What This Project Does
In large organizations, documents arrive every day — research papers, news articles, regulatory updates, internal reports. Someone has to read each one and decide who should care about it. That takes time and it is easy to miss things.
AssortApp automates this. You give it a document — paste text, drop a URL, or upload a PDF. It reads it, decides which team it is relevant to, and writes a summary just for that audience. It also draws a mind map of the key concepts, pulls out important claims, and flags anything that looks like a compliance risk.
How It Works
The app runs every document through seven steps, all connected using LangGraph.
Step 1 — Ingest The user submits content. The app checks size limits and runs safety checks.
Step 2 — Route the Audience This is the core of the project. A small, fast ML model (TF-IDF + Logistic Regression) tries to classify the document first. If it is confident, it routes immediately. If not, it passes to an LLM to make the call. Documents that are unclear go to a Cross-Functional bucket.
The three main audiences are:
- Commercial
- Medical Affairs
- R&D
Step 3 — Generate the Summary The LLM writes a one-line summary, 3–5 decision-ready bullet points, topic tags, and key signals — all tailored to the routed audience.
Step 4 — Evaluate A second LLM checks the output. Are all required sections present? Is the length right? If something is wrong, the pipeline retries with corrective feedback.
Step 5 — Citations and Risk Flags Two tools run back to back — one extracts factual claims and traces them to the source, the other flags compliance issues by severity.
Step 6 — Revision Loop If the output fails evaluation, the system tries again with corrective context, up to a set retry limit.
Step 7 — Save Everything Every attempt, every piece of feedback, and every final output is saved to a database. Nothing is lost. You can always go back and see exactly what happened and why.
What I Am Most Proud Of
Hybrid routing. Using ML first and only calling the LLM when needed keeps the app fast and cheap to run. Most documents are easy to classify — no need to spend LLM tokens on them.
Full audit trail. Every version of every output is stored. In professional environments, being able to explain how a decision was made is just as important as the decision itself.
Tag intelligence. As more documents are processed, the app starts to surface patterns — which topics are trending, which concepts appear together. It gets more useful over time.
Live Numbers
| Metric | Value |
|---|---|
| Documents processed | 89 |
| Success rate | 97% |
| App version | 3.0.0 |
Try It Yourself
The app is live and open to try: