This is really interesting and impactful piece of work. Well done! I guess this is similar to agent based analysis or discrete network analysis, both of which in the past required a very specialised skills, with a good dose of large and complex unstructured data extraction and normalisation which was completely out of reach unless you had serious amount of cash.
On a practical side, you mention 70 citations providing source data. What was the effort to source, prepare and judge the data at this scale? What percentage of that work is potentially automatable?
Thanks Jiri, it was great to realise another side of Constellation's value beyond pure agentic workflows!
Regarding data, the first pass was maybe half the final set, sourced in hours rather than weeks, with Claude handling search, collation and first-pass grading. The rest were pulled in as the model build demanded them, for example to resolve unit mismatch or to fill gaps in the data. Sometimes the first source was not authoritative enough to carry the figure it was holding up, which only came to light when we started to run the model.
The real effort was getting everything onto a baseline and deciding which figures withstand scrutiny. As for how much is automatable, search and collation already is, near fully. The AI surfaces the discrepancies well, which is half the reconciliation battle, but resolving them and judging what stands stayed human, and I'd expect it to stay that way.
Direction, what to point the engine at and what counts as a plausible shock, was fully human.
This is really interesting and impactful piece of work. Well done! I guess this is similar to agent based analysis or discrete network analysis, both of which in the past required a very specialised skills, with a good dose of large and complex unstructured data extraction and normalisation which was completely out of reach unless you had serious amount of cash.
On a practical side, you mention 70 citations providing source data. What was the effort to source, prepare and judge the data at this scale? What percentage of that work is potentially automatable?
Thanks Jiri, it was great to realise another side of Constellation's value beyond pure agentic workflows!
Regarding data, the first pass was maybe half the final set, sourced in hours rather than weeks, with Claude handling search, collation and first-pass grading. The rest were pulled in as the model build demanded them, for example to resolve unit mismatch or to fill gaps in the data. Sometimes the first source was not authoritative enough to carry the figure it was holding up, which only came to light when we started to run the model.
The real effort was getting everything onto a baseline and deciding which figures withstand scrutiny. As for how much is automatable, search and collation already is, near fully. The AI surfaces the discrepancies well, which is half the reconciliation battle, but resolving them and judging what stands stayed human, and I'd expect it to stay that way.
Direction, what to point the engine at and what counts as a plausible shock, was fully human.