AI Automation  ×  On-Chain Engineering

Where intelligence meets value.

Software that acts on its own, and money that runs on code. We're a senior, forward-deployed engineering team that builds both — and the secure seam where they connect.

Book a discovery call →See it work ↓
◣ INTELLIGENCEthe convergenceVALUE ◢
REASON & ACTMOVE & SETTLEAI AGENTSAUTOMATIONON-CHAINSMART CONTRACTSintelligence × value
Built with LangGraph · Temporal · MCP · Solidity · Rust · Move — across EVM, Solana, Stellar & Sui
see it work

An agent decides. The chain enforces. Value moves.

The hard part isn't getting a model to suggest an action — it's letting it move real money without giving it the keys. Here's a live run: the agent reasons off-chain, the smart contract enforces the rules on-chain, and nothing settles outside the bounds we wrote in code.

agent-run · treasury-rebalance.flowrunning…
◣ OFF-CHAIN · AGENT REASONS
the seam
ON-CHAIN · CONTRACT ENFORCES ◢
require(...) — or revert
healthFactor ≥ 1.0
slippage ≤ 0.5%
perTxCap ≤ 25,000 USDC
tx 0x9f3c…a17
rebalance 12,400 USDC → vault
block 21,403,118 · 1.2s · $0.004
✓ value moved — strictly within on-chain bounds
agent never holds custody rules live in the contract, not the prompt every step provable on-chain
Reliability compounding

Why agent reliability collapses on long tasks.

A long task is a chain of steps. If each step is independent, end-to-end success is pN — it decays exponentially. You don't fix that by waiting for a smarter model; you fix it with architecture: a verifier that catches a bad step and a bounded retry. On-chain, the contract is that verifier.

100%75%50%25%0%coin flip01020304050STEPS IN THE TASK (N)naive agent · pⁿbounded · verify + retry
per-step reliability p0.95
verify + retries k3
AT 20 STEPS
35.8%
naive agent succeeds
99.8%
bounded agent succeeds
coin-flip at N = 14
90%-horizon: 2 800+ steps
naive = pᴺ · bounded = (1 − (1−p)ᵏ)ᴺ · the gap is the value of verification
Two disciplines · one team

Most teams pick a side. We engineer both.

01 · Intelligence

AI & Agentic Automation

From a single automated workflow to large-scale, multi-agent systems that plan, call tools, and act under supervision. We ship for enterprises and prototype fast for startup teams.

Enterprise agentic systems — orchestration, tools, memory, evals & observability.
Domain automation — for manufacturing, accounting, and legal operations.
Product prototypes — zero-to-one agent products for founding teams.
Guardrails & safety — bounded autonomy, human-in-the-loop, audit trails.
LangGraphTemporalMCPClaudeopen models
02 · Value

On-Chain & DeFi Engineering

Core protocol work — not generic dApp glue. We design and ship value-bearing systems where the economics and the security have to be right the first time.

Lending & money markets — collateral, liquidations, rate models.
DEXs & derivatives — AMMs, order books, and perps/futures venues.
Yield & RWA protocols — vaults, strategies, tokenized real assets.
Economic risk dashboards — stress testing, monitoring, assessment.
SolidityRustSorobanMoveEVM · Solana · Stellar
Embeddings & semantic search

How AI turns text into searchable meaning.

An embedding model turns text into a long vector of numbers, so that meaning becomes distance: related ideas land near each other. Search stops being keyword-matching and becomes geometry — find the nearest neighbors. It's the retrieval step behind RAG (retrieval-augmented generation) — how an agent pulls the right context before it answers, matched by cosine similarity.

1Encode
input
"chargeback dispute"
sub-tokens
embedding vector · ℝ¹⁵³⁶
‖v‖ = 1.00
2Place
1,536-d projected to 2-D · near = similar meaning
3Retrieve
query —
cosine(a, b) = (a · b) / (‖a‖ ‖b‖) · retrieve the k nearest neighbors
Automation · by industry

The back office, run by agents.

From a single workflow to a supervised fleet — the same engineering rigor we bring on-chain, applied to the work that actually runs a business. A few of the places it fits:

Manufacturing · procurement & QA

Keep the plant moving

Agents read ERP and sensor data, flag quality anomalies, draft purchase orders, and chase approvals — the loop that keeps procurement and the line in sync.

Explore use case →
Accounting · close & reconciliation

Surface only the exceptions

Month-end close, ledger reconciliation, and audit-prep agents pull from the books and hand a human only the line items that genuinely need judgment.

Explore use case →
Legal · contracts & intake

Redline, summarize, route

Contract review, client intake, and discovery agents that draft and triage at volume — with a lawyer on every final call and a full audit trail.

Explore use case →
Marketing · content & SEO

Hand the busywork to agents

Research, briefs, content production, and performance reporting run by agents — so the team spends its hours on strategy, not the grind.

Explore use case →
Customer ops · support & triage

Resolve routine, escalate the rest

Ticket triage, response drafting, and knowledge agents that close the routine cases around the clock and route the hard ones to a person.

Explore use case →
Back office · documents & data

Fewer handoffs, fewer errors

Invoice processing, data entry, and document workflows automated end to end — extraction, validation, and routing without the copy-paste.

Explore use case →
…and dozens more. If it's repetitive, rule-bound, and high-volume, an agent can probably run it.
Unit economics

Human ops scale with volume. Agent ops scale toward the gas fee.

Humans cost the same per task forever — a flat line. An agent is the opposite shape: expensive to build, then almost free to run, its cost-per-task collapsing toward the settlement fee as volume grows. Drag your own numbers and read what it's worth.

$0.04 floor — the settlement feebreakeven101001k10k100k1MMONTHLY TASK VOLUME (log)COST / TASKhumanagent
monthly volume20,000
human $/task$4.20
build cost$60k
agent $/task
breakeven
monthly saving
payback
agent(V) = F / V + m · floor m = $0.04 · breakeven V* = F / (c_h − m)
The convergence principle

Bounded by code, not trust.

Autonomy is only safe when the limits are enforceable. Every system we build at the seam follows the same four moves.

01
Observe
The agent reads on-chain state — prices, positions, health factors — as ground truth.
02
Sign
It reasons, then submits a typed, signed action. It proposes — it never takes custody.
03
Enforce
The contract asserts the rules. Outside the bounds, the transaction simply reverts.
04
Settle
Value moves and finalizes in seconds — every step transparent and provable.
Selective autonomy

The agent knows when it doesn't know.

Every decision carries a confidence. Set a gate τ: above it, the agent acts; below it, a human does. Drag the gate and watch the trade — push it right and almost no error ships, but humans review more; push it left for full autonomy and errors slip through. There's a cost-optimal place to stand.

τ◂ ESCALATE TO HUMANACT AUTONOMOUSLY ▸0.25.50.751.0AGENT CONFIDENCEwould be rightwould be wrong
base accuracy85%
gate τ0.70
automation rate
errors shipped
human reviews
cost-optimal gate τ* =
act if confidence ≥ τ, else escalate · cost(τ) = c_err·errors + c_human·reviews
The stack we go deep on

We speak the whole stack — fluently.

Agentic AI
LangGraph orchestrationTemporal durable workflowsModel Context ProtocolClaude & open modelsMulti-agent planningRAG & tool useEvals & observability
Chains
EthereumBase · ArbitrumBNB ChainSolanaStellarSuiAny EVM L2
Languages
SolidityRustSoroban (Rust)MoveVyperTypeScript · viemPython
DeFi & Risk
AMM & order-book DEXsPerps & futuresLending marketsYield vaultsTokenized RWAsRisk dashboardsAudits & formal checks
Self-attention

Not every word matters equally.

The mechanism behind every transformer: for each word, the model scores how much every other word matters, runs those scores through a softmax so they sum to one, and blends accordingly. Watch the focus move along the sentence — thicker links and hotter cells mean more attention.

FOCUS TOKEN ATTENDS TO THE SENTENCE — thicker link = more weightTheagentsignedthetxon-chainledgerattention weights · softmax → Σ = 1.00
focus: weights always sum to 1the operation inside every transformer
Selected work

Systems that move real value.

In-house protocols and confidential client work. Some names are under wraps.
Vanna
In-house protocolvanna.finance ↗

Vanna Protocol

A capital-efficient DeFi protocol with on-chain risk controls and agent-ready execution — our reference implementation of the convergence thesis.

LendingRisk engineMulti-chain
Building · name tentative

Auri

A tokenized-gold protocol bringing a real-world asset on-chain with verifiable reserves and transparent settlement.

RWAProof of reserves
Confidential · enterprise

Agentic ops for a leading manufacturer

Multi-agent automation across procurement and quality operations for a top Indian manufacturing group.

Enterprise agentsOps automation
Global by default

We work where our clients are.

Remote-first, senior and forward-deployed, we partner with founders and enterprises across the Americas, Europe, the Middle East and Asia. Engineering follows the sun — you get one team, aligned to your time zone, wherever you are.

AmericasEuropeMiddle EastIndiaAPACRemote-firstTime-zone aligned
USACanadaUKUAEIndiaSingaporeclients on four continents · one team
How we work

Three ways to start.

01 · 2 weeks

Discovery sprint

We map the seam: architecture, threat model, and a costed plan you own — whether or not we build it.

02 · project

Protocol & agent build

A senior, forward-deployed pod ships your protocol or agentic system to production — audited, monitored, and handed over clean.

03 · ongoing

Forward-deployed engineers

Senior, forward-deployed engineers working inside your team — to raise the bar on the systems that move value.

Writing

Notes from the seam.

All writing →
cover · AI safety
AI · 6 min

Why agents need on-chain guardrails, not bigger context windows

Bounded autonomy beats blind trust — a case for putting limits in the contract.

cover · DEX design
DeFi · 9 min

Designing a perps DEX an AI can trade without rugging itself

Risk limits, oracle hygiene, and the invariants that keep automation honest.

cover · stablecoins
Stablecoins · 7 min

Stablecoins in 2026: the settlement layer for agentic commerce

When software pays software, the rails have to be instant, cheap, and final.

Have something at the seam of AI and value?

Tell us what you're building. We'll tell you, honestly, whether we're the right team — and how we'd start.

Book a discovery call →hello@zonymous.com