Update README.md
Browse files
README.md
CHANGED
|
@@ -1,10 +1,49 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Signeasy AI
|
| 2 |
+
|
| 3 |
+
We're the AI team at [Signeasy](https://signeasy.com) β building
|
| 4 |
+
the machine-learning systems that turn unstructured contract data
|
| 5 |
+
into something a sales rep, a paralegal, or a finance lead can
|
| 6 |
+
search, summarise, and act on in seconds.
|
| 7 |
+
|
| 8 |
+
## What we work on
|
| 9 |
+
|
| 10 |
+
- **Document AI** β extracts contract types, key terms, and signing
|
| 11 |
+
metadata from PDFs at ingest time, so the rest of the product can
|
| 12 |
+
treat a contract as structured data instead of a binary blob.
|
| 13 |
+
- **Natural-language search** β turns one English query
|
| 14 |
+
(*"top 10 NDAs by value"*, *"expiring contracts over $500K"*) into
|
| 15 |
+
a structured filter spec the contract-management UI dispatches into
|
| 16 |
+
its existing data layer. Resolution-only β no LLM in the hot path.
|
| 17 |
+
- **Contract categorisation** β classifies a freshly-uploaded
|
| 18 |
+
document into one of ~20 first-class contract types so the right
|
| 19 |
+
extraction pipelines fire automatically.
|
| 20 |
+
|
| 21 |
+
## Models hosted here
|
| 22 |
+
|
| 23 |
+
| Model | Purpose |
|
| 24 |
+
|---|---|
|
| 25 |
+
| [`Signeasy/nl-search-joint-bert`](https://huggingface.co/Signeasy/nl-search-joint-bert) |
|
| 26 |
+
DistilBERT-based JointBERT classifier (intent + slot + span heads) powering natural-language
|
| 27 |
+
search over contract data. Private. See the model card for eval results, OOD defences, and version
|
| 28 |
+
history. |
|
| 29 |
+
|
| 30 |
+
## How we work
|
| 31 |
+
|
| 32 |
+
Most of our models are deliberately small β DistilBERT, MiniLM,
|
| 33 |
+
distilled cross-encoders β so they run on CPU at the latency the
|
| 34 |
+
product needs. Search has to feel instant; an end-user doesn't wait.
|
| 35 |
+
|
| 36 |
+
We use Hugging Face Hub as our model registry: training happens in
|
| 37 |
+
our research repo, the artifact gets pushed here as a private repo
|
| 38 |
+
at a tagged revision (`v3.5`, `v3.6`, β¦), and production pulls it
|
| 39 |
+
via `snapshot_download` at image-build time so cold starts are
|
| 40 |
+
self-contained.
|
| 41 |
+
|
| 42 |
+
## Audience
|
| 43 |
+
|
| 44 |
+
This org is primarily for the Signeasy AI team and the engineers
|
| 45 |
+
integrating our models into the product. The repos here are not a
|
| 46 |
+
general-purpose release; they are the internal artifact registry.
|
| 47 |
+
|
| 48 |
+
If you're a Signeasy engineer who needs access to a private repo
|
| 49 |
+
here, ping the AI team on Slack.
|