Side project · AI diagramming

From messy AI outputs to clean, editable diagrams.

Bhava is my nights-and-weekends AI tool built with one part-time engineer. I lead everything from product discovery to shipped experiments—driving 60% activation, 5% landing conversion, $150–200 MRR, and −40% cost per diagram in four weeks.

Prompt strategy Trust signals Pricing & retention AI evals

Product discovery

Mapped workflows, ran 8 interviews, and sized activation gaps before touching UI.

Trust-led UX

Turned the prompt into a guided demo, added progress states, and rebuilt onboarding.

Growth experiments

Removed free mode, shipped usage-based pricing, and tracked retention + MRR weekly.

AI evaluation

Logged 100+ failed diagrams, clustered errors, and guided sub-agent strategy to lift accuracy.

60%

Activation lift

1,600+

Early users

25

Paying teams

4

Renewals sealed

Context

Why we built this

I'm a product designer at an ad-tech startup. By day, I'm deep in B2B dashboards. By night, I watch my team waste hours redrawing the same system diagram in Figma, Draw.io, Excalidraw, and Miro.

Same workflow. Four different tools. Different versions. Complete chaos.

So I started building Bhava—an AI tool that generates diagrams instantly. But more importantly, one that doesn't feel like a black box.

This is early stage. We're 4 weeks post-launch with ~1,500 users and $150–200 MRR. I work on this part-time alongside my full-time job. One engineer friend helps part-time. Between us, I handle design, product, evals, UI fixes, pricing experiments, and customer interviews. He handles optimization and infrastructure.

This is the story of how we went from a fuzzy idea to 60% activation—and what I learned about building AI products people actually trust.

The problem

Broken workflows, broken trust

🔀

Diagramming was fragmented

Engineers used Draw.io
Designers used Figma
PMs used Miro
Same diagram, 4 places, 30+ minutes each time

🤔

AI tools fell short

Vague prompts → broken diagrams
No feedback while AI "thought" → mistrust
Failures with zero explanation

The activation problem: In early tests, only 38% of users created their first diagram. The other 62% bounced without trying.

Our bet: Build on top of Draw.io (largest user base) and make AI feel reliable, not random.

Approach

Designing for trust

Every design decision mapped back to a trust framework for AI research

Ability

Can the AI actually do the task?

Benevolence

Does it feel like it's helping me?

Integrity

Is it honest about what it can and can't do?

Reliability

Does it work consistently?

Discovery

Understanding the drop-off

Before redesigning anything, I spent 2 weeks analyzing user behavior—watching session recordings, tracking prompts, and interviewing people who churned.

Blank canvas paralysis

Users landed on an empty editor with no guidance, no examples. They froze.

Mode confusion

"Intelligent" vs "Basic" results varied wildly. Trust eroded fast.

No progress feedback

3–8 seconds of spinner. No updates. Pure anxiety.

Hidden export

Only 15% exported their first diagram. The happy path was invisible.

"I don't know what to type, so I just close the tab."

— Product manager, B2B SaaS

What I shipped

Six experiments that moved metrics

Each redesign tackled a specific trust or activation gap. Here's what worked.

Experiment 01

Homepage prompt became the product demo

Problem: Vague CTAs meant visitors signed up without understanding what to type.

Solution: Elevated a giant prompt box with example chips and a mini walkthrough so users preview the experience before creating an account.

1% → 5% Landing conversion (+4pp)

45s → 12s Time-to-first-prompt

Experiment 02

Guided prompt experience replaced the blank chat

Problem: New users froze on an empty chat and churned without generating anything.

Solution: Added diagram-type cards, contextual hints, and a three-step progress indicator that nudges people into action.

38% → 60% Activation (+22pp)

+18% Prompt quality scores

Experiment 03

Removing free mode protected trust

Problem: The legacy "Basic" mode produced low-quality diagrams that tanked perceived reliability.

Solution: Sunset the free mode, offered one premium try, and introduced usage-gated access to keep output quality consistent.

30% Day-7 retention (stabilized)

−55% "Bad diagrams" support tickets

Experiment 04

Usage-based pricing matched value to spend

Problem: Unlimited $10/month plans were unprofitable and encouraged abuse.

Solution: Swapped to a $10 base plan with transparent credit packs and real-time usage tracking.

$30 First enterprise add-on purchase

−22% → +14% Contribution margin flip

Experiment 05

Usage clarity reduced support debt

Problem: Pricing changes created confusion—users couldn't tell where credits went.

Solution: Built an always-available tutorial and a usage dashboard detailing credits, modes, and expiry.

−60% Billing questions

+1.2 pts NPS on transparency

Experiment 06

Manual evals powered sub-agent quality

Problem: Diagram quality varied by type and we lacked clarity on failure patterns.

Solution: Logged ~100 failed diagrams, clustered errors, and routed high-volume types through specialized sub-agents.

+70% Flowchart success rate

Flat Costs (thanks to caching)

4 weeks post-launch

Current metrics

A snapshot of where things stand after the first month of shipping.

Activation

60%

+22pp from 38%

Landing conversion

Visitors → Sign ups

+4pp from 1%

Cost per diagram

$0.048

After prompt caching

−40% reduction

MRR

$150–200

~30 paying customers

First month baseline

Day-7 retention

30%

Active after one week

Stabilized post-pricing shift

Generation speed

3.2s

p50 latency

7.8s at p95

Reflection

What broke, what worked

Failures

Template gallery: zero usage, removed quickly
Three free tries: people churned emails, costs spiked
Basic mode: poor quality eroded trust
Five-step onboarding: too long, trimmed to three

Wins

Homepage prompt lifted conversion 4pp
Progress indicators stopped mid-gen refreshes
Caching reduced costs by 40%
Quality-first approach kept users returning

What I'd do differently

Set up analytics on day one (not week two)
Start manual evals earlier to catch quality issues
Test pricing experiments faster—waiting cost us margin

What's next

Future bets

Shipping soon (2–4 weeks)

"Explain this diagram" overlays
One-click export to PNG, SVG, PDF
Diagram versioning

Exploring

Team collaboration (comments, shared workspaces)
API access (10 inbound requests)
Template starter packs

Said no to

Custom branding (low demand)
Slack/Notion integrations (unclear ROI)
White-label offering (not before $2K MRR)

12-month thesis: AI will replace 30–40% of manual diagramming. Winners will prioritize speed, transparency, and trust. Draw.io integration gives distribution. Usage-based pricing aligns incentives.

Honesty

Risks I'm owning

Four weeks is too early to claim product-market fit
Need to track day-30 and day-60 retention before declaring success
Export rate stuck at 15%—my next focus area
Eval rubric is v1; still manually labeled and somewhat subjective
Some metrics estimated; refining instrumentation
Pricing tests ongoing; willingness-to-pay still forming

Explore more work

Product-Led Growth

PushOwl PLG

Start fast. See value soon. Decide later.

→

AI Product Design

PushOwl AI

AI features that reduce time‑to‑first‑win

→

Want to dive deeper?

Let's chat about designing trustworthy AI, running growth experiments, or how I can bring this playbook to your team.

Get in touch View more work