BITVAVO_ENG // DEFI.EXE — Q2 2026
▸ INSERT AI TO CONTINUE
Shipping DeFi
at AI Speed
A brand-new team, mixed skills, a hard deadline — and a system that made it repeatable.
Shipping DeFi
LEVEL 1 — PLAYER SELECT
A team that didn't exist
until end of February.
Crypto ✗ FE
Deep crypto knowledge — new to mobile & frontend
FE ✗ Crypto
Strong frontend devs — never touched crypto
+ Non-eng
Engineering managers & a designer in the mix
Different backgrounds, big skill gaps — nobody held the whole map.
THE MAIN QUEST
DeFi: a crypto wallet,
built on the frontend.
// WHERE IT RUNS
Crypto logic in-app
Cryptographic operations handled in-house, with select third-party services.
// THE DOMAIN
Deep & novel
Blockchain, wallet mechanics, transactions, crypto-specific UX. Not an integration.
// THE DEADLINE
Ship in Q2
Steep learning curve, no slack in the schedule. On time was the only option.
★ QUEST COMPLETE
DeFi.
Delivered.
On time.
Shipped by the end of Q2 — final release the week of this talk. Now the interesting part: how.
NOT JUST DEFI
Three big wins.
Same team, same quarter.
DeFi
Complex crypto wallet — delivered on time
Design System
Helped lay its foundations & AI setup — still in progress
RN → Expo
Full Bare → Expo migration, native code & all
Two of these are migrations teams defer for quarters. We did all three together.
THE HUMAN STORY
The people who
"couldn't" — did.
- Three of our developers rank among the repo's top contributors by commits & PRs.
- A crypto expert — new to the codebase, new to mobile — became one of the highest shippers.
- A designer and an engineering manager wrote code that reached production.
Not about replacing engineers — about removing the friction between what you know and what you can build.
🏆 RECEIPTS — LAST 30 DAYS
Three of the repo's top four are DeFi engineers.
#2 · #3 · #4 last month — all DeFi engineers. 100+ commits each, in their first full quarter.
BY THE NUMBERS
Shipping velocity.
0
DeFi pull requests merged
0
merged every day — since February
PRs merged / month · 2026
1
32
58
139
155
111*
JANFEBMARAPRMAYJUN
* June partial — through the 25th
NON-ENGINEERS SHIPPING
A designer,
shipping real UI.
- Daniel — a designer, not an engineer.
- Working in the code taught him the whole flow — what's possible, and what isn't.
- AI closed the gap between the Figma and the codebase.
Real pull requests. Real reviews. In production.
NON-ENGINEERS SHIPPING
An EM, fixing a
gnarly crypto bug.
- Bozho — an engineering manager.
- Took on a genuinely hard, chain-specific crypto bug.
- Root-caused, fixed, tested, merged to main.
The kind of bug you'd expect only a senior crypto dev to touch.
HOW WE DID IT
A system, not luck.
Six phases the AI could follow.
01Knowledge foundation
02Production architecture
03Live business-logic docs ★
04AI governance — rules & skills
05UI & design-system
06The defi-developer agent ★
01
PHASE 01 / 06
Knowledge foundation
- Fed Claude every doc: legal, PRD, specs, third-party & design.
- Two engineers ran a rapid POC: swap, transfer, balances, history.
- Docs + POC + third-party references → a first-draft architecture doc.
Output → deep domain context, a throwaway POC, and v1 of the architecture.
02
PHASE 02 / 06
Production architecture
- Took the Phase 1 draft + mixed in our production codebase → an extended dialogue with Claude.
- Layering, module boundaries, isolated crypto logic, state, integrations.
- Mapped onto our stack: React · Sagas · Redux · secure storage.
Output → project-specific architecture blueprints. Still just docs — no production code yet.
03
PHASE 03 / 06 — THE ONE PEOPLE SKIP ★
Live business-logic docs
- Code says what. It doesn't say why.
- Rules & skills auto-update a doc every time a feature changes — never stale.
- LLMs read intent first → output is more correct and aligned.
- Published to the wiki → PMs, design & new joiners all read it.
04
PHASE 04 / 06
AI governance
- DeFi Architecture Skill — enforces the layers & patterns.
- Boundary Rules — hard module lines; no leaking into other products.
- Project rules — DeFi conventions that accumulate over time.
Result → almost anyone generates code that looks & behaves like ours.
05
PHASE 05 / 06
UI & design system
- Rule: always use design-system components when they exist.
- Missing one? Build it — and contribute it back.
- We helped lay the foundations & enable the design system for production.
- Designer → Figma (design tokens) → code: one pipeline.
Result → fast, consistent UI + a still-in-progress design system, pushed forward.
06
PHASE 06 / 06 — THE PAYOFF ★
One agent to rule them all
- Wrapped it all into one subagent: defi-developer — a senior DeFi-engineer persona.
- It indexes everything: the architecture, every skill & rule, the live docs, the design system.
- Picks the right layer, respects module boundaries, reads a canonical example before scaffolding.
The capstone → the whole process, in one agent that writes DeFi code the way we would.
INSERT COIN — WHAT WENT WRONG
Not every quarter
is a high score.
- PR quality was rough in the early weeks.
- Then PRs piled up into a review bottleneck.
- And AI is no substitute for understanding crypto.
Here's what tripped us up — and what we did about it.
WHAT WENT WRONG · 01
Early friction, then improvement.
Week 2: messy PRs, many iterations. Week 6: the same developers shipping clean, production-ready code — because the rules caught up.
Reviewer comments per PR · relative
WEEK 1WEEK 2WEEK 3WEEK 4WEEK 5WEEK 6
The system — not just developer skill — was the multiplier.
WHAT WENT WRONG · 02
Then: too many PRs.
Volume exploded. Human review became the bottleneck.
- Turned on automatic Copilot reviews.
- Bigger win: gave Copilot instructions generated from our Cursor skills & rules — so it reviews in our conventions.
- Added more linting rules to catch the nits before a human looks.
Push routine review down to the machine; keep humans for judgment.
WHAT WENT WRONG · 03
You can't outsource
understanding.
We can't fully trust AI on crypto — and we can't add value we don't understand ourselves.
- Team sessions walking through our own features.
- Real documentation — including the self-generated docs from Phase 4.
- AI as co-pilot, not autopilot for anything cryptographic.
Domain understanding is still the job. AI accelerates it — it doesn't replace it.
TAKE THIS HOME · 01
Treat PR comments
as training data.
// THE LOOP
Mine the reviews
Every 1–2 weeks a script pulls recent PRs and their review comments; Claude finds the repeated mistakes and turns them into new or updated rules.
// BEST SIGNAL
Non-engineer code
Where EMs and designers get corrected is the sharpest signal for where the AI needs better guidance. The single most impactful practice we adopted.
Not theoretical: messy week-2 PRs became clean by week 6 as the rules evolved to match what reviewers kept flagging.
TAKE THIS HOME · 02
Scripts > prompts
for repeatable work.
- Prompts are powerful but non-deterministic.
- Write a deterministic script for boilerplate — then wrap it in a skill.
- Run script → validate → let AI adapt on top when needed.
- Codify the nits reviewers repeat as lint rules — enforced automatically, every time.
Faster · consistent · far fewer tokens · and you can actually test it.
BONUS ROUND
…and the math just changed.
TAKE THIS HOME · 03
Remove friction
for non-engineers.
// SETUP
Zero-config onboarding
Cloud environments, one-command setup. Nobody survives a two-day local config before their first line of code.
// TESTING
Easy to verify
One command, a visual preview, or a clear pass/fail. They shouldn't need to grok the whole test suite.
// SHIPPING
Short path to prod
Fewer gates means fewer places to get stuck — and fewer chances to lose confidence.
Great AI tooling + a painful process = they quit. Remove the friction and you unlock contributors with deep product, design & business context.
TAKE THIS HOME · 04 — THE BIG ONE ★
Never stop looping.
- Your AI setup is a living system — keep polishing it.
- Add, update, retire skills & rules as the project evolves.
- Learn from where the prompts went wrong.
- Constantly mine insights from your own sessions.
It's never "done." Looping on the system itself is the advantage that compounds.
THE TAKEAWAY
AI didn't replace the team.
It amplified it —
and it's repeatable.
Docs, rules, live feedback loops, deterministic scripts. A methodology any project can adopt — not a one-off.
END OF DEMO
CONTINUE?
Questions, war stories, and live demos welcome.
THANK YOU
#ask-defi
Q2 2026