I build the whole system. With a team of AI agents, engineers when the project needs them, and rules none of them can rewrite.
Fifteen years in distributed systems. I design the system and own every decision; agents and engineers build, test and ship. Three production platforms are running tonight.
- Running in production
VeoCabs
VivaTrades
Shubh Bhagya- mcp-multi-dbnpm
Pick the one that matches where you are
AI agents that run the product, inside rules they cannot rewrite
Shubh Bhagya is built and maintained by fifteen Claude Code agents on a schedule: they read the analytics, file the work, build it, put it through a separate QA agent and ship it, then score each change against the growth objective it claimed to serve. The merge is done by a workflow, and only on evidence bound to the exact commit.
How the agents stay reliable
Authority lives in a workflow, not an agent
A workflow does the merge, on evidence bound to the exact commit rather than on an agent's word that the work is done. Anything irreversible still comes to a person.
Who holds merge authority →The agent that writes the code never grades it
QA is a separate agent with its own context. It failed 44% of first submissions, on a site where a merge is a deploy and there is no staging.
What goes wrong with agent teams →Limits enforced below the prompt
Read-only database connections and denied tools, not instructions a model can be talked out of.
Read-only by default: mcp-multi-db →Monitoring built to notice silence
Did each loop run, did anything come out, and which agent stopped? A zero is treated as a question, not a clean result.
The growth loop and its ops dashboard →
Is the AI you already run actually under control?
Two questions, and most AI features only answer the first. Is it secure: where untrusted text reaches your tools and data, what leaves for a model provider, what would stop a runaway bill. And is it governed: who can approve or merge what agents do, whether every decision leaves a record, and whether the controls on the dashboard exist in code or only on paper.
- Series · 11 parts
OWASP LLM Top 10, explained for startup teams
Every risk with real incidents and the control that matters most.
- Audit
Governance theatre: auditing an AI support bot
An audit log, SSO and a confidence score that were mostly decoration.
- Audit
From read-only to governed
Five gaps found in an MCP server that gives a model a brokerage account.
- Open source
Governing an MCP database server
Plan-checked table policy, alias-proof redaction, and the two bypasses an attack pass found.
Three platforms, built end to end
VeoCabs
Build your product, end to endAn intercity cab lead-generation platform across 150+ Indian cities: instant fare estimates from Google Routes, enquiries pushed to Telegram, and a CRM the business runs on daily. Built and still maintained by me, and run day to day by the business itself.
AgentsBuilt from scratch with Claude Code. The architecture and every decision that matters are mine; much of the typing is not.
150+ cities served37 curated route pages₹0 monthly infrastructure costRead the case studyVivaTrades
Add AI to your productMakes a stock decision take two minutes instead of an evening. Plain-English analysis of 500 NSE stocks, refreshed on a schedule, plus free strategy backtesting. No paywall, no signup.
AgentsAgents run in production: each night coding-agent sessions write the reports, with a sub-agent per analysis dimension.
500 nse stocks covered100 analysed per night · 5-day rotation₹0 cost to use · no signupRead the case studyShubh Bhagya
Add AI to your productGives away what other astrology sites charge for: kundli, matching, dashas, numerology, muhurat and dosha calculators, and unlimited questions answered — computed properly, cited, and free with no signup.
AgentsMaintained day to day by a loop of 15 AI agents: their cycles produced 43 merged pull requests in 25 days.
₹0 for everything, no signup76 cited explainers13s→1s homepage, 32 storage ops → 0Read the case study
Tools that came out of the work
- mcp-multi-dbTypeScript
Read-only MCP server over PostgreSQL, MySQL and SQLite. Safe by default, because a query tool with write access is an incident waiting for a prompt.
- promptevalPython
LLM eval framework with first-class token-cost tracking. Answers the question everyone skips: did that prompt change save money without losing quality?
- ai-support-botTypeScript
Upload documents, get a RAG-backed support assistant you can embed on any site as a widget.
- multi-agent-blogTypeScript
Cooperating agents gather context, draft, critique, then revise against the critique. A human still approves before anything publishes.
- mcp-zerodhaJavaScript
Read-only MCP server putting broker data — prices, portfolio, history — in front of AI agents, with an enforced allowlist and an audit line per call.
Notes from building the whole thing
- 2 November 2026 · 17 min
Autonomous AI agents that grow an app: the loop behind shubhagya.com
Fifteen Claude Code agents, eighteen GitHub workflows and one growth objective. What the loop does, what it shipped in 25 days, how it is monitored, and what still breaks.
- 2 November 2026 · 8 min
Governing an MCP database server, and what an attack pass broke
Read-only stops an AI agent changing data, not reading the wrong table. How mcp-multi-db adds policy, redaction and an audit log, and the two bypasses an attack pass found.
- 2 November 2026 · 11 min
AI agent pull requests: who holds merge authority
Fifteen Claude Code agents, a merge workflow, one loop that still holds a merge grant, and the event triggers that kept stranding green pull requests.
What are you building?
A short description of the problem is enough to start. If it is not something I should take on, I will say so and point you somewhere better.