Skip to content
Naresh Muppalla

BusinessOnBot · Dec 2025 – Present · Bengaluru

Customer Success Manager

I own 42 D2C brands, from onboarding and setup through go-live and daily success.

  • Configure WhatsApp AI agents in Sim Studio and Langflow.
  • Turn brand issues that repeat into one root-cause fix.
  • Write up every fix, so the team and the AI assistants answer it the same way.
Journey →Playbook →About →Contact →Résumé ↗

(L)The lab

AI that does the busywork.

The tools and workflows I built so I can spend my time on customers, not on reports. Most of them run every day. Brand names and customer data are left out.

  1. 01AI Ops Workspace→
  2. 02Codebase + logs investigator→
  3. 03WhatsApp AI agents→
  4. 04Guardrails library→
  5. 05Conversation review→
  6. 06Help centre + SOP knowledge base→
  7. 07Spec-to-site workflow→

01 / 07

AI Ops Workspace

Claude Code · runs every day

My daily support desk, run with an AI agent. Every ticket I handle is logged once and turned into the right output.

  1. 1Log the day's tickets in one file
  2. 2Claude drafts Jira-ready bug (CRB), change (CFR) and product-feedback (PFR) reports
  3. 3Repeated issues roll into a pattern tracker, and ready replies cite their KB source

Result

Less time writing reports, more time with brands. 7 bug reports, 5 feature/product reports and 9 recurring patterns tracked so far.

Claude Code · Markdown · Jira format · Postman

02 / 07

Codebase + logs investigator

Claude Code · New Relic · read-only

For tickets nobody can explain, I trace the answer in the company's own code and logs before replying.

  1. 1Read 28 company repos read-only to follow the exact code path
  2. 2Query New Relic (NRQL) for proof: error codes, counts, timelines
  3. 3Write the root cause, a plain customer reply and an engineering escalation

Result

Root causes confirmed with evidence: a routing cap, a stale template, a platform-wide AI outage. Five are written up as case files.

Claude Code · New Relic NRQL · GitHub · Meta error codes

Read the case files →

03 / 07

WhatsApp AI agents

Sim Studio · Langflow

AI agents that sell and support for D2C brands on WhatsApp. I build and launch them as part of onboarding.

  1. 1Design the prompt, the conversation flow and the tool calls
  2. 2Test against scripted conversations before go-live
  3. 3Example: a clinic appointment agent that collects details and tags tickets

Result

Brands go live with an agent that already handles the real questions.

Sim Studio · Langflow · OpenAI · Tool calls

04 / 07

Guardrails library

9 brand bots

The rules that keep brand agents safe: what they must never say, when to hand off to a human, and how to stay on-brand.

  1. 1No medical, legal or refund promises
  2. 2Clear human hand-off triggers
  3. 3Brand tone and scope written per bot

Result

Guardrails written for 9 brand bots. Safe agents are the ones brands keep switched on.

Prompt design · Policy · Hand-off rules

05 / 07

Conversation review

Python + LLM

A way to find where an agent fails, without reading thousands of chats by hand.

  1. 1A Python script cleans exported chat logs and removes test traffic
  2. 2Messages are grouped into whole conversations
  3. 3An LLM review prompt lists where the agent went wrong

Result

A short list of prompt fixes, based on real conversations, not hunches.

Python · LLM review prompt

06 / 07

Help centre + SOP knowledge base

Docusaurus · llms.txt

A public help centre and the internal operations manual I own, both readable by AI assistants.

  1. 1Troubleshooting guides: WhatsApp deliverability, New Relic tracing, templates
  2. 2llms.txt so AI assistants answer from the same source
  3. 3“Was this helpful?” and an issue reporter on every page

Result

Customers, the support team and AI assistants all answer from one source.

Docusaurus · Astro · llms.txt

See the knowledge base →

07 / 07

Spec-to-site workflow

Claude Code · Astro

How I ship websites fast without losing the design: a written spec goes in, and I review every build in the browser.

  1. 1Write the spec: audience, sections, copy, design tokens
  2. 2Claude Code builds against it
  3. 3Check it in a real browser at real breakpoints before it ships

Result

Nine live sites, including this portfolio, built this way.

Claude Code · Astro · Vercel

See the sites →

(✓)Responsible by default

How I use AI safely.

  • Read-only access

    The codebase and logs are read, never changed. Fixes go through engineering.

  • Human review before go-live

    Every agent is tested against scripted conversations first.

  • No customer data in prompts

    Logs are cleaned and test traffic removed before any review.

  • Evidence, then claims

    A root cause is only stated once the logs confirm it.