Headline Impact
22,000+
Hours of Manual Labor Eliminated with LLM-Powered Classification
FinTech
Text Analytics
LLM Classification
22,000+
Hours of manual labor eliminated through AI classification
2M+
Employee reviews processed and classified at scale
~20
Topic classes compressed for cost-efficient classification
The Client
Kelp — FinTech in Alternative Investments
Kelp — a FinTech company in alternative investments maintaining a proprietary database of 2M+ employee reviews. A small fraction contained investor-critical data on misconduct or workplace bias. Identifying them manually would have required 22,000+ hours of labor.
The Challenge
Finding the Needle in 2M+ Haystacks
Two problems: (1) Finding the needle in 2M+ haystacks — only a small fraction of employee reviews contained material information for investors, but manual identification was a 22,000+ hour impossibility. (2) Business analysts spent excessive hours extracting investment-sensitive information from news articles, market reports, and financial documents, creating delays in time-critical decisions.
What We Built
LLM-Powered Classification & Extraction Pipeline
1. LLM-Powered Classification
Tested KNN with BERT embeddings, topicGPT, and OpenAI sentence embeddings before implementing GPT-4 with prompt engineering for maximum accuracy.
2. Data Cleaning Pipeline
Automated filtering to eliminate low-probability records before classification — dramatically reducing cost and processing time.
3. Topic Compression
Compressed classifications to ~20 topic classes for cost-efficient processing without sacrificing signal quality.
4. Model Evaluation Framework
Built pipelines enabling rapid iteration and easy switching to newer, cheaper LLMs as they become available.
5. Text Extraction Tools
Combined prompt tuning with post-processing to extract investment-sensitive information from unstructured documents.
Technology
Powered By
GPT-4
BERT Embeddings
Prompt Engineering
Topic Classification
Model Evaluation Pipelines
Automated Data Cleaning
The Results
22,000+ Hours Eliminated — Investor-Critical Data Surfaced at Scale
Eliminated 22,000+ hours of manual classification work. Processed 2M+ employee reviews, surfacing the investor-critical subset with high accuracy while building a reusable framework that can switch to cheaper models as they emerge.
"Tailored AI engages with Kelp on tasks involving building LLM-powered products for text analytics and extraction from unstructured datasets."
— Kelp
Ready to Transform Your Operations?
We've delivered $100M+ in business impact across IT services, healthcare, HR tech, and fintech.
Book a Scoping Call
