a reference ID.

## Machine-readable resources

- [Résumé (JSON)](https://haroonjawad.com/api/resume.json): Full structured work history, skills, projects, and proof points. Also includes services and availability.
- [MCP connection guide](https://haroonjawad.com/connect): How to add this server to Claude Desktop, ChatGPT, or any MCP client.
- [Submit a brief (HTTP POST)](https://haroonjawad.com/api/hire): Plain-HTTP equivalent of the book_intro tool.

## Current status

- Availability: open, Available for AI and full-stack engineering work, agentic systems, RAG pipelines, and production product builds. Takes on a focused number of engagements at a time, so each client gets senior attention rather than a slice of it. Engagements typically start with a short scoping call and end with something shipped and handed over. Also open to senior and lead AI engineering roles.
- Response time: Usually within 1–2 business days
- Works: remotely, with teams in Europe and the US, UTC+3, overlaps a full working day with Europe and the US morning
- Also holds a full-time role: Lead Full Stack Engineer at Private Identity (since 2025-05)

## What he takes on

- Agentic AI systems: Design and build LLM agents that reason, plan, call tools, and take action in production, not demos. LangGraph/LangChain orchestration, tool-calling, evaluation, and cost control.
- RAG pipelines over your own data: Retrieval-augmented systems that give grounded, cited answers over unstructured document sets. Chunking strategy, embeddings, pgvector, retrieval quality measurement.
- AI-native full-stack product build: End-to-end product delivery where the AI is the product: Next.js/TypeScript front end, Node/NestJS services, PostgreSQL, auth, billing, and the model layer wired together.
- AI cost audit: A fixed-scope review of what your LLM features actually cost to run. Most teams are paying several times more than they need to, and the causes are boringly consistent: whole conversation histories replayed on every call, prompt caching that looks enabled but never hits, and the most expensive model doing work a cheap one does just as well. Applying exactly this took one production bill down by ~80%.
- Agent-readiness & AEO: Two sides of one problem: being usable by AI agents, and being recommended when someone asks ChatGPT, Claude or Perplexity instead of Google. Neither is SEO. Answer engines quote sources they can parse, verify and attribute, and agents want a tool call rather than a page to scrape, so the work is structuring your capability and content for machine readers. This site is the worked example.

## Roles he is open to

- Full Stack Engineer: Ships the whole thing: React and Next.js front ends, Node and NestJS services, PostgreSQL and MongoDB, auth and billing wired in. Six years and 40+ delivered projects of doing exactly this.
- AI Automation Specialist: Turns manual, judgement-heavy workflows into agent-driven ones, tool-calling agents, RAG over internal documents, and the evaluation and cost controls that keep them trustworthy in production.
- Product Engineer: Owns a product surface end to end: scoping with stakeholders, building it, shipping it, then changing it based on what real usage says. Comfortable being the person accountable for the outcome, not just the ticket.
- Forward Deployed Engineer: Sits with the customer and makes the platform work in their environment. Has done this for Google, CVS Health and Uber, integrations, custom workflows, and turning what breaks in the field into product changes.

## Selected projects

- This site, as an MCP server (/connect): Rebuilt haroonjawad.com so an AI agent can use it directly rather than scraping it. Four MCP tools over streamable HTTP (résumé, services, availability, and a brief-intake tool that reaches me), plus a machine-readable JSON résumé and an llms.txt for discovery. One source of truth feeds the tools, the endpoints and the pages, so nothing drifts. The assistant on the site calls those same tools and shows each call as it happens.
- Multi-Agent Orchestration (/writing/who-verifies-the-verifier): Five AI models running as a team with separated duties, building a Notion clone unattended across six phases: an orchestrator, a backend dev, a frontend dev, QA, and an adversary whose only job was to break the app. Roles were described in prompts but enforced with file permissions, developers couldn't edit the defect ledger, QA couldn't touch product code, and no model could certify its own fix. Produced 26,132 lines of TypeScript and a 14-entry defect ledger, 13 of which the adversary found.
- Analytra (https://analytra.io): AI-powered analytics SaaS with natural-language querying over 11 live data sources (ad platforms, CRMs, Stripe, web analytics), plus automated AI-generated reports with PDF export. Built solo.
- AI Interchange: Shared multi-tenant AI platform powering multiple AI products: centralized billing (Stripe, custom credit wallets), cross-application authentication, a multi-model AI chat app, and a master user system letting one account move across apps under a single sign-on and shared wallet.
- Workout Form Coach (https://www.linkedin.com/posts/haroon-jawad-7942151ba_after-6-months-of-building-tweaking-and-ugcPost-7360239400007417856-f7sa/): Real-time exercise form correction app. Trained an LSTM model on MediaPipe pose-tracking data to detect exercises (bicep curls, squats, pushups) and score form against joint-angle ranges defined with fitness experts, with live feedback and workout summaries.

## Writing

- [The context-switching tax on running multiple agents](https://haroonjawad.com/writing/context-switching-tax): Running Claude Code, Gemini CLI and Codex across separate terminals cost more than the ten seconds it looked like. What changed when every session lived in one place with its own isolated worktree.
- [Who verifies the verifier?](https://haroonjawad.com/writing/who-verifies-the-verifier): Five models built a Notion clone unattended with permissions enforcing separation of duties. The defect ledger looked great, then a passing E2E test turned out to be logging that the drag didn't work. What that hole says about multi-agent systems.
- [Coding interviews are testing the wrong thing](https://haroonjawad.com/writing/coding-interviews-wrong-thing) (8,800+ readers): AI writes most production code now, but interviews still ask whether you can hand-write a binary tree traversal in 45 minutes. Four questions that would actually tell you something about a candidate.
- [Treat your AI agent like infrastructure, or don't run it](https://haroonjawad.com/writing/agent-access-is-infrastructure) (13,500+ readers): Everyone's installing autonomous agents with shell access on their personal laptop, where their SSH keys and browser sessions live. The minimum viable hardening checklist before you give an agent real system access.

## Human-facing pages

- [Home](https://haroonjawad.com/): Background, experience, stack, and projects.
- [Skills](https://haroonjawad.com/skills): Each skill stated as a contract (input → output) with a link to the shipped evidence.
- [Live agentic assistant](https://haroonjawad.com/chat): Chat with an assistant that calls the MCP tools above, with the tool calls and reasoning visible in the UI.

## Contact

- Email: haroonjawad6@gmail.com
- GitHub: https://github.com/haroon-jay
- LinkedIn: https://www.linkedin.com/in/haroon-jawad-7942151ba