Direction 01

AI-accelerated tracker pipelines

Your data scientists build s from , , and models. AI helps them draft and run those pipelines faster, grounded in ODP methods, with a human sign-off on every .

Tap or hover any underlined word for a plain-English explanation.

01

Spin up a new company KPI tracker

Each new means wiring , , your , , and . Most of that work follows patterns your team already uses, but it still takes days of manual .

Solution: A agent indexed on ODP's reads the scientist's intent, retrieves similar trackers and internal formulas, drafts the full , and opens a . The scientist reviews, edits, and signs off before anything ships to .

Tools: , , , ,

How it looks in practice

Claude Code · New tracker draftODP codebase

1 · Scientist intent

"Build an orders tracker for RetailCo from UK receipt panel + web scrape. Match our retail methodology."

2 · Codebase blocks retrieved

odp/normalize/receipt_panel.py

Reused from ApparelCo tracker

odp/models/orders_v3.py

Proprietary blend formula

odp/backtest/point_in_time.py

House validation harness

3 · Pipeline scaffold generated

Scrape

UK receipt + web

Clean

ODP normalize blocks

Model

Multi-signal blend

Backtest

Out-of-sample check

Also generated

  • · Unit tests for ingestion + normalization
  • · Methodology note for audit trail
  • · Backtest report vs last 8 quarters
PR #412 · draft ready
Scientist sign-off

Nothing ships to production until a data scientist reviews, edits, and approves the merge.

02

Keep alternative data feeds running

, web , and break when a site or changes. A silent gap can pollute before anyone notices.

Solution: Always-on monitoring detects breaks. An agent diagnoses the issue using ODP's patterns, proposes a fix as a , and pings the owner. The scientist approves the . No more finding broken feeds by accident downstream.

Tools: , , ,

How it looks in practice

Feed monitor · UK receipt panelSelf-heal loop

1 · Break detected

Scraper stopped · 6 hours ago

Vendor portal HTML changed. Selector .txn-row no longer matches. Zero rows ingested since 02:00 UTC.

2 · Impact on estimates

3 retail trackers depend on this feed. Daily orders estimate for RetailCo would gap without intervention.

3 · Agent diagnosis + fix

Matched ODP scraper pattern · retail_receipt_v2

Before

.txn-row

After

.transaction-item

PR #891 · fix selectors
Owner notified

Scientist approves merge. Feed resumes before the next daily estimate run.

Effect: Scientists spend less time on and more on . New trackers move from intent to reviewable draft in hours, not days, without sacrificing .

Direction 02

Data quality & research automation

ODP ingests a huge volume of every day. AI watches those feeds, flags what looks wrong, and drafts research output your analysts can review before it reaches clients.

Tap or hover any underlined word for a plain-English explanation.

01

AI data quality & anomaly detection

Dozens of update daily across 250+ companies. When a source goes , KPIs , or a signal looks wrong, scientists cannot check every feed by hand.

Solution: AI automatically checks incoming feeds and flags broken sources, missing data, strange movements, or possible . The outcome is alerts, priority tickets, and short explanations for analysts: what changed, why it looks suspicious, and which company/ needs review.

Tools: , , , ,

How it looks in practice

Data quality · Live alternative feedsAnalyst queue

Flagged for review

RetailCo · orders

P1

Volume down 68% vs 7-day baseline

What changed: UK receipt panel stale. Last row: Tue 02:14 UTC.

Suggested action: Check scraper + confirm before publish

TravelCo · bookings

P2

KPI moved +22% with no sector corroboration

What changed: Single vendor spike. Possible false signal.

Suggested action: Compare against flight + app panels

Delivered to analyst

Linear ticket · RTL feed break
Slack alert · #data-quality

Plain-language summary

"RetailCo orders look wrong because the receipt feed stopped updating. Do not publish today's estimate until the feed is restored or manually validated."

Problems surface at ingestion, not after a client sees a bad number.

02

AI research copilot for investor reports

ODP publishes daily revenue and estimates for 250+ plus deep . Analysts spend hours turning moves into client-ready commentary for and .

Solution: AI reviews daily changes across companies and sectors, then drafts short investment-style comments, chart summaries, client-ready , and Q&A notes for calls. The analyst edits and approves before anything goes out.

Tools: , , ,

How it looks in practice

Research copilot · Daily KPI briefDraft for analyst

1 · KPI moves scanned

CompanyKPIMoveVs
RetailCoOrders-4.2%consensus
BrandCoRevenue-2.8%prior week
AppCoMAU+6.1%sector avg

US Retail sector

Broad softness in mid-tier apparel

2 · Draft client outputs

Investment comment

Receipt panel softness in mid-tier apparel. Watch RetailCo and BrandCo into earnings. Talking point: channel shift vs macro headwind.

Also drafted

  • · Chart summary for sector deck
  • · 3 talking points for hedge fund call
  • · Q&A: "Why did AppCo MAU diverge?"
Analyst edits before sendHuman approval

Effect: Data problems surface before they hit . Analysts start from a draft, not a blank page, and ship client-ready research faster.

Case studies

Real automations we have shipped. Each one follows the same shape: the AI does the heavy lifting, and a person stays in control of the decisions that matter.

Tap or hover any underlined word for a plain-English explanation.

01

AI document and data intake

Messy documents in, clean data out, a person makes the final call.

The situation

A US real estate fund manages more than $500 million. Every week it gets a flood of offers to buy new buildings. The offers arrive as PDFs, spreadsheets, and slide decks, and every broker formats them differently. Only about 1 in 100 offers turns into a deal, but someone still has to read all of them. That reading used to be done by hand, by expensive analysts.

How it works

01

Pull in the offer

The system watches the inbox and pulls in every attachment: PDFs, Excel models, and slide decks.

How it looks in practice

Inbox · New offersAuto-watch

1240 Market St offer

Broker · Kessler

New

Riverside portfolio · 4 assets

Seller · Nova RE

Parsing

Oak Business Park

Broker · Hale

Queued

Automation

  • · Watches broker inbox 24/7
  • · Pulls PDF, Excel, and deck attachments
  • · Queues each offer for extraction
  • · No analyst copies files by hand
02

Read it and fill the model

AI reads the documents, pulls out the key numbers, and fills the fund's own model. Then it works out , , risk, and fit.

How it looks in practice

Investment modelAI filled

Extracted fields

Address1240 Market St, SF
Asking price$28.4M
NOI$1.9M
Cap rate6.7%
Occupancy92%
Source docsOM.pdf + model.xlsx

Projected ROI

11.2%

Portfolio fit

7.5 / 10

Fund model populated automatically

AI reads the broker pack, maps numbers into the fund's own template, and runs yield, ROI, and risk calculations.

03

Summarize for a person

It writes a short summary with the highlights, the pros and cons, and the risks. A manager reads it and makes the call.

How it looks in practice

Deal summary · Manager view1 page

AI summary

Strong location with steady tenant mix
Cap rate above fund target range
!Two leases expire within 12 months
!Broker ask is 4% above recent comps

Recommended next step

Worth a deeper look. Request rent roll and confirm lease rollover before scheduling a site visit.

Manager decides
Look closerPass

Human makes the investment call. AI only prepares the first review.

What it delivers (illustrative)

  • A first review drops from about 30 to 60 minutes down to about 3 to 5 minutes.
  • The team can handle 5 to 10 times more offers without hiring more people.
  • Senior staff stop copying numbers between PDF and Excel, so each review costs less.
  • The fund replies to brokers faster, which is a real edge in a fast market.
02

A sales agent that feels human

Talks to leads by email, consults them, and sends the proposal for signature.

The situation

A company sells beverage services for events and parties. Sales reps used to spend their days answering the same email questions and writing proposals by hand. We built an AI sales agent that handles the whole conversation. People do not realize they are talking to an AI.

How it works

01

Read and reply to leads

The agent reads each incoming email, works out what the client wants, and replies in a natural, human tone.

How it looks in practice

Inbox · SalesAI agent

events@vinerie.co

“Need a drinks package for a 200-person launch in June. Budget around $2k, mostly mocktails.”

Event cateringHot leadBudget fit

Agent action

Replies in a natural tone within minutes. Asks 2 clarifying questions about venue and timing. Client does not know it is AI.

02

Consult and bring in help

It answers questions and recommends the right packages. When a request needs a person, it quietly loops in a teammate.

How it looks in practice

ConversationConsulting

Client: what can we do under $2k, no alcohol?

Agent: recommends 3 mocktail bar packages with staffing, glassware, and timing for a 200-guest launch.

Client: can you include a bartender and setup by 5pm?

Packages compared

  • · Classic mocktail bar · $1,450
  • · Premium botanical bar · $1,780
  • · Self-serve station · $980
Loops in a human rep only for custom pricing
03

Send the proposal to sign

When the deal looks good, it builds a proposal in and sends the commercial offer for on its own.

How it looks in practice

PandaDoc · ProposalSent for signature

Summer Launch Beverages

Mocktail bar · 200 guests$1,450
Staff · 4 hrs · setup 5pm$400
Total$1,850
Commercial offer sentAwaiting sign

Agent builds and sends the proposal without a sales rep touching the document.

What it delivers (illustrative)

  • Leads get a helpful reply in minutes, any time of day.
  • The agent writes and sends proposals without a rep touching them.
  • People step in only for the tricky conversations.
  • The team handles far more leads without growing .