Aurora
Glass-box Quantitative AI for Humans and Agents
11 followers
Glass-box Quantitative AI for Humans and Agents
11 followers
Aurora is glass-box quantitative AI that runs locally. Drop in a dataset and 24+ research-grade methods do the real math — anomalies, causality, regimes, forecasts — then every finding is cited to the method and source behind it. A live "0 fabricated" chip guarantees no invented numbers. Cloud LLMs guess; Aurora computes. Includes an MCP server + Python SDK so AI agents can call it for verified math. 100% local, no cloud, no API keys. Open source, Apache 2.0.
Free
Launch Team
World model AI : Sharm SEKAISimulate your idea on a living world.
Promoted
How does Aurora actually guarantee no fabricated numbers when it pulls from local datasets, and does that guarantee still hold if the data itself is incomplete or messy?
@durudly5 Great question, and it's really two questions, so let me take both.
On the guarantee itself: it's architectural, not a filter. The LLM in Aurora never generates a number. All quantitative claims come from classical methods that actually compute on your data (robust z-scores, change-point detection, HMM regimes, Granger tests, etc.), and each finding is a structured object carrying its method, threshold, and evidence. The LLM's only job is narrating those already-computed findings, and the synthesis layer is restricted to claims it can cite back to a computed artifact or a knowledge bank entry. So "0 fabricated" isn't the model behaving well, it's the model never being asked to produce numbers in the first place. There's nothing to hallucinate because generation and computation are separated by design.
On messy or incomplete data, the honest answer: the guarantee holds, but it's important to be precise about what's guaranteed. Aurora guarantees no invented numbers, not that messy data yields great findings. Garbage in still means limited insight out. The difference is what happens at that boundary. When data lacks structure a method needs (no time axis, too few samples, negative values where a method can't handle them), Aurora skips that method and tells you why, visibly, in the output. When it samples or times out, that's disclosed too. So on messy data you get fewer findings with honest skip reasons rather than confident conclusions built on quicksand. Frankly, messy data is where the design matters most, because that's exactly when a generative tool will happily fill the gaps with plausible-sounding numbers.
If you want to check the claim rather than take my word for it, the whole thing is open source, and every run produces a signed bundle you can audit. Would genuinely love to hear what happens if you throw your messiest dataset at it.
How does the "0 fabricated" chip actually work under the hood, like is it catching hallucinations at the method level or just flagging outputs that lack source citations?
@ftavukcuyo70638 Good question, and the honest answer is it's neither a hallucination catcher nor just a citation check. It's layered, and the chip is the audit result at the end.
Layer one is prevention by architecture. The LLM in Aurora is never asked to produce a number. Every quantitative claim originates from a deterministic method that actually ran (Hampel z-scores, change-point detection, HMM, Granger, etc.), and each finding exists as a structured artifact first: method, value, threshold, evidence, citation. The synthesis model writes around those artifacts under strict grounding, so there's no generation step where a number could be invented in the first place.
Layer two is post-hoc verification. After synthesis, the output is checked against the run's artifacts and the knowledge bank: numeric claims have to trace to a computed finding, and cited claims have to resolve to a real entry. Anything that can't be grounded gets flagged or dropped rather than shipped. So it's not "flag outputs lacking citations" cosmetically, it's "verify every claim resolves to something that actually exists in the run."
The chip is the count of claims that failed that verification. It reads 0 because the architecture makes fabrication structurally hard and the verifier confirms it, not because a classifier guessed the output looked truthful. That's also why I call it contractual: if it ever reads nonzero, that's the system telling you something got through, visibly, instead of hiding it.
Finally tried Aurora on a messy returns dataset and the regime detection flagged a transition I'd been manually squinting at for weeks. Citations to the actual method make it feel honest in a way most quant tools don't.
@adelicay27612 appreciate you taking the time to try this out!
Ran the changepoint and causal inference tools on a messy sales dataset and the citations to the underlying method made it really easy to sanity check the output. Love that the 0 fabricated chip is front and center instead of buried somewhere.
@nilgnxmc0 appreciate the feedback and support!