CoinyFarm: AI Yield Strategist
DeFi yield farming promises returns, but the reality for most users is a wall of protocol names, APY figures, and risk parameters that mean nothing without context. CoinyFarm came to us in 2023 with a dashboard that ranked farms by APY and lost users within seconds. The data was correct. The experience was hostile.
Context
CoinyFarm aggregates yield opportunities across multiple chains. The underlying data — TVL, APY, impermanent loss estimates, protocol audit status — was already being pulled from Subgraph endpoints into an Airtable-backed catalog. The frontend was built on Tilda. The problem was not data quality. It was comprehension. Users landed, saw a ranked table, and bounced.
Problem
DeFi yield options are genuinely hard to compare. A farm offering 40% APY on an unaudited protocol with $200K TVL is not the same as 12% on a battle-tested pool with $400M. Experienced traders read those signals instinctively. New users do not. They either chase the highest number and get rekt, or they see the complexity and leave.
The bounce rate confirmed this. Users were not rejecting the product — they were rejecting the cognitive load. The question was whether we could reduce that load without dumbing down the data.
Approach
We kept the existing data pipeline intact. Subgraph continued to feed Airtable, which remained the source of truth for farm metadata. What we added was a reasoning layer powered by OpenAI that sat between the data and the user.
For every farm in the catalog, the system generated a plain-English risk explanation. Not a generic disclaimer — a specific breakdown: what protocol this farm runs on, what the TVL implies about liquidity, whether the contract is audited, what the APY actually means given current reward token inflation. The model was prompted to be concrete and brief, and its output was cached per farm, refreshed when underlying parameters changed.
The dashboard was restructured on Tilda to lead with the explanation, not the APY. Ranking still existed, but each farm card opened with a one-sentence risk summary in natural language, followed by the numbers. The APY was still there — it was just no longer the first thing you saw.
What was built
- A Subgraph-to-Airtable ingestion pipeline, already in place, feeding farm metadata and on-chain parameters.
- An OpenAI-powered explanation layer that generated and cached per-farm risk summaries, refreshed on parameter change.
- A Tilda frontend reorganized around plain-English risk context, with APY and TVL as supporting detail rather than the headline.
- Ranking logic that factored in audit status and TVL alongside raw yield.
Results
Within the first month after launch:
- TVL up 9%. Users who understood what they were buying into stayed and deposited.
- Bounce rate down 8%. The plain-English lead reduced the exit-on-first-glance pattern.
- Average session length +42 seconds. Users were reading, not just scanning and leaving.
The numbers were modest and that was the point. This was not a viral growth hack. It was a comprehension fix. Users who understand risk make larger and more durable deposits than users chasing a number they do not understand.
Takeaway
The highest-leverage AI application in DeFi is not a trading bot. It is a translation layer. The data already exists and is already accurate. The gap is between what the data says and what the user understands. An LLM that explains risk in plain English, grounded in real on-chain parameters, closes that gap without inventing anything new. The user still makes the decision. They just make it informed.