Applied AI / 04
grounded answers
Credit policy assistant
Answers from a lending policy manual with every claim cited, and weak matches declined instead of guessed.
The decisionWhat the credit policy says about a case, and whether the answer can be trusted.
01 / Overview
The problem, and the approach.
AThe challenge
Loan officers lose time searching a long policy manual, and a chatbot that paraphrases policy can be confidently wrong in exactly the places that create compliance risk.
BThe approach
Retrieval over 25 policy sections, tuned with lending synonyms. Every claim in an answer cites its section, a quoted mode cannot drift from the source text, and a refusal threshold routes weak matches to Credit Policy instead of guessing. An evaluation set scores retrieval and refusals before any change ships.
02 / The decision journey
4 decisions, one story.
Each stage asks one business question and answers it with a model you can open and change. Any card opens that stage of the live demo.
-
01
Ask
What does the policy say?
Retrieval with lending synonyms finds the sections that answer it.
Open this stage
-
02
Cite
Can I check the answer?
Every claim carries its section id, and quoted mode cannot drift from the source.
Open this stage
-
03
Decline
What if the manual doesn’t cover it?
Weak matches are declined and routed to Credit Policy instead of guessed.
Open this stage
-
04
Evaluate
How do we know it is right?
A test set scores retrieval and refusals before every change, misses included.
Open this stage
03 / Under the hood
How it’s built.
Algorithms
- TF-IDF retrieval
- Refusal threshold
- Evaluation set
- Claude (optional)
Stack
- Python
- TF-IDF retrieval
- Refusal threshold
- Evaluation set
- Claude (optional)
- Streamlit
Methods, in order
- Retrieval
- Cited answers
- Refusal thresholds
- Evaluation set
- Claude (optional)
Data
A 25-section credit policy manual written for the demo; no real policy text.
04 / Outcomes
What it changed.
- 01
89% of questions find the right policy section first.
- 02
75% of out-of-scope questions declined rather than answered.
- 03
Every answer carries a section id that a reviewer can check in seconds.
Results come from synthetic or public data with known answers. They show the method, not a client result.
05 / Try it
The live demo.
The real app, running on synthetic data. Change the inputs and watch the decision move.
Interactive, synthetic data. If nobody has used it for a while it takes about 30 seconds to wake.
Open in a new tab ↗06 / Related
Keep exploring.
Data & decisions / 10
Lead prioritization
Production lending work rebuilt on synthetic data: who the team should call first, why, and which product fits.
View project
Interactive lab / 01
Lending decision lab
A fictional credit union and 50,000 synthetic members, followed through nine connected lending decisions: who to approve, what to charge, who will leave, and whether it is all fair.
View project
Applied AI / 03
Analytics copilot
Ask the warehouse a question in plain English and get governed SQL, a chart and a one-line answer, or an honest decline.
View project
Open conversations