Siarhei Mardovich

Uncertainty as Signal

Climate-risk decision intelligence for capital markets. A 90-meter flood simulation shows whether water reaches a building’s electrical substation or only inundates the parking lot. That single physical distinction is the underwriting decision.

  • Climate risk
  • Probabilistic signal detection
  • Capital decisions
  • Jupiter Intelligence

Substation or parking lot

01 · The consequence

The approach crystallized in the flood mesh view: a 90-meter hydrodynamic simulation that shows whether a projected 2-meter flood reaches a building’s electrical substation or merely inundates the parking lot.

That single physical distinction changes the underwriting decision, the loan terms, and what an analyst can defend in front of a credit review board. It is not a difference in score. It is a difference in what breaks.

Fig. 01 Hydrodynamic Mesh Inspector

90-meter hydrodynamic simulation. Substation vs. parking lot. That distinction is the underwriting decision. .

Swap the peril for a volatility regime or a default cascade and the logic holds: widen the distribution at the point of decision so capital lands on a known shape of loss.

What 90-meter resolution buys

02 · The design decision

The mesh inspector runs a 90-meter hydrodynamic flood simulation at building-footprint resolution. It determines whether water reaches the electrical substation or only the parking lot, and lets an underwriter model how a sea wall or a raised pad changes Average Annual Loss. Accept or reject becomes a negotiable engineering conversation.

The most consequential design decision in the Analyze phase was not the resolution of the flood mesh. It was what to put next to it. The easy path was a side panel of summary statistics: probability, expected damage, a color-coded risk band. It would have looked complete and loaded fast. It would also have hidden the exact thing an underwriter needs to defend a covenant: the physical mechanism of loss.

I chose to anchor the view on a building-footprint simulation with toggles for depth, velocity, and scenario, paired with an asset-overlay layer that distinguishes critical structures from parking and landscaping.

The trade-off

Cognitive load. The screen asks more of the analyst than a single score would. The gain is that the analyst can answer a credit committee question with a physical fact rather than a rating.

The figure an analyst can defend

03 · The numbers

Water reaches the substation at 1.8 meters under SSP5-8.5 by 2050, and a raised pad would shift the loss curve by an estimated 35 percent. That is the sentence an analyst carries into a credit committee, and every term in it is traceable back to the simulation that produced it.

90 m Hydrodynamic simulation resolution, at building-footprint scale
1.8 m Water depth at the electrical substation under SSP5-8.5 by 2050
35% Estimated loss-curve shift from a raised pad
22,000 Data points generated per location across peril, scenario, and horizon

Underwriters I worked with during the Jupiter engagement did not ask for prettier charts. They asked for defensible detail they could put in front of a credit officer without becoming hydrologists. The mesh inspector is the answer to that request. It keeps the model honest by forcing the interface to show the spatial assumption behind the number, which is also where model risk lives in climate analytics.

Keeping the range visible

04 · The disclosure layer

Most climate-risk products flatten a fat-tailed distribution into one score and invite a decision-maker to treat it like a credit rating. That collapse turns a known risk into an unpriced one. And the bank must still defend the resulting capital decision to a credit board, a model-risk auditor, and a TCFD or ISSB examiner.

An analyst trained on the precision of historical financial statements is handed a tool that says there is a 1-in-100-year flood probability under a high-emissions scenario by 2050, and is asked to justify a billion-dollar capital deployment to an investment committee. Each asset in a global portfolio carries 3 analysis axes (peril, scenario, and time horizon) and generates roughly 22,000 data points per location. The design problem was never how to display them. It was where, along the path to a decision, the cognitive load becomes dangerous.

The answer is to make uncertainty legible rather than hide it. An analyst who can see the range of outcomes and the mechanism that produces them is far more equipped to make a defensible decision than someone handed a single score with no idea where it came from.

Trust architecture as an operational layer

The platform is structured as a logical progression, not a static dashboard. Every output is traceable from raw signal to scenario to model lineage to financial outcome, so an analyst can answer “where did this number come from” in a credit review, a model-risk audit, or a regulatory filing. Reporting is layered rather than flat: high-level scores for screening, engineering-grade flood meshes for site-level decisions, and a unified control panel that toggles peril, scenario, and horizon without losing analytical context.

The connective tissue is an ontology: an operational layer for the institution. The 12 instruments below walk that layer as the analyst’s actual decision journey, from orientation to continuous governance. Every instrument exposes both what the risk is and what the analyst does about it.

Fig. 02 Control panel

Peril, scenario, and time-horizon parameters in 1 decision-ready view. .

Where the commitment point is chosen

The decomposition studio splits a loss metric into model, scenario, data, and temporal uncertainty, so a committee can target its skepticism. If 40 percent of the uncertainty is scenario choice, explore the range; if it is asset data, commission a survey. Then choose where on the distribution to commit capital, and record that choice for audit.

12 instruments across 6 phases

05 · The decision journey

The instruments are organized into 6 phases that mirror how an analyst actually moves from orientation to continuous governance. Each one states the decision it supports, how it reads on screen, and opens standalone for a closer look.

Phase I · Orient

01

The digital twin of the balance sheet under climate stress. Renders the climate-exposed book as objects, links, and actions. A commercial mortgage links to a flood zone, which links to a scenario, which links to a capital reserve. Click a node and the risk pulses through its supply-chain links to a revenue-dependent loan, so propagation is visible rather than assumed.

Supports TCFD Governance: know what you own and how risk flows through it. Reads as an interactive map of the balance sheet under climate stress.

02

Where did this number come from? Traces any disclosed figure back through model transform and scenario parameters to raw signal, with model version, validation status, and a compounding confidence halo at every node. An audit becomes a navigable conversation instead of a document hunt.

Supports model risk management and audit defense under SR 11-7. Reads as: click a reported loss number and watch its derivation unwind.

Phase II · Detect

03

22,000 outputs per location, compressed into a single read. A polar instrument where spokes are scenario pathways (RCP 2.6 through RCP 8.5, NGFS orderly through hot-house) and rings are horizons from 2030 to 2100. Scenario divergence becomes spatial separation, surfacing the assets where the pathway assumption, not the asset, drives the decision.

Supports portfolio screening: find where scenario choice changes the answer. Reads as scenario divergence made visible as geometric distance.

04

The portfolio’s nervous system. Most tools only address chronic risk. This closes the acute gap: it streams near-real-time hazard signals (fire detection, flood gauges, hurricane tracks) against portfolio positions, estimates impact probability at 24, 48, and 72 hours, and recommends protective action while the event is still unfolding.

Supports acute physical risk: respond within hours, not quarters. Reads as a command center for events as they happen.

Phase III · Analyze

05

Substation vs. parking lot. That distinction is the underwriting decision. 90-meter hydrodynamic flood simulation at building-footprint resolution. It determines whether water reaches the electrical substation or only the parking lot, and lets an underwriter model how a sea wall or a raised pad changes Average Annual Loss. Accept or reject becomes a negotiable engineering conversation.

Supports underwriting: deductible, premium, and protective-investment terms. Reads as a physical mechanism of loss you can defend, not a score.

06

Widening the signal back out at the point of decision. Decomposes a loss metric into model, scenario, data, and temporal uncertainty, so a committee can target its skepticism. If 40 percent of the uncertainty is scenario choice, explore the range; if it is asset data, commission a survey. Then choose where on the distribution to commit capital, and record that choice for audit.

Supports investment-committee skepticism, directed where it belongs. Reads as: choose your commitment point on a known distribution.

Phase IV · Decide

07

Scenario-based capital deployment as a navigable conversation. Reweight sectors, set concentration limits, and buy protection while risk-adjusted return, the efficient frontier, and a TCFD-alignment score update live. Every action produces both a financial outcome and a filing narrative, so the strategy discussion and the regulatory filing evolve from one instrument.

Supports TCFD Strategy: test strategy resilience under scenarios. Reads as a decision forge for portfolio-level capital response.

08

A policy change made visible before it becomes reviewer behavior. A dual-handle instrument bins the book into Accept, Review, and Act. Drag a threshold and see how many loans move zones, and whether the bank has the analyst capacity to process them. The calibration becomes a visible, auditable decision rather than a number in a memo.

Supports risk governance: set the trigger and own its operational load. Reads as: calibrate the trigger, see the workload it creates.

09

Translating climate signal into contractual protection. Bridges the gap between a risk team that speaks in AAL and TVaR and a legal team that writes binary covenants. Define triggers (flood AAL over 75bps, wildfire probability within 10km over 20 percent annualized), link them to protective actions, and stress-test against historical and projected scenarios.

Supports climate-adjusted lending: encode risk into binding terms. Reads as a rule builder that turns a signal into a covenant.

Phase V · Disclose

10

From analysis to audit-ready narrative in one continuous flow. Generates TCFD-aligned (and ISSB-transition-ready) narrative across all 4 pillars, with every claim linked to its supporting evidence and a reviewer mode that shows exactly what an external examiner would see. Not a generic template: a traceable document where the governance section references logged committee actions and the metrics section cites model versions.

Supports compliance: produce defensible disclosure from the analysis itself. Reads as the same analysis, attested and filing-ready.

11

What did the reviewer see, weight, and sign at the moment of record? An immutable, tamper-evident timeline of every capital decision, threshold change, covenant set, and filing approved. Each event captures who decided, what data they saw, what weights they applied, what confidence they accepted, and a signature. An examiner does not trust the model; they trust the process around it.

Supports examination: defend the process, not just the model. Reads as a defensible record, replayable to the moment of decision.

Phase VI · Evolve

12

The model that does not know it is drifting is the most dangerous model. Tracks 3 kinds of drift against calibrated thresholds: data drift (are the inputs still representative?), model drift (is predictive performance degrading?), and scenario drift (are the chosen scenarios still aligned with the latest IPCC and NGFS guidance?). When drift crosses a threshold, it opens a review with the diagnostic context pre-populated.

Supports model governance: know before the regulator does. Reads as a system that watches itself produce outputs.

What a Tier 1 bank or insurer does with it

06 · The institutional payoff

The layer is designed so the bank can price a flood signal as a basis-point input and an analyst can defend it without being a climate scientist.

Basis points, not surprise
The bank prices a flood signal as a basis-point input rather than absorbing it as an unexpected loss.
Mechanism visible
The investment committee sees the physical mechanism of loss, not a black-box risk rating. Evidence: the flood-mesh view.
Passes MRM review
Analysts defend a model-risk audit by tracing lineage and assumptions, without being climate scientists. Evidence: the model-lineage tracer.
A line item to evolve
Average Annual Loss becomes a number the firm can act on, disclose, and evolve.
AI model risk aware
Drift, validation status, and scenario alignment are surfaced so the model does not become the unpriced risk.
Filing-ready
TCFD and ISSB narratives link directly to the analysis that produced them.

How each phase maps to what an examiner asks for

Each phase maps to clauses that examiners actually ask for. The table shows which interface output answers which regulatory expectation.

Phase TCFD / ISSB pillar SR 11-7 expectation What the interface produces
I · Orient Governance: know what you own Model inventory and scope Portfolio ontology explorer with traceable asset-to-model links
II · Detect Strategy: identify risks Model use and limitations Hotspot radar and signal dashboard with scenario divergence visible
III · Analyze Metrics and targets Validation and testing Mesh inspector and decomposition studio exposing assumptions
IV · Decide Strategy: manage risks Monitoring and exception management Allocation sandbox, threshold instrument, and covenant designer
V · Disclose Disclose across all pillars Audit trail and documentation TCFD composer and audit defense timeline with immutable records
VI · Evolve Metrics and targets Model validation and ongoing monitoring Drift monitor with data, model, and scenario drift thresholds

Shipped Jupiter Intelligence evidence

07 · Provenance
Role
UX Designer, Design Systems and Data Visualization, focused on the uncertainty-visualization layer. Owned information architecture for multi-peril, multi-scenario, multi-horizon climate data.
Approach
Trust architecture for a probabilistic input that must defend a deterministic capital decision. Every output traceable from raw signal to scenario to model lineage to financial outcome.
Outcome
Underwriter requests for defensible detail shaped the control panel and flood-mesh inspector, alongside the domain model, service blueprint, and journey work shipped during the engagement.
Context
Jupiter Intelligence, January 2022 to July 2023. One engagement in 15 years designing decision interfaces across capital markets, climate risk, and enterprise data products.

From a domain model to a flood mesh

Before the first wireframe, I built a domain model of how a physical climate event at a remote site cascades through global trade flows to a balance sheet in London. It became a service blueprint, then an end-user journey mapped as a mental space: the sequence of questions, hesitations, and commitments an analyst moves through from the first hazard flag to a signed capital position. That map is why reporting is layered rather than flat.

Fig. 03 Domain model

Domain model mapping signal to financial outcome. Built before the first wireframe. .

Fig. 04 End-user journey

The end-user journey as a mental space. It locates where confidence breaks down. .

Fig. 05 Service blueprint

Service blueprint: hazard awareness to financial impact to resilience strategy. .

Engagement timeline

The shipped Jupiter artifacts are separated from the current case studies so the reader can tell source engagement evidence from portfolio demonstrations.

Q1 2022
Discovery. Domain interviews, regulation mapping, analyst shadowing.
Q2 2022
Domain model. Signal-to-balance-sheet ontology. Shipped Jupiter artifact.
Q3 2022
Blueprint. Service blueprint and mental-space journey. Shipped Jupiter artifact.
Q4 2022
Hi-fi. Control panel, surface views, filing layouts. Shipped Jupiter artifact.
Q1 2023
Flood mesh. Hydrodynamic mesh inspector at building scale. Shipped Jupiter artifact.
2026
Case study. 12 single-file instruments and this case page.
Build and tooling
  • React, TypeScript, Vite single-file prototyping, no-backend prototypes
  • D3.js custom visualization, Leaflet and Chart.js, geospatial rendering
  • Live climate models and satellite imagery wired into the prototypes from day one, so the designs were tested against real institutional data before engineering handoff
  • WCAG AA accessibility

From signal to signature

08 · Close

The 12 instruments are a decision journey, not a feature list. Orientation establishes what is owned and where each number comes from. Detection watches for chronic and acute risk. Analysis makes the mechanism legible. Decision commits capital on a known shape of loss. Filing turns it into an immutable record. Evolution keeps the models honest.

Uncertainty is not a problem to eliminate. It is a signal to make legible. In any high-stakes financial interface, trust architecture is not a layer on top of the model. It is the structure that decides whether the model gets used at all. We learned that analysts did not want another score; they wanted the spatial assumption behind the number exposed.

What I would measure next: committee questions answerable from the screen without escalation to a climate specialist; time from hazard flag to a filing-ready position; and audit queries resolved inside the lineage tracer without a document hunt. The engagement shipped the artifacts; measuring their operating effect belongs to the next deployment.

Contact

Open to principal and staff product design roles. Greater New York City Area, remote or hybrid preferred, onsite flexible.