← Back to work
Shipped · Vercel · Three.js

FinSight + Web

Not a dashboard — a scroll-driven research monograph that turns the FinSight pipeline's findings into something you can explore, told in eight acts.

ProblemA research pipeline's real findings usually stay locked in a notebook that only the author can read.
ApproachRebuild the same 14,584-transcript corpus as an explorable, scroll-driven story — eight acts, one shared 3D dataset.
ResultThe same honestly-reported signal from the research pipeline — concentrated in the energy sector — made spatially explorable.
SignificanceProof that rigorous research and an inviting, explorable interface aren't a tradeoff.
8Narrative acts
+0.31Energy-sector IC
3DExplorable signal field
1Shared dataset, 8 views

What Is This

This is the explorable counterpart to FinSight's research pipeline — not a dashboard with filters and dropdowns, but a guided, scroll-driven monograph. The same earnings-call corpus that backs the research is restaged eight times, each act revealing it from a different angle.

Why It Was Built

A walk-forward backtest and an honest null result are easy to bury in a PDF. FinSight + Web exists to give those findings a form a non-technical reader can actually sit with and explore — without ever overstating what the underlying research found.

How It Works

CORPUS & READING RETRIEVAL & SIGNAL FIELD HONEST MODELS & ATTRIBUTION SECTOR & VERDICT

The eight acts group into four movements: the corpus is archived and read for tone, then turned into a searchable signal field; walk-forward models are tested honestly and their reasoning attributed; finally the signal is located sector-by-sector and condensed into a verdict. One dataset, four movements, eight ways of looking at it.

Key Result

+0.31 ICEnergy sector — where the signal lives

The same finding from the research pipeline, now visualized as a sector constellation on the live site — not a new number, a new way to see it

Why It Matters

This is what deploying honest research looks like when it's given a real interface: the same walk-forward discipline that refuses to overstate a thin edge, presented so a reader can feel where the signal is strong and where it genuinely isn't — archival aesthetic, real numbers, no inflated claims.

↗ View live monograph → See the research pipeline