Reconstructed Slides: Make your LLM app a Domain Expert: How to Build an Expert System — Christopher Lovejoy, Anterior
Source Video
Make your LLM app a Domain Expert: How to Build an Expert System — Christopher Lovejoy, Anterior
Method
This deck is reconstructed from the existing video frame captures by detecting likely slide regions with OpenCV, cropping/upscaling those regions, deduplicating similar crops, and OCRing the cropped slide images locally. It is a cleaner companion to the full-stage frame deck.
Reconstructed Slides

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.91 - Text source: agent_vision.
Slide text:
Own your vertical: A playbook for building a domain-native LLM application
Dr Christopher Lovejoy, MD
Head of Clinical AI
2025-06-04
Anterior

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.93 - Text source: advanced OCR
rapidocr-live/left-72/contrast. - OCR decision: ready — small body text and logos are better handled by OCR
Slide text:
Anterior
Were a New York-based, clinician-led
AIE reasoning tools andsolutions to accelerate andautomatehealthcare company thatprovides clinical
administration
BACKEDBY NEA SEQUOIA
andFOUNDERSOF
DoosMad GoogleAi
Google ALUMRECE mMeta
Microsoft amazon MeKnecy &Company
IMPERIAL NHS
aws

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
OUR BET FOR VERTICAL AI APPLICATIONS: system for incorporating domain insights > the sophistication of your models and pipelines

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
Why is it hard to successfully apply LLMs to specialized industries?

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.95 - Text source: advanced OCR
rapidocr-live/left-72/contrast. - OCR decision: ready — small case details and question text are better handled by OCR
Slide text:
An example clinical case processed by Florence
AIE Doctor recommends a knee arthroscopy Presents with right knee pain 78-year-old female
Q: Is there documentation of unsu
conservative therapy for at least 6
Microsoft

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
KLAS CASE STUDY
We achieved a baseline performance of 95% for approving care...
95.73%

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
KLAS CASE STUDY
We achieved a baseline performance of 95% for approving care...
then iterated with a specific customer to >99% within 8 weeks
95.73%
99.24%

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.96 - Text source: advanced OCR
rapidocr-live/left-72/contrast. - OCR decision: ready — Dense diagram labels and small node text are better suited for OCR than direct transcription.
Slide text:
AIE maintain a failure mode ontology (2) Empower domain experts to define citraetion Photo ortastion Tae
1ozruoievy location
Handuriting ertraction Medealrecord extract'on
extraction checkbox Medical Necessity Review Failure Modes reasontg
chrorolog'
mtereretotion Rles
Ruhe courct
seeetion
Microsoft

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.95 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — The slide contains a product screenshot with dense small UI text that OCR can extract more reliably than direct vision transcription.
Slide text:
(3) Measure metrics and failure modes together
AIE on production data to unlocktargeted insights
ToooTe
Fax
Microsoft smol 3i

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.94 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — Flowchart-style slide with many small labels and arrows is OCR-suitable.
Slide text:
AIE
MEASURE Review AI outouts AralyTe perfornarce to prioritise work Specify failure mode' to fix and perforrance threshold Make changes
opplcation Producton Perfornance insights ChnicaPx engineer(s) Chnico! enginear(s) AI datoset'eval "Failure mode
about naking Iive Make decisions Subnit inproverents See inpact of changes
aws
S

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: agent_vision.
Slide text:
Failure mode datasets enable targeted product iteration

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — Dense diagram slide with multiple small labels and arrows; OCR is likely better than manual transcription here.
Slide text:
Empower your domain experts to make
AIE improvements directly with tooling and evals Application Pipelines
P45366 deternie vhether to make live Ewlate to
changes
domain cxperts Tooling for of releunt donin Reatns retreunl baocdedge Domain evals
Lonain deternine Vhether Mgoo Evohiate to
to pipebnes
Knowdedae Base uIOwOq
Microsoft smol ai

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
Takeaways
To build domain-native LLM applications you need to solve the last mile problem.
Using the best models isn't enough - you should build an adaptive domain intelligence engine.
Domain experts power this system by reviewing AI outputs to generate performance metrics, failure modes and suggested improvements.
This takes production data and uses it to give your LLM product a nuanced understanding of customer workflows.
The result is a self-improving, data-driven process that can be managed by a domain expert PM.
Hidden Non-Slide Evidence
- `slide-001.jpg` —
sponsor_logoconfidence0.99; sponsor/logo wall, not a presentation slide - `slide-014.jpg` —
title_cardconfidence0.99; Section divider / title-only slide with no substantive content. - `slide-016.jpg` —
title_cardconfidence0.94; Closing thank-you/contact slide.
Classification audit: raw/sources/slide-ai-classification/reconstructed/MRM7oA3JsFs/audit.json