Analyst Desk

One source of truth for the fund's deals and portfolio companies.

  1. Received files in by 7 AM
  2. Mapped by known layouts and by the intake agent
  3. Checked checks with flagged
  4. Tied out control totals tied
  5. PublishedEvery figure traced

Eleven files are waiting

This is one demo morning at a fund with three synthetic portfolio companies. Their systems sent eleven overnight files before 6 AM. Run the morning to receive, map, check, and tie them out, or press the right arrow for a guided tour.

  • sideoats_CR1_2026-11-04.csv
  • sideoats_CR2_2026-11-04.csv
  • sideoats_CR3_2026-11-04.csv
  • sideoats_CR4_2026-11-04.csv
  • sideoats_battery_CB1_2026-11-04.csv
  • sideoats_settlement_2026-11-04.csv
  • sideoats_market_price_2026-11-04.csv
  • kiskadee_invoicing_2026-11-04.csv
  • kiskadee_tms_2026-11-04.xlsx
  • possumhaw_night_audit_2026-11-04.txt
  • possumhaw_deposits_2026-11-04.csv

Brightrow Field Services

A carve-out of a commercial solar operations and maintenance business from a larger electrical contractor

The data room

Kiskadee Freight

Board meeting for October 2026

Headlines and bullets drafted by the board deck agent from a run recorded on October 29, 2026. The charts come from the ledger.

Kiskadee FreightBoard meeting for October 20261 / 5

October finished better than budget on loads, revenue, and the EBITDA estimate.

  • Loads totaled and finished .
  • Revenue was and finished .
  • Gross margin reached or of revenue, and the EBITDA estimate was at .
Loads
Revenue
Gross margin
Gross margin share
EBITDA estimate
On time
Every figure traces to the fund's ledger. All data here is synthetic.
Kiskadee FreightBoard meeting for October 20262 / 5

Gross margin finished as gains from loads and revenue per load outweighed the effect of costs.

  • Gross margin moved from a budget of to an actual of .
  • Loads added , and revenue per load added .
  • The effect of costs was .
Gross margin from budget to actual
Budget margin
Volume
Rate
Cost
Actual margin

Each step is a ledger value and the steps add up to the actual margin.

Table view
Gross margin from budget to actual
StepAmount
Budget margin
Volume
Rate
Cost
Actual margin
Every figure traces to the fund's ledger. All data here is synthetic.
Kiskadee FreightBoard meeting for October 20263 / 5

Operations delivered of loads on time in October with an empty miles share of .

  • On-time delivery was for the month.
  • The empty miles share was .
  • Loads totaled , and revenue per load was .
On-time share by week
Oct 1 to 7Oct 8 to 14Oct 15 to 21Oct 22 to 28

The partial last week of October appears only in the table view.

Table view
On-time share by week
WeekValue
Oct 1 to 7
Oct 8 to 14
Oct 15 to 21
Oct 22 to 28
Oct 29 to 31 partial
Every figure traces to the fund's ledger. All data here is synthetic.
Kiskadee FreightBoard meeting for October 20264 / 5

Cash collected of exceeded invoicing of  in October.

  • The company invoiced and collected .
  • Receivables were at month end, and days sales outstanding was .
Invoiced and collected by week
Oct 1 to 7Oct 8 to 14Oct 15 to 21Oct 22 to 28
  • Invoiced
  • Collected

The partial last week of October appears only in the table view.

Table view
Invoiced and collected by week
WeekInvoicedCollected
Oct 1 to 7
Oct 8 to 14
Oct 15 to 21
Oct 22 to 28
Oct 29 to 31 partial
Every figure traces to the fund's ledger. All data here is synthetic.
Kiskadee FreightBoard meeting for October 20265 / 5

The board may decide whether to set targets for on-time delivery, empty miles, and days sales outstanding.

  • On-time delivery was , and the empty miles share was .
  • Days sales outstanding was .
  • Gross margin finished .
Every figure traces to the fund's ledger. All data here is synthetic.

Hours back

Every workflow the desk runs is timed against the hours it takes by hand. Run times are measured, and the hours by hand and the review minutes are estimates you can change in the table.

Hours back this week
25.8

Hours by hand less review and run time

Hours back today
3.6

The morning report and the new layout

Hours by hand this week
30.0

What the same runs take when they're built by hand

Hours of review kept
4.2

A person still reviews every draft

WorkflowRuns this weekMeasured run timeEstimated hours by handEstimated review minutesHours back
The daily portfolio report317 seconds6.7
A new file layout mapped110 seconds1.3
The deal model's first draft114 seconds8.5
The memo's first draft172 seconds5.0
The board deck161 seconds4.2
This week25.8

Hours back is the runs times the hours by hand less the review and the run time. Blue marks an estimate as it does in the deal model, and totals update as you change them.

Hours by hand against hours with the desk

The daily portfolio report
6.7 h back
A new file layout mapped
1.3 h back
The deal model's first draft
8.5 h back
The memo's first draft
5.0 h back
The board deck
4.2 h back
  • By hand
  • With the desk and a review

Each pair compares the hours a workflow's runs take by hand with the hours they take with the desk and a review.

How it works

One ledger sits under every deliverable. Agents do the reading and the drafting, code checks every number, and nothing reaches a person without passing the gate.

The architecture

Portfolio files

Eleven files a morning

The data room

The seller's financials, customers, and teaser

Intake

Parsers for known layouts and an agent for new ones

Checks

Ranges, keys, and tie-outs

The ledger

Every file, fact, and derived value with its lineage

1,065 files, 7,555 facts, and 1,902 derived values

The gate

No typed numbers

Daily report

Commentary by an agent

Deal model

Assumptions by an agent

Investment memo

Drafted by an agent

Board deck

Headlines by an agent

The spark marks a step where an agent works, and code bounds every agent.

The agents and their limits

Each agent has one job, a fixed input, and an output schema, and code enforces its limits. None of them can publish anything on its own.

AgentWhat it seesWhat it returnsLimits enforced in code
Intake mapperA new file's header and sample rowsMappings and formatsAccepted only when every check and the control total pass, and saved as a layout after that
Daily commentaryEach company's KPIs, variances, and checksTwo to four sentences eachOnly placeholders get past the gate
Deal analystThree years of data room historyCited operating assumptionsEach value must sit in a band set by history, and manual overrides win
Memo writerThe model's outputs, the adjustments, the customers, and the claims checkEleven cited sectionsThe gate checks every figure, and the risks section must cite every claim
Board deck writerKiskadee's month and its variancesFive headlines with two or three bullets eachCharts come only from the ledger

Recorded once

The morning's five agents ran on the models routing picked for them, and every reply passed the same validation before it was saved.

In the recorded runs a morning's five agent calls came to $0.52 at API prices.

Replayed here

The page replays those runs and checks each one again on every build. Nothing on it calls a model or waits on the network, and a visit costs nothing. The invented company names changed after the runs were recorded, and the swap touched nothing else in them.

Model routing

Every model here took the same tests with the same prompts and the same checks. Each agent runs on the cheapest model that clears its quality bar, and the next one that clears it takes over when a call fails. The daily commentary, the board deck writer, and the memo writer stay on Claude Opus 5.5 because passing validation isn't the same as reading well.

The blended morning
$0.51

27% below Claude Opus 5.5 on every agent

Every agent on Claude Opus 5.5
$0.70

The same five agent calls on one model

Intake on its route
12 of 12

Layouts mapped exactly by Claude Haiku 5.5

How the blend adds up

ModelRunsModel time a morningCost a morning
Claude Opus 5.5Commentary, board deck, and memo2.4 min$0.51
Claude Haiku 5.5Intake12 s$0.0030
GPT-6 LunaDeal analyst18 s$0.0013
The blended morning2.9 min$0.51
Claude Opus 5.5 on every agent2.8 min$0.70
Claude Haiku 5.5 on every agent7.3 min$0.065
Intake quality against cost

Layouts mapped exactly

89101112
Within one of the best
Claude Haiku 5.5Gemini 3.1 ProClaude Opus 5.5Gemini 3.8 FlashGPT-6 AstraClaude Sonnet 5.5GPT-6 Luna
$0.001$0.01$0.10

Cost of one mapping on a log scale

  • Anthropic
  • OpenAI
  • Google
  • The ringed dot is the intake agent's model

2 of the 7 models cleared the intake bar, and Claude Haiku 5.5 did it for the least at $0.0030 a mapping.

Table view
Intake quality against cost
ModelVendorIntake exactCost of one mapping
GPT-6 LunaOpenAI9 of 12$0.00055
Claude Haiku 5.5Anthropic12 of 12$0.0030
Gemini 3.8 FlashGoogle10 of 12$0.010
Gemini 3.1 ProGoogle12 of 12$0.030
Claude Sonnet 5.5Anthropic9 of 12$0.041
GPT-6 AstraOpenAI9 of 12$0.043
Claude Opus 5.5Anthropic10 of 12$0.075

Where each agent runs

AgentRuns onWhyIf a call fails
Intake mapperClaude Haiku 5.5The cheapest of the 2 models that cleared its barGemini 3.1 Pro
Daily commentaryClaude Opus 5.5Held for writing quality because a person reads itGPT-6 Luna and 5 more
Deal analystGPT-6 LunaThe cheapest of the 7 models that cleared its barClaude Haiku 5.5 and 5 more
Board deck writerClaude Opus 5.5Held for writing quality because a person reads itGPT-6 Luna and 5 more
Memo writerClaude Opus 5.5Held for writing quality because a person reads itClaude Haiku 5.5 and 4 more

How each model did

ModelVendorIntake exactProse valid first tryModel time a morningCost a morningRouted agents
Gemini 3.1 Progemini-3.1-pro-previewGoogle12 of 124 of 43.6 min$0.52None
Claude Haiku 5.5claude-haiku-5-5Anthropic12 of 123 of 47.3 min$0.065Intake
Gemini 3.8 Flashgemini-3.8-flashGoogle10 of 124 of 45.8 min$0.20None
Claude Opus 5.5claude-opus-5-5Anthropic10 of 124 of 42.8 min$0.70Commentary, board deck, and memo
GPT-6 Lunagpt-6-lunaOpenAI9 of 123 of 42.3 min$0.0086Deal analyst
GPT-6 Astragpt-6-astraOpenAI9 of 123 of 42.1 min$0.55None
Claude Sonnet 5.5claude-sonnet-5-5Anthropic9 of 122 of 44.7 min$0.65None

The intake agent has to land within one layout of the best model's score, and none of its misses can be a mapping the checks would have let through. Each prose agent has to pass validation on every run within the one repair. The comparison ran on October 28. Costs use each vendor's list prices that day, and Claude's come from what the CLI reported.

Code excerpts

These excerpts of the desk's code are copies made for this page. The page reads only the copies and never the code that runs the desk. Open one below or click a box in the diagram above to see its excerpt.

Results from the tests and evals

Each result here is read from a file the tests, the evals, or the recordings wrote. None of them is typed.

Planted errors caught
18 of 18

On held-out seeds 3119, 6427, and 8831 after the checks were frozen

False alarms
0

Across the same held-out seeds

Look-alikes left alone
9 of 9

Normal days that only look like errors

New layouts mapped exactly
10 of 12

Held-out layouts the agent had never seen, and the other 2 still parsed every row like the answer key

Tests passing
148 of 148

In the last full run on November 20

Workbook against Python
74 values

Each formula's value matches the Python model to the cent

LibreOffice recalculation
383 formulas

Recalculated from scratch, and each one matched to the cent

Agent runs recorded
17

With Claude Haiku 5.5, Claude Opus 5.5, and GPT-6 Luna from October 27 to 29

How I'd roll it out at a fund

  1. The first thirty days

    One portfolio company's daily report on its real files with every tie-out in place.

  2. The first sixty days

    The model and memo template on the fund's real deal process with the gate on every figure.

  3. The first ninety days

    Hours back measured on every automated workflow.