Salix

Salix listens to the calls you can't.
Built for B2B sales teams. Managers can't sit in on every call.
Every call scored, with the reason in plain words.
Findings quote the transcript, with timestamps.
A coaching plan aimed at each rep's weakest skills.
Reps practise against an AI buyer built from real objections.
The buyer only commits when the rep shows the skills.
Credit packs, paid with PayPal.
A real PayPal sandbox payment through APIMatic's SDK.
Managers see who's ahead and which deals need attention.

Inspiration

A deal can be lost in one sentence. The buyer says "it looks expensive", the rep answers with "we have won several awards", and nobody on the team ever hears it. If you run a B2B sales team, you can't sit in on every call, so the same mistake repeats and you only find out when the quarter ends. Salix listens to the calls a manager can't.

What it does

Scores real calls in English, Urdu or Roman Urdu, detecting the language automatically. It measures talk time, questions, objections and deal risk.
Shows its evidence. Findings quote the transcript with a timestamp, so nobody has to take the AI's word for it.
Writes a coaching plan aimed at each rep's weakest skills.
Tests the rep against an AI buyer built from objections in real calls. The buyer only commits once the rep shows the skills, and the rep gets a verdict.
Gives managers a team view of who is ahead, which deals are at risk, and the plays that win.
Charges credits through PayPal. Each analysis costs one credit, and a failed analysis gives it back.

How we built it

The backend uses FastAPI, SQLAlchemy and SQLite. The frontend uses Next.js and React.

PayPal: credit packs are bought through PayPal Orders using paypal-server-sdk, the SDK APIMatic generates. Salix creates the order in PayPal's sandbox, the buyer approves it on PayPal's own page, and Salix captures it and grants the credits once. The payment in the video is sandbox order 4GM35893VB994222R, captured as 4M2861976M379154E for USD 40.00, status COMPLETED, read back from PayPal's API. No real money moved. Braintree is also supported.

APIMatic: three tools were used. The generated PayPal SDK runs the payments. The Context Plugin reviewed my integration and led to real fixes. The APIMatic CLI generated a Python SDK and docs portal for Salix's own API. The evidence is in docs/APIMATIC.md.

AI: the analysis runs on real AI models, with automatic failover when a provider is overloaded.

New since the submission period opened: the credit paywall (PayPal Orders through the APIMatic SDK, and Braintree), Google or email sign-in with isolated teams, automatic language detection, the deal score and coaching threshold, the closing test roleplay, AI failover, and the generated SDK and docs portal for Salix's own API.

To try it: follow the Quickstart in the README. It needs no Docker, no database server and no API key. To see the purchase flow without any accounts, set SALIX_PAYPAL_SIMULATOR=true. The local simulator is clearly labelled. Add your own sandbox keys for real PayPal sandbox payments.

The call in the video is a sample I scripted, analysed for real. The team data is sample data.

Challenges we ran into

A model provider overloaded mid-analysis. Google's AI was overloaded while I recorded the demo, and one analysis took 2 minutes 20 seconds. I shortened how long Salix waits before switching models. The video shows and discloses this.
An SDK error could look like an empty order. The Context Plugin review showed the SDK also reports errors through is_error(). My code only handled exceptions, so I fixed that and added retries for safe requests.
The first generated SDK for Salix's API crashed. APIMatic scored my API description 100/100, but the SDK still failed when it ran. I fixed the description and tested the SDK against a running server until all 17 live checks passed.

Accomplishments that we're proud of

A real PayPal sandbox payment runs end to end through APIMatic's generated SDK: order, approval on PayPal's own page, capture, and credits granted once.
Salix handles English, Urdu and Roman Urdu with no language menu.
Findings point back to the exact line of the transcript.
A rep doesn't just get told what went wrong. They practise against a buyer until they can close.
A fresh clone runs with no API key, and the backend test suite has more than 1,400 passing tests.

What we learned

A generated SDK isn't proven until it runs against a real server. A high score on the API description didn't stop the first SDK from crashing. I also learned that AI output earns trust when every finding points back to the evidence, and that an AI product has to plan for its provider being overloaded.

What's next for Salix

Moving PayPal payments from the sandbox to live.
Trying Salix with real sales teams on real calls, especially teams selling in Urdu and Roman Urdu.
Connecting it to the tools sales teams already use, so calls and deals flow in automatically.