Price action is not the full picture
Market prices, liquidity, indicators, news and sentiment have to be combined into one consistent context that agents can reason over.
FIELD NOTES / DIGITAL ASSETS & WEB3
The AI forms the thesis.
Code decides if it trades.
An AI-powered crypto trading platform where autonomous agents research markets and form trading decisions, deterministic risk controls decide whether the trade is allowed, and a dedicated execution layer handles blockchain interaction.
WEB3 CASE STUDY
Quant
Multi-agent research.
Deterministic risk control.
Quant / Agent decision workspaceACTUAL PRODUCT INTERFACE
THE SEPARATIONRecommendation is not authorization.
An LLM is a useful reasoning system, but not an appropriate authority for enforcing financial limits.
Follow the decision ↗ResearchReasonProposeAuthorizeExecute
DOMAIN
Digital asset trading
WHAT WE BUILT
AI + Web3 trading platform
OUR ROLE
Architecture, AI systems & blockchain engineering
01 / THE REAL ENGINEERING PROBLEM
“BUY ETH — 82% confidence”
is not a decision you can audit.
Crypto traders have access to sophisticated market data, decentralized exchanges and AI models, but these normally exist as disconnected tools — centralized-exchange prices, DEX liquidity, technical indicators, news and sentiment all monitored separately.
A generic AI-generated signal is difficult to evaluate when real capital is involved. Users need to know what evidence was considered, which analytical perspectives agreed or disagreed, what the final thesis was, and why risk controls approved or rejected it.
So AI reasoning had to become part of the product experience rather than a black box that emits a verdict.
An LLM is the wrong authority for limits
Prompting a model with “do not exceed a 5% position size” does not provide the same guarantee as implementing that restriction in deterministic code.
Two exchange models, one product
The platform had to work across fundamentally different venue architectures, each with its own liquidity and settlement behaviour.
02 / INSIDE THE PRODUCT
See the reasoning.
Not just the verdict.
Five views of the delivered platform.
Each shows another stage between research and execution.
One context, many sources.
Market prices, liquidity, indicators, news and sentiment are combined into the consistent context agents reason over.
THE DECISION PATH
Gather context→Agent research→Form thesis→Risk check→Authorize→Execute on-chain
03 / RECOMMENDATION VS AUTHORIZATION
AI proposes.
Deterministic code disposes.
The platform separates three responsibilities: AI generates the trading thesis, deterministic software decides whether the trade is allowed, and a dedicated execution layer handles blockchain interaction.
The result is a foundation for building AI-native trading products without giving a language model unrestricted authority over user capital.
Look under the hood ↘01
Agents form a thesis
Multi-agent research combines market data, indicators, news and sentiment into a reasoned trading proposal.
Routed to execution
The decision passes to the execution layer with its parameters fixed by the risk tier, not by the model.
Recorded with its reason
A rejected proposal keeps its reasoning and hash, so why the controls refused is as inspectable as why the agent proposed.
02
Deterministic controls decide
Position limits and risk rules are enforced in code, so approval or rejection is a guarantee rather than an instruction.
03
The execution layer acts
Only an authorized decision reaches the venue, through a layer built for blockchain interaction.
04 / THE TECHNOLOGY & THE REASON
Three tiers.
Three different jobs.
The stack follows the separation the product depends on: an experience tier, a reasoning tier, and an execution tier that is the only part holding authority to interact with a chain.
N
EXPERIENCE
Next.js + React Native
Web and mobile surfaces carrying market context, agent activity, decision records and portfolio state.
Why both?
Traders monitor from a desk and act from a phone; the decision record has to be legible on either.
APPLICATION
TypeScript + Fastify
The application tier serving the product surfaces and coordinating between reasoning and execution.
Why a separate tier?
It keeps the product experience independent of both model behaviour and chain latency.
PY
REASONING
Python + FastAPI
Runs the multi-agent research pipeline that combines market data, indicators, news and sentiment into a thesis.
Why isolate the agents?
Reasoning is probabilistic and slow; it must not sit in the path that enforces limits or submits a trade.
DECISION RECORD
PostgreSQL + Redis
Preserves every decision with its reasoning and content hash, alongside the fast-moving state the platform reads continuously.
Why hash the decision?
A preserved reasoning trail is only trustworthy if it can be shown not to have changed after the fact.
VM
CHAIN INTERACTION
viem
The execution layer's interface to blockchain networks and decentralized venues.
Why a dedicated layer?
Chain interaction is where authority over funds lives, so it is deliberately the narrowest, most explicit part of the system.
712
SIGNING
EIP-712
Typed structured signing for wallet interactions across the supported networks.
Why typed signing?
A user should be able to see what they are authorizing, not sign an opaque payload.
The architecture and product narrative reproduce the supplied project documentation. Interface values are product screen data and are not presented as investment results or evidence of future performance.
Explore the documented platform stack
- Experience
- Next.js · React Native · TypeScript
- Application
- Fastify · TypeScript
- AI & research
- Python · FastAPI · multi-agent research pipeline
- Data
- PostgreSQL · Redis · preserved decision records with content hashes
- Web3
- viem · EIP-712 typed signing · multi-chain venue integration
Technology names identify the documented implementation, not partnerships or endorsements.
A CONVERSATION ABOUT YOUR TRADING PRODUCT
Where should the model
stop having authority?
Bring one AI-assisted decision your product makes. In 20 minutes, we will map which part is reasoning, which part must be deterministic, and where the boundary belongs.
20 minutes · Your venues, controls and constraints