Key takeaways
- Suzu Technology is run by one datacenter architect and a set of AI agents that do the repeatable work: this website (137 URLs, 44 free tools) was designed, coded and deployed in days, not months, with a Python generator and a Git-based deploy.
- The same pattern used in enterprise infrastructure — a tool server the LLM can call, with authentication and an audit log — is what makes AI useful for operations. Chat alone does not run servers; MCP tool servers do.
- A scheduled agent researches the last 72 hours of vendor news every morning, writes a sourced article in the house style, builds the site and pushes it. A human reviews; nobody writes from a blank page.
- Browser automation (Chrome driven by an agent) turns multi-day ops chores — OEM warranty lookups, quote comparisons — into a background job.
- The real cost of this stack for a small Indian business is roughly ₹2,000–4,500 a month in AI subscriptions plus a Hostinger plan; the real constraint is knowing what to automate first.
- Everything below is reproducible. The section on "where to start" is for founders and engineers who want the same setup.
1. Who is writing this
I am Sushil Kumar, founder of Suzu Technology. For fifteen-plus years my job has been datacenter architecture: target-state designs, SAN and backup estates, HA/DR to RPO and RTO numbers, ISO 20000 and 27001 programmes, and the migrations between VMware, Hyper-V, OLVM and Proxmox that keep CFOs calm. Much of that work was inside India's state data centers and e-governance programmes — Punjab, Karnataka, Haryana, Himachal Pradesh, a greenfield design in Gujarat — where an outage is a newspaper story. The full profile is on the About page; employer details are deliberately not on this site.
In parallel I built two small businesses: Suzu Travels, a Himachal destination management company, and SuzuCart, an e-commerce store. Both are real, with staff, payroll and ad budgets. They are also my lab. Everything Suzu Technology sells under AI & automation and digital marketing was first run on my own money.
Suzu Technology exists because the lab worked better than expected. One architect, working with AI agents as colleagues, now produces what used to need a five-person team. This article explains the stack, honestly, including what still needs a human.
2. The stack, in one table
| Layer | What runs it | What it does | Human role |
|---|---|---|---|
| Thinking partner | Claude (chat, Cowork mode, Claude Code) | Architecture reviews, HLD/LLD drafting, code, copy, analysis | Sets the brief, reviews output, owns decisions |
| Tool access | MCP servers (infrastructure, WordPress, Meta Ads, GitHub, browser) | Lets the model do things — query hosts, publish a post, read ad spend, push a commit | Grants scoped credentials; approves write actions |
| Scheduled agents | Cloud-run tasks on a cron | Daily Insights article; planned: weekly site QA, nightly tools health | Reads the morning report; edits if needed |
| Build & deploy | python3 build/gen.py → GitHub main → subtree split to deploy → Hostinger auto-deploy |
Turns markdown and Python page definitions into 137 static URLs with schema, sitemap and feed | Reviews diffs; releases |
| Business software | React + Node.js + SQLite app; Excel dashboards | Leads, tour operations, payroll, accounts, CEO dashboard for Suzu Travels | Uses it daily; requests changes in plain language |
| Content & video | Generative video tools with locked AI avatars; Meta Ads via MCP | A Himachal places-and-stories reel series; lead-gen campaigns | Picks topics, checks facts, approves spend |
The row that matters most is the second one. A model that can only talk is an expensive intern. A model that can call get_uptime, list_servers or patch_compliance_report over an authenticated tool server, and whose every call is logged, is an operations engineer on the night shift.
3. Infrastructure agents over MCP
The Model Context Protocol is a simple idea: you expose functions ("tools") over a server; the LLM discovers them and calls them with structured arguments; results come back as text or JSON. I have built MCP tool servers for infrastructure that expose read-only checks across a fleet — uptime, OS inventory, multi-host patch-compliance reports, Proxmox cluster and Proxmox Backup Server status — with Bearer-token authentication, a least-privilege service account, and every call written to an audit log. The dashboard and the MCP server share one tools.py, so a human in a browser and a model in a chat see the same truth.
Three rules from doing this in production-like environments:
- Read before write. Ship read-only tools first. Add power/create/delete tools only once you have the audit trail and a change process that the agent must follow (ours: the agent prepares the change record; a human approves it in the CAB).
- Containerise the server, keep the gateway thin. The tool server should be a Docker image with its credentials in a secret manager, not a pet VM with keys in a config file. Access to on-prem hosts goes through one SSH proxy you already monitor.
- The model does not get the jump host password. It gets a scoped account that can run exactly the commands the tools wrap. That is what makes "AI in operations" survivable in an ISO 27001 audit.
The same design is now a Suzu service: we build the tool server for your estate, connect it to the chat platform your team already uses, and hand over the runbook.
4. The content engine that wrote this site's blog
Every morning a scheduled agent wakes in the cloud, reads the repository's house-style notes, lists the articles already published so it does not repeat a topic, searches the last 72 hours of vendor blogs, release notes and advisories, picks one topic, and writes 900–1,400 words with inline links to primary sources. It then runs the static-site generator, checks that the URL count went up by one and that the HTML rendered, commits with a conventional message, pushes main, updates the deploy branch, and appends a line to a content log. If anything blocks it — a push rejected, a build error — it notifies my phone with exactly what failed.
The guardrails are in the prompt and the repo, not in hope: no invented numbers, no copied text, a required "Key takeaways" block, a comparison table where sensible, an India angle with costs in rupees, category rotation across days, and a hard rule on what may never be mentioned. This morning's article on the Next.js security update came out of that pipeline with nine sources; I read it over chai.
What it cannot do: decide what Suzu should say about a topic. Opinion, pricing and promises stay human.
5. Browser automation for the work nobody wants
Some tasks have no API. Checking warranty and support status for a large server estate on an OEM portal means a serial number at a time in a web form. With Chrome driven by an agent, that became a background job producing a spreadsheet, with the misses flagged for a human — days of clicking turned into a coffee break. The same approach compares vendor quotes line by line, drafts the clarification emails, and writes the comparison workbook the finance team asked for.
The rule here is the same as for infrastructure: the agent reads and compiles; a person signs the purchase order.
6. What it costs in India, and what it saves
| Item | Monthly cost (approx.) | Notes |
|---|---|---|
| AI assistant subscription (Pro/Max tier) | ₹1,700–₹17,000 | Billed in USD; GST applies. The lower tier is enough for most founders. |
| Hosting (shared business plan, 100 sites) | ₹300–₹700 | One plan runs suzutechnology.com, suzutravels.com and suzucart.com. |
| GitHub, GA4, Search Console, PageSpeed API | ₹0 | Free tiers. |
| Generative video tools | ₹0–₹2,000 | Depends on the plan and how many reels a month. |
| Ads MCP connector / scheduler | ₹0–₹1,500 | Free tiers exist; paid for scale. |
Against that: a three-page brochure site from a Noida agency starts near ₹50,000; a "standard" business site with blog and SEO setup is commonly quoted at ₹1–1.5 lakh; a junior content writer costs ₹25,000–₹40,000 a month. Suzu Technology's site, with 44 interactive tools, a learning hub and a daily blog, went from an empty repository to a live site in its first 48 hours (the Git history is public), built in the evenings alongside a full-time schedule. The saving is not the money — it is that the work actually got done.
The honest caveat: this works because the person steering it can read a diff, a Lighthouse report and a DMARC record. AI compresses the time of someone who already knows the job. It does not replace knowing the job.
7. Where a founder or engineer should start
- Put one repeatable task in an agent's hands this week — the weekly report, the backlink check, the invoice reminder. Measure hours saved.
- Give it tools, not passwords. If the task needs a system, wrap the three commands it needs in a small MCP server with a scoped account and a log. Our AI & automation page describes the engagement if you want it built.
- Keep everything in Git. Prompts, generators, content, release notes. An agent that can read the repo's rules is an agent you can correct once.
- Schedule, then supervise. Cron the task, read the report every morning for a month, tighten the prompt each time it does something you would not. After a month, you have a colleague.
- Write down what it must never do. Ours is one paragraph long and has never been broken.
If you are a student or early engineer: the roadmap in how to become a cloud / infrastructure engineer in India is what I would do today, and the free tools on this site are the kind of small, finished things that get you hired.
Suzu Technology designs infrastructure, builds websites and AI agents, and runs the marketing — one accountable team under one phone number. Meet the founder on the About page or start a project.