Text classification
Route, tag and prioritize documents, messages and alerts, including moderation and risk categories.
Natural language processingSoftware Sushi field brief
Text in. Decisions out.
We build NLP systems that read reports, tickets, documents and feeds, pull out what matters, and route it to the right person, with evidence attached.
- Classification, entity extraction, search and summarization
- Classic ML and LLMs, chosen for accuracy, cost and privacy
- Human review built into the workflow, not an afterthought
Recognized by the platforms we build on
Select Tier Services Partner
Certified team AWS, Claude and moreThe text problem
Your best signals are buried in text nobody has time to read
Too much to read
Analysts spend their day triaging reports, tickets and alerts instead of acting on the few that matter.
Knowledge stuck in documents
Contracts, research and case notes hold answers, but keyword search cannot find them.
Rules that keep breaking
Hand-written keyword rules miss new phrasing and flood the team with false positives.
What you walk away with
Your own text, measured before you build
NLP projects live or die on the labels and the evaluation, so that is where we start.
- A labeled sample and a baseline
- Built with your analysts, so every later model has something honest to beat.
- Model comparison on your data
- Classic ML and LLM approaches side by side: precision, recall, latency and cost.
- An integration plan
- Where predictions land, who reviews them, and how the system learns from corrections.
What we build
NLP that fits the decision, not the hype
Sometimes a fine-tuned classifier beats a large model on cost and accuracy. We test both and show you the numbers.
Entity and relation extraction
Pull names, organizations, places, dates and events into structured records your systems can use.
Semantic search
Find passages by meaning across large document collections, with sources you can open.
Summarization
Condense long reports and threads into briefs that keep the evidence linked.
Risk and signal scoring
Combine text signals into scores and alerts, with explanations an analyst can check.
Human-in-the-loop review
Review queues and feedback loops so analysts correct the system and it improves.
Proof
NLP in production for risk intelligence
Named with the client’s permission and backed by their own review.
Software Sushi, Inc has delivered AI products and a working UI for our platform. Our user experience has increased at least tenfold, which we've measured through customer feedback and engagement.
Client engagement
ThreatMinder
We support a risk-intelligence product that turns research, signals and evidence into structured workflows for security, risk and compliance teams: AI-assisted research and risk workflows, plus the web app, API and orchestration around them.
See our clientsTechnical write-up
Building a content moderation classifier
How we approach text classification with scikit-learn and LightGBM: features, evaluation and the trade-offs against larger models.
Read the articleHow we work
From raw text to a reviewed signal
Define the signal
Agree on the categories, entities or answers that drive a decision, and how to judge them.
Label and baseline
Build a small labeled set and a simple baseline so every later model has something to beat.
Model and evaluate
Try classic ML and LLM approaches, compare accuracy, latency and cost, and pick the winner.
Integrate and monitor
Ship into your workflow with review queues, drift monitoring and a feedback loop.
Your first step
Measure it on your own text first
We take a labeled sample of your documents, build a baseline and a stronger model, and show you precision and recall side by side before you commit to a build.
- A baseline and a candidate model on your data
- Accuracy numbers you can check, per category
- A plan for review queues, integration and cost
A good fit if
- Your team reads or triages large volumes of text
- You can share a representative sample, even a small one
- Mistakes matter, so you want evidence and review
Probably not if
- There is no text data to start from yet
- You only need a one-off keyword search
What text do you need to make sense of?
The secure contact form is temporarily unavailable. You can still reach our team directly.
Email our team- 1We reply within 24 hours.
- 2An intro call of about 30 minutes on your use case, data and constraints.
- 3You leave with a possible way forward, scoped to a measurable outcome.
Rather talk first? Book a call
FAQ
Questions, answered.
Do you use LLMs or traditional NLP?
Both. Fine-tuned classifiers and extraction models are often cheaper and more predictable for well-defined tasks. LLMs shine for open-ended reading and summarization. We compare them on your data.
How much labeled data do we need?
Less than most teams expect to get started. We often begin with a small labeled sample, measure, and label more only where the model is weak.
Can you handle sensitive or regulated text?
Yes. We design access controls and data handling for regulated work and can deploy in infrastructure you own or control.
How do you measure accuracy?
With a held-out test set and per-category precision and recall, agreed with you before we build, and tracked after launch.
Can it plug into our existing tools?
Yes. We deliver NLP as APIs and workflow steps that connect to your apps, ticketing and data platforms.
Have a pile of text that should be a signal?
Tell us what you read today and what decision it feeds. We will show you what NLP can take off your team’s plate.