Planned stories
Every piece of work starts with scope, acceptance criteria and dependencies settled up front. Scope questions get answered in planning, not halfway through implementation.
South Africa (SAST) · Remote or hybrid
Justin Heyns
Senior Engineer: Full-Stack & Smart Contract Delivery with AI Agents
I am the sole engineer on a live financial protocol on Ethereum mainnet. I built it with a delivery system where agents do the work and something other than the agent checks every result before it ships.
- 800+
- reviewed work items since July 2025
- 30+
- repositories under one pipeline
- 40+
- deployed contracts, none exploited
- 10+
- years of production software
Summary
I designed and built more than 40 deployed contracts, the web app, the deployment tooling and the operations for a live stablecoin protocol. That is work that used to need a team. Every contract since 2025 has gone through my own audit pipeline.
AI agents handle whole stages of my delivery cycle: planning, implementation, testing and first-pass review. My job is designing the system they work in. That means the guardrails, the independent checks and the standard every change must meet before it ships.
I can bring this to a team however it uses AI today. I can ship features quickly myself. I can also set up the ticket structure, review gates and CI checks that let your engineers delegate more to AI without AI-quality bugs reaching production.
Before crypto I spent six years writing C# / .NET and React at Digiata, a financial-services software firm. I was a Code4rena audit-contest judge, and I hold a master's in economics, which I use in mechanism and incentive design.
How I work with AI
Everything an agent produces is untrusted until something other than that agent has verified it. This is what "Claude Code" on my CV means in practice.
- PlanScoped story with acceptance criteria
- IsolateOwn worktree and branch per sprint
- ExecuteAgent implements and tests
- ReviewFresh agent judges the result
- ValidateCheckboxes and file changes verified
- ApproveMachine and human sign-off kept apart
Isolated execution
Each sprint gets its own git worktree and branch. Every commit is tagged with its story number, so any line of code traces back to the decision behind it.
Independent review
A separate review agent that has never seen the implementer's reasoning judges the result. Validator agents confirm every ticked checkbox was really done and flag suspicious file changes.
Reward-hacking detection
Agents sometimes weaken or delete tests instead of fixing the code. Dedicated checks look for exactly that, so a green test suite means what it says.
Rules in the system, not the prompt
Critical invariants are enforced by CI gates, dedicated tests and tool-call hooks. An agent cannot argue its way past them.
Unattended batch runs
Runs chain stories and projects together and merge finished work. They stop at the first result a human should see: a failed review, a refused approval, or a story out of attempts.
What I can do for you
Ship features end to end
Design, implementation, tests, frontend, deployment and production operations.
Raise your team's AI maturity
Move engineers from autocomplete to delegated tickets to autonomous pipelines, with review gates that keep quality where it is now.
Find bugs before attackers do
Static analysis, symbolic execution, invariant fuzzing, and a runnable proof of concept for every finding.
Design incentives that hold up
Pricing, yield and reward mechanics judged by an economist as well as an engineer.
Experience
Stablecoin protocol on Ethereum
A yield-bearing stablecoin protocol, live on Ethereum mainnet and listed on DefiLlama. Protocol name and links are available on request.
- Designed minting at a fixed 1:1 rate against dollar stablecoins, which removes any dependence on a price oracle.
- Built a pluggable yield-strategy layer that wraps external ERC4626 vaults behind one interface, so funds move between yield sources without user action.
- Wrote a staking system with a fixed migration interface, enforced by a CI gate and dedicated tests. Deployed code can't be patched, so the guarantee that users can always be moved has to be built into the structure.
- Built NFT-based yield claims, NFT staking, an emergency pause system and Uniswap / Balancer liquidity integrations.
- Run deployments from a staging repository that pins every contract repository as a submodule, so the exact source behind each live contract can be traced.
- Built the React / wagmi web app and the Next.js marketing site, and wrote the DefiLlama TVL adapters.
DeFi liquidity protocol
A token-launch and liquidity protocol, launched on Ethereum mainnet in February 2021 after a year of development.
- Built a multi-token automated market maker and its React frontend.
- Built a liquidity-seeding system for new tokens, which went through a public competitive audit contest in January 2022.
- Built yield-backed token wrappers and a yield-aggregator line that later became the stablecoin protocol above.
Code4rena
- Judged four public smart contract audit contests (PoolTogether, FactoryDAO, FIAT DAO and Fraxlend), ruling on the validity, duplication and severity of 436 submitted findings.
Digiata Technologies
- Full-stack development of business applications for financial services, insurance and banking clients, on C# / .NET, ASP.NET MVC and SQL Server, with React front ends.
Earlier
Founder and developer of a gift card auction platform (2012 – 2013). Economics lecturer at the University of Cape Town (2010) and TSiBA Education (2008).
AI engineering systems
Private · live demo on request
Multi-agent delivery workflow with review gates
The system described above: planned stories, isolated worktrees, independent review, validator agents and separate machine and human approval.
800+ completed stories across 30+ repositories since July 2025.
Claude Code plugin · repository on request
AI-assisted smart contract audit pipeline
22 specialised agents and 16 commands covering scope discovery, static analysis (Slither, Aderyn, Semgrep), symbolic execution (Halmos), invariant generation, proof-of-concept exploits, severity review and report writing.
No finding is reported until it has a runnable Foundry proof of concept and a second-opinion severity check, which filters out plausible-sounding false positives. It also reviews deployment scripts by running them against a mainnet fork before they go live.
Skills
AI-assisted engineering
- Multi-agent pipelines
- Claude Code agents, skills, hooks, workflows
- Agent-executable tickets
- Independent review & validation
- Reward-hacking detection
Frontend
- TypeScript
- React
- Next.js
- wagmi / viem
- Tailwind
Backend
- C# / .NET
- ASP.NET MVC
- SQL Server
Smart contracts
- Solidity
- Foundry (unit, fuzz, invariant)
- ERC20 / 4626 / 721 / 1155
- Upgrade & migration design
Security
- Slither
- Aderyn
- Semgrep
- Halmos
- Medusa
- PoC exploit tests
Infrastructure
- AWS (CloudFront, Lambda)
- Git worktree workflows
- CI gates
- Production deployment
Education
University of Cape Town, 2005 – 2011
- Master of Commerce in Applied Economics
- Bachelor of Commerce (Honours) in Economics
- Bachelor of Science in Computer Science and Economics, with a Science Faculty entry scholarship
Let's talk
Public code, published writing on token economics and a live demo of the delivery pipeline are available on request.