Aampe
Madhuri Maram was Head of Design & Product at Aampe from January 2025 to June 2026. She designed Aampe's agentic content flywheel, NX, Relay (LLM copy generation with evals), Reward Functions and Label Buddy, and led a six-person design team. Aampe was acquired by MoEngage on 24 June 2026.
Aampe gives every user their own AI agent. I designed what those agents learn from, and the tools the team uses to run them.
A partial list of Aampe's customers
I joined as Head of Design & Product and stayed until MoEngage acquired Aampe. I led a team of six designers. I also owned product calls, and I shipped code when that was the fastest way to learn.
My contribution
Led design
Built and ran a team of six. SHARP, Aampe's design system. A design process rebuilt for AI.
Owned product
The content flywheel, NX, Relay, Reward Functions and content coverage, from pitch to launch.
Shipped code
apps.aampe.com, Label Buddy, Patches, and the harnesses and skills behind them.
- 3 months →1 month
- Time to value for a new customer, once the flywheel was live
- 70% →95%
- First-pass copy quality from Relay, scored by evals
- 40 hours →1 hour
- A month, per CS member, spent checking copy labels by hand
Aampe was acquired by MoEngage on 24 June 2026. Read the announcement
About Aampe
Agentic infrastructure for personalised experiences. A shipped product, documented publicly on kb.aampe.com. Acquired by MoEngage.
When a brand like Grab brings on a million users, Aampe gives each one an agent. It watches what that person taps, and when, across push, WhatsApp, email and in-app.
Every message is assembled from labeled parts: a greeting, a call to action, an offer, a value prop. Each agent learns its user's weights for those labels and writes copy to match.
Tag a part wrong and every agent learns the wrong thing, at 100k combinations a day. That is why so much of my work went into content quality.
Agents explore what to send, when and where, then exploit what works. Reward functions define a win. Label weights set priorities.
The agentic content flywheel
Most AI content tools generate once and stop. I designed a loop instead. Six stages, NX at the centre, and each pass gives the next one more to work with.
Brand voice, offers, legal constraints, events, topics and attributes, plus past messages from CDPs like Braze and MoEngage. Relay's input layer turns them into one prompt structure, so every agent has enough distinct content to choose from.
- Aampe engineering
- Customer data from CDPs
- My part
- Relay's input layer
Marketers create and label content in Composer. I moved it from one form per message to the NX canvas, where every variant sits in one view. Relay generates the copy from components, labels and context profiles.
- Aampe engineering
- Composer, Aampe's message builder
- My part
- NX canvas and Relay's creation flow
Agents make explore and exploit decisions and deploy content against business goals. That engine is Aampe's ML and engineering work, and it does a lot more than this. My contribution was the controls people use to steer it: reward functions and label weights.
- Aampe engineering
- The agent engine: Aampe ML and engineering
- My part
- Reward functions and label weights
This is an add-on on top of Aampe's engineering. Relay's debug log, programmed by Rajat, records every step the LLM took and why, so anyone can ask “why did that happen?” and read the trace. We paired it with agent fleet dashboards, built and tested with CS in a few hours.
- Aampe engineering
- Debug log, programmed by Rajat
- My part
- Readable traces and fleet dashboards, with CS
Content coverage shows how labels spread across message groups and where the gaps are. Every gap is an opportunity for agents to experiment. Audience Coverage does the same across segments.
- Aampe engineering
- Analytics platform
- My part
- Content coverage and Audience Coverage
Fix labels, fill gaps, add topics. Label Buddy runs semantic QA on new and existing copy, so the next pass of inputs starts cleaner and the loop turns again.
- My part
- Label Buddy, built with Mohana
Wanderly helps travellers discover, plan and book personalised trips in one place.
Warm, inspiring and adventurous. Confident, never pushy.
No guarantees on prices, visas or weather. No implied airline partnerships.
- Aampe engineering
- Customer data from CDPs
- My part
- Relay's input layer
NX canvas- Aampe engineering
- Composer, Aampe's message builder
- My part
- NX canvas and Relay's creation flow
Reward Functions- Aampe engineering
- The agent engine: Aampe ML and engineering
- My part
- Reward functions and label weights
- RequestGenerate alternates · Value proposition
- StateLLM connection established (claude)
Before
status: testingAfterstatus: success+ provider: claude - Response24 alternates returned
- ActionLabels checked · 2 sent to review
- Aampe engineering
- Debug log, programmed by Rajat
- My part
- Readable traces and fleet dashboards, with CS
Content coverage- Aampe engineering
- Analytics platform
- My part
- Content coverage and Audience Coverage
| Type | Review |
|---|---|
| Value proposition | 127 |
| Tone | 94 |
| Greeting | 93 |
- ReviewGreeting / Warm welcomeCould not be classified automatically✓ ApproveFlag for rerunExclude
- ReviewValue proposition / EfficiencyBorderline match, human look needed✓ ApproveFlag for rerunExclude
- ReviewCall to action / UrgencyBrand voice clash flagged✓ ApproveFlag for rerunExclude
- My part
- Label Buddy, built with Mohana
I frankly don't know how Madhuri does it. She has an incredible vision for what the product could be, and is both excellent at leading a team to execute that vision and getting in the trenches herself.
The product, and the brand around it
The MVP had one form per message, and patching it was not helping. I mapped the spectrum of concerns, then moved Composer to a canvas. Marketers write, label and preview every variant of a message in one place.
Relay is Aampe's LLM copy engine. I took it from MVP to product: brand inputs, components and labels, context profiles and modifiers, with Prompt Shaper upstream as the conversational builder.
Every line is scored by evals on label fit, distinctness, brand voice, functional correctness and fact safety. First-pass quality went from 70% to 95%.
Relay's engine came from Schaun Wheeler. Mohana, Rajat and I designed and built the platform around it.
Madhuri is one of those rare product designers who makes you feel like a genuine collaborator. I used Relay and Email Configurator in real customer work.
Agents need a clear definition of a win. I designed how teams set one: pick the target events, weight them, review. Label weights then let teams nudge what agents favour.
I joined as the sole designer and led Aampe's rebrand and the new aampe.com with huegrid: identity, brand guidelines, web, social and events. It launched in April 2025.
Lead brand and design engineer: Arjun (arjunphlox) at huegrid.
In seven months, Madhuri and the design group have deeply transformed Aampe's aesthetics and helped make Design part of Aampe's DNA.
Before · one form per message
Sign in · Aampe ComposerWanderly helps travellers discover, plan and book personalised trips in one place.
Warm, inspiring and adventurous. Confident, never pushy.
No guarantees on prices, visas or weather. No implied airline partnerships.
- Value proposition60 labels
- Call to action12 labels
- Offering46 labels
- Greeting9 labels
- Default▾
- Festive
- Win-back
Your next escape is waiting. Plan it in minutes ✈️
Start
Pick target eventsReview the weights
Label weights
Brand guidelines
Event boothProduct features and tools I shipped
Agents learn from message labels, so one wrong label teaches every agent the wrong thing. I built an 8-stage semantic QA pipeline where the LLM runs sixth, on the smallest input. The monthly audit went from 40 hours to 1.
Email templates lived in Braze or Salesforce, so CS set them up by hand. This tool turns a customer's HTML template into an agent-ready format. It went from concept to live customer use in one sprint.
Building tooling at the intersection of design and AI personalization is genuinely hard. Madhuri navigated all of that with a level of thoughtfulness that made the products better.
A Figma plugin that turns frames into Aampe in-app surfaces in a few clicks. Pick the elements to personalise, then preview variants in Aampe. Live on the Figma Community.
Describe an audience in plain language and get the rules. Audience Coverage shows where messages are missing across segments. I built Audience Coverage with Kalyan Kola Cahill, a forward-deployed engineer, and tested it with two live customers in under a week.
She brings together exceptional product instincts, strong design judgement and the ability to lead projects with clarity and confidence.
A Chrome extension and an AI coding agent. Designers file a ticket with a screenshot, URL and context; Patches reads it, writes the code, deploys to staging and iterates on QA. 99 issues closed in a month.
| Type | Review |
|---|---|
| Value proposition | 127 |
| Tone | 94 |
| Greeting | 93 |
- ReviewGreeting / Warm welcomeCould not be classified automatically✓ ApproveFlag for rerunExclude
- ReviewValue proposition / EfficiencyBorderline match, human look needed✓ ApproveFlag for rerunExclude
- ReviewCall to action / UrgencyBrand voice clash flagged✓ ApproveFlag for rerunExclude
Under NDA. Walkthrough on request.
Request a walkthrough
She pays close attention to detail, from layout and typography to the overall feel of a product, and her output consistently reflects a high level of craft.
Building the design practice
I led a team of six, including me: four product designers, one PMM designer and one brand and web designer. We moved from PRD, wireframes and handoff to PRD plus prototype, shipped to apps.aampe.com and tested with CS, GTM and customers inside a two-week sprint.
She speaks through artefacts. Rapidly prototyped, iterated work that cuts through ambiguity and gets everyone on the same page.
Aampe's design system, built from zero. It ships as a Figma library and as an LLM-readable file, so Claude Code and Figma Make output looks like Aampe, not generic. Built with Arjun (arjunphlox) at huegrid.
A product lab for MVPs, modelled on Google Labs. I built the harness that made it work: aampe-template, a starter with Aampe's APIs and the SHARP design shell, ready for humans and agents. Every app starts from the same base, so designers ship working tools with Claude Code, and 7+ are live with customers.
The Pain Depot was a FigJam board one person kept every day. Two hours a day, always two to three weeks behind. I replaced it with a Cowork pipeline: a Chrome Slack export, a shared notes repo and synthesis skills that cite every claim.
It runs every morning across six sources and feeds KBs, PRDs and CS guides.
I wrote or drove 15+ KB articles and CS resources. I also wrote the harnesses and skills behind them: a Slack extract, a pain-point analysis skill and Aampe's platform context, packaged for Claude. CS gets call-ready KBs. Engineering gets cited PRD evidence.
She sought out customer feedback, turned it into actionable insights, and transformed tribal knowledge into resources customers and teams could actually use.
Designers get pulled between the current track and the future one. Conceptual Hour protected the future track. Every week we found and framed problems at least a quarter ahead, and each one got a problem statement and a PRD before any design started.
She generates strong ideas quickly, explores multiple directions before settling on one, and challenges the status quo.
Me · Madhuri
Our process, rebuilt for AI
The lab
The harness · aampe-template
One path to live
Before · the Pain Depot
After · the Cowork pipeline
Systems, not discipline
Skills and plugins
FAQ
What did Madhuri Maram do at Aampe?
She was Head of Design & Product from January 2025 to June 2026. She led design, owned product for the content flywheel, NX, Relay, Reward Functions and content coverage, and shipped tools herself.
What is Aampe?
Aampe is agentic infrastructure for personalised messaging. It gives every user their own AI agent, which learns what that person responds to across push, email, WhatsApp and in-app. MoEngage acquired Aampe on 24 June 2026.
Which AI products did she design and ship?
Relay and Prompt Shaper (LLM copy generation with evals), NX, Reward Functions, Label Buddy, apps.aampe.com, Email Variable Configurator, the Surface Creator Figma plugin and Patches for Linear.
What results did the work produce?
First-pass copy quality went from 70% to 95%. The monthly label audit went from 40 hours to 1 hour per CS member. Time to value for a new customer went from 3 months to 1 month.
How big was the design team, and how did it work?
Six designers including Madhuri. The team moved from PRD, wireframes and handoff to PRD plus prototype, shipped to apps.aampe.com and tested with customers inside a two-week sprint.
Is Madhuri available for design leadership roles?
Yes. Book a 15-minute call.
In seven months, Madhuri and the design group have deeply transformed Aampe's aesthetics and helped make Design part of Aampe's DNA.
She generates strong ideas quickly, challenges the status quo, and her design work is beautiful, polished and consistently high craft.
I'd frequently come back from a weekend off to Madhuri presenting a new agentic workflow she created for the design team on Monday.
What sets Madhuri apart is her ability to bridge the gap between customers, product, and design. Any team would be fortunate to have her.
Want the full story?
Case study briefs are on the site. Full case studies on request.