Tensor OKO: Trade Whisperer
Whale tracking is one of the oldest signals in crypto trading. When large wallets move, price follows. The data is public — every transfer is on-chain — but raw on-chain data is a firehose. Tensor OKO launched in 2022 to solve the problem that every whale-alert bot has: the signal is there, but by the time a human reads the raw stream, figures out what it means, and decides whether to act, the move is over.
Context
Tensor OKO is a Telegram bot built for active crypto traders. The stack is Python and aiogram for the Telegram interface, ClickHouse for high-throughput event storage, and OpenAI for the reasoning layer. The bot monitors on-chain whale movements in real time and pushes alerts to subscribed traders. The core challenge was not capturing the data — that problem is solved. The challenge was making the data actionable before it goes stale.
Problem
Traders subscribed to whale-alert feeds drown in noise. A large transfer from an unknown wallet to an exchange could mean an imminent sell, or it could be an internal rebalance. A series of withdrawals from a known accumulator wallet could signal conviction, or could be movement to cold storage. The raw data does not tell you which. It gives you an address, an amount, and a timestamp. The trader has to do the interpretation, and they have to do it fast.
Most whale-alert bots solve this by adding more filters: only show transfers above $1M, only show flows to known exchanges, only show wallets on a watchlist. Filtering reduces noise but also removes signal. The trader still has to interpret whatever makes it through. The result is a feed that is either too noisy to use or too filtered to be useful.
Approach
We did not filter harder. We added interpretation. Every significant on-chain event that enters ClickHouse is passed to an OpenAI-powered reasoning layer that produces a plain-English TLDR within 60 seconds of the event. The TLDR is not a summary of the transaction — it is an interpretation: what wallet, what pattern, what it likely means given recent activity from that address, and whether it correlates with other movements in the same window.
The reasoning layer has access to historical context stored in ClickHouse. When a wallet moves, the model can see that wallet's recent behavior — is this a one-off transfer or part of a pattern? Is this an exchange inflow that historically preceded a dump? The model does not predict price. It compresses the interpretation work the trader would have done manually into a few sentences, delivered before the window closes.
The alert is pushed to the trader's Telegram chat via aiogram with the TLDR, the raw transaction details, and a link to the on-chain explorer. The trader still decides. They just decide with context, in seconds, instead of starting from raw data.
What was built
- A Python service ingesting on-chain whale events into ClickHouse for high-throughput storage and historical querying.
- An OpenAI reasoning layer that generates a contextual TLDR for each significant event within 60 seconds, using ClickHouse history for pattern context.
- An aiogram-based Telegram bot that delivers formatted alerts with TLDR, raw transaction data, and explorer links to subscribed traders.
- A subscription and alert-routing system supporting approximately 120 active traders.
Results
- Analysis time cut 32% for approximately 120 active traders. The TLDR eliminated the manual interpretation step that consumed most of the reaction window.
- 5,000+ alerts delivered monthly. Volume that would be unmanageable as raw data became scannable as interpreted summaries.
- 41% retention after 30 days. For a trading tool in a market where most bots are installed, used once, and forgotten, holding nearly half of users past the 30-day mark indicated the alerts were genuinely useful, not just novel.
Takeaway
The bottleneck in real-time trading signals is not data access. It is interpretation speed. Raw on-chain data is instant and free. What costs time is turning that data into a decision. An LLM that reads the event, pulls historical context, and returns a plain-English interpretation within 60 seconds does not replace the trader's judgment — it gives them the context they need to apply that judgment before the opportunity closes. The trader still pulls the trigger. The AI just loads the gun faster.