Systems live in production right now

I design and ship AI systems that run real operations.

Not demos, not notebooks. A multi-tenant CRM handling real patient cases, an 11-agent AI operating system that runs itself, and growth pipelines that scrape, score and act without me in the loop. I'm a growth marketer who learned to build the machines.

AI orchestration Multi-agent systems NestJS · React · MongoDB Python · Node · TypeScript Growth engineering
Explore the systems → Work with me
saurabh@ai-os ~ %
➜ status --systems
medtour-crm    ● live # real patients
hermes-ai-os   ● 11 agents # always-on
livance        ● 448 pages
job-pipeline   ● nightly
outreach       ● draft-gated
➜ whoami
role = builds AI that ships
based = Dubai, UAE
➜
5+
Production AI systems
11 agents
Autonomous AI OS
170+ ✓
Automated test checks
9+ yrs
Growth + engineering
01 · The thesis

Marketing taught me what to build. AI let me build it.

For nine years I ran growth for brands: $20M+ in ad spend, 24+ companies, healthcare to D2C. That work is a different portfolio. The lesson it left me with: most businesses don't lose on strategy, they lose on execution that doesn't scale: leads that go cold, follow-ups nobody sends, ops that live in one person's head and one spreadsheet.

So I started building the systems that close that gap. Real software, in production, holding real data, plus an AI operating system that runs my own work: agents that research, write, tailor CVs, draft outreach and file everything for approval before anything risky happens.

Claude authors  ·  Hermes operates
Chromium renders  ·  Sheets govern
GitHub records  ·  Saurabh approves

That one line is how everything below is wired. Nothing risky executes on its own; it stops at a file-based approval gate with a five-level task classification. The systems are autonomous where it's safe and human-gated where it counts. That's the whole design philosophy: ship real automation without handing over the wheel.

02 · Featured systems

Five systems, all running.

Each of these is deployed and doing real work · not a proof of concept. Roles range from full-stack engineer to AI architect.

SYS_01 · SaaS · full-stack + AI

MedTour CRM

Sole architect & engineer · greenfield build
Live · real patient data

A multi-tenant SaaS operating system for medical-tourism companies: the full patient journey from first WhatsApp message to aftercare. Built from zero and deployed to production for a real 5–6 person team working live cases at crm.livancehealth.com.

What it does

  • Lead capture + SLA engine · every lead gets a case ID, routes to a coordinator, and fires reminders if unactioned
  • Recommendation + quote workflow · hospital matching, tokenized hospital reply portal so quotes update with zero coordinator typing
  • Follow-up & nurture engine · cadence auto-scheduling, auto-nurture, going-cold alerts, built for weeks-long real sales cycles
  • Country playbooks + complexity scoring · the data moat: 0–100 case scoring routes hard cases to senior staff
  • Patient & partner portals, treatment-plan builder, payments + commissions, 4 role-based dashboards
  • Automatic tenant isolation at the data layer · a cross-tenant fetch returns 404, enforced so no query can leak

Proof

132 ✓
Automated checks green
Live
In production
42
Real cases migrated
175+
Doctors in catalog

Stack

NestJSMongoDBReact + ViteTypeScript JWT + HelmetWhatsApp Cloud APINginx · PM2Hetzner
crm.livancehealth.com ↗
SYS_02 · AI orchestration · flagship

Hermes · a personal AI operating system

Architect of the whole stack
Always-on · self-running

An always-on, 11-agent AI operating system on a cloud VM that runs my actual work. A central orchestrator routes tasks to specialist agents; ~30 cron jobs keep it alive; an Obsidian knowledge vault is its long-term memory. It writes code, runs SEO, hunts jobs, drafts outreach and reports back · and stops at an approval gate before anything risky.

How it's built

  • 11 specialist agents under one orchestrator · SEO, engineering, job-hunting, outreach, branding, research, reporting, automation
  • Approval gate · 5-level task classification; drafts and reads run free, anything that sends/deploys/publishes stops and waits for a human OK
  • CLI bridge routing all model calls through local Claude + Codex CLIs · no API keys, no usage caps
  • Mission Control · a 3-agent command center with chat history to see and steer every agent
  • INDEX-first memory · agents read a 50-line index and fetch one note instead of loading thousands of lines every run

Scale

11
Autonomous agents
~30
Cron jobs
5-level
Approval gate
24/7
Uptime

Stack

Claude + Codex CLINode · PythonCron orchestration Obsidian + MCPGoogle APIsTelegramAzure VMNginx
SYS_03 · AI-run product · growth

Livance · an AI-operated growth platform

Product owner + AI ops architect
Live · agent-run

A medical-travel platform whose content and SEO are run by an AI agent swarm. An SEO Head orchestrates 10 sub-specialists on a weekly wave cycle; a separate engineering agent implements approved specs and opens pull requests. The result is a site that grows itself · with a human only at the approval step.

What the agents run

  • 3-tier international SEO cluster · country → treatment-hub → treatment-leaf, all prerendered and indexable
  • 448+ dynamic pages · 210 doctors, 41 hospitals, 179 treatments · plus a TipTap blog CMS
  • 10-specialist SEO agent · research, technical, on-page, schema, internal-linking, backlinks, AI-search visibility
  • Unit-economics modelling · a 90-day P&L that found the true blended CPL and the single biggest funnel leak

Footprint

448+
Pages prerendered
10
SEO sub-agents
30+
Target countries
Weekly
Autonomous cycle

Stack

React SPANode APIMongoDBPuppeteer prerender GSC · GA4GitHub Actions CICloudflare
livancehealth.com ↗
SYS_04 · Autonomous pipeline

Nightly job-search pipeline

Designer & engineer
Runs nightly

A fully autonomous pipeline that runs every night: scrape → score → tailor → track. It pulls ~100 fresh roles, has an LLM rank each one, then writes a JD-matched CV per strong match (headline, summary and an ATS-clean layout), hosts it, and drops a morning briefing to Telegram. Zero manual steps until I choose to apply.

The loop

  • Scrape · JobSpy pulls ~98 LinkedIn roles a night plus regional boards
  • Score · LLM classifies each role Strong / Moderate / Stretch / Pass against my profile
  • Tailor · auto-generates a per-role ATS CV with a JD-matched headline and summary, then hosts it
  • Track · a Google Sheet is the system of record; nothing is auto-submitted

Throughput

~100
Roles / night
Auto
CV per match
4-tier
LLM scoring
03:30
Telegram briefing

Stack

PythonJobSpyLLM scoringGoogle Sheets APICronTelegram bot
SYS_05 · Multi-channel engine

Outreach engine

Designer & engineer
Draft-gated by design

A draft-only, multi-channel outreach engine across LinkedIn, recruiter email, cold BD and partnerships. Every message is personalised, rate-limited and pushed to an approval queue · Gmail drafts and gated LinkedIn posting, never a blind auto-send. It's the clearest example of the whole philosophy: automate the work, keep the human on the trigger.

What it handles

  • 5 outreach specialists · LinkedIn, recruiter email, backlink, cold email, partnerships, each with follow-up tracking
  • Approval queue → Gmail drafts · nothing sends without a human tick in the control sheet
  • Gated LinkedIn posting via a saved-auth browser automation, only after sheet approval
  • Guardrails baked in · daily rate limits, identity locks, mandatory personalisation

Controls

5
Channels
Draft
Only, by design
Rate
-limited daily
Sheet
Approval control

Stack

PythonPlaywrightGmail APIGoogle SheetsCron
03 · How it's wired

One approval gate. Everything runs behind it.

The systems above aren't five islands · they share a spine. Autonomous where it's safe, human-gated where it counts.

You chat · Obsidian
│
Orchestrator routes by keyword
├──────────┬──────────┤
SEO ×10
Jobs
Outreach
│
Approval gate 5-level classify
│
Sheets govern
GitHub record
01

Cheap by default, deep on demand

Deterministic file reads cost zero model calls; a lightweight classifier routes most tasks to a single executor. Multi-agent "deep mode" is opt-in with a cost check.

02

Nothing risky auto-runs

Sending email, deploying, publishing, DNS changes, financial actions · all classified MANUAL_ONLY. The agent prepares everything and stops.

03

Memory that scales

An indexed knowledge vault means agents stay cheap and consistent across thousands of notes · read the index, fetch one note, act, write the result back.

04

Everything is logged

Google Sheets are the control panels and GitHub is the record. Every approval, draft and deploy is auditable after the fact.

04 · What I work with

The toolkit behind the systems.

AI & agents

  • Claude & Codex (CLI + API)
  • Multi-agent orchestration
  • MCP servers & tools
  • Prompt + workflow design
  • RAG / indexed memory

Backend

  • NestJS · Node · Express
  • MongoDB · schema design
  • Multi-tenant architecture
  • REST APIs · JWT auth
  • Python automation

Frontend & infra

  • React · Vite · TypeScript
  • Nginx · PM2 · deploy scripts
  • Azure · Hetzner · Cloudflare
  • GitHub Actions CI/CD
  • Puppeteer · Playwright

Growth engineering

  • Technical + programmatic SEO
  • GSC · GA4 · attribution
  • CRM & funnel automation
  • Unit economics / P&L models
  • WhatsApp / email pipelines
05 · Let's build

Have a system that should run itself?

I'm open to roles and projects where AI does real operational work · CRMs, agent systems, growth automation. Based in Dubai, working globally.

hello.saurabhvrm@gmail.com LinkedIn ↗
◆ Marketing portfolio ◆ MedTour CRM ◆ Livance