Inspiration
Leaderboards lie by omission: a wallet can top the monthly PnL board while being deeply negative over its lifetime (we verified exactly this in our own research). What a researcher actually needs is not another ranking, but a boring, reliable answer to one question: what changed since last week, and is this wallet still the same kind of actor it used to be?
What it does
WatchDigest monitors a watchlist of wallets using only public data:
- A deterministic rule engine compares each wallet against its stored baseline (position value, trading cadence, price distribution, concentration) and fires typed alerts — dormancy, sweep behaviour, cadence collapse to bot-like intervals, concentration spikes.
- An LLM layer (Nemotron, served via Nebius Token Factory) turns the structured alert data into a readable weekly briefing.
- A verification pass then checks the briefing: every number in the generated text must exist in the source data. If a number can't be traced, the briefing is rejected and never published — the LLM narrates, it never calculates.
How we built it
Python (standard library core) for collection and the rule engine — already working and self-verified against hand-computed values from real public data. The new layer for this hackathon is the Nemotron briefing module on Nebius Token Factory (OpenAI-compatible API), plus the numeric-claim verifier that guards its output. Data sources: public APIs only; no accounts, no keys to the markets, no trading — this is a read-only research tool, and it says so in its own output.
Challenges we ran into
Keeping an LLM honest about numbers. Our fix is architectural, not prompt-based: the model receives only the alert JSON, and a separate checker rejects any figure that does not appear in it.
Accomplishments we're proud of
The base engine already caught real behavioural changes in live data — a formerly human-paced wallet whose median trade interval collapsed to zero seconds while 99.8% of its trades became near-settlement sweeps.
What we learned
For research agents, "the model said so" is not a source. Splitting detection (deterministic) from narration (LLM) from verification (deterministic) makes the whole system auditable.
What's next
Scheduled weekly runs, more wallet cohorts, and the same detect→narrate→verify pattern applied to other public-data domains.
Live-run verification (2026-10-28)
WatchDigest was run end-to-end against the real Nebius Token Factory API using nvidia/Nemotron-3.5-Super-VL. The generated briefing passed the numeric verifier with all 39 figures traced to the source data. Measured usage was 1,459 prompt tokens and 2,418 completion tokens (2,073 of them reasoning tokens), about $0.0026 per briefing at list prices. One practical finding: this model spends most of its budget on reasoning tokens, so a 1,200-token cap returned an empty message; the pipeline now requests 4,000.
Log in or sign up for Devpost to join the conversation.