OnboardIQ

Constraints

• Highly sensitive data: SSNs, immigration document numbers, signatures.
• I-9 fields change based on citizenship status.
• Output had to be the official PDFs, with signatures and dates in the right boxes.
• Users are often not technical, so the interview runs by voice.
• Every signature needs an audit trail: who signed, when and from where.

After each answer the AI reads it back, letter by letter for names and email, and digit by digit for numbers like the mobile number, date of birth and SSN, and asks the hire to confirm before saving. A wrong read-back means the hire simply says it again. The assistant first collects basic details (name, address, date of birth, SSN, phone, email, emergency contact). It then goes through the DE 4, the I-9 and a short W-4 step, and asks only for the fields that are still missing. It refers to each one only as "an employment form".

Next, a dialog opens to upload photos of the identity documents. Then a review screen shows everything collected. The hire confirms, or says it is wrong and the assistant fixes it. Finally the hire signs, either by drawing with the mouse on a canvas or by typing their name. The signature is stamped on every form with the date, time and IP address. A closing dialog ends the session and opens a short feedback form.

Key features

1. 01

   Voice-only interview in English or Spanish, with speaking and listening animations and an animated transcript.

2. 02

   Read-back confirmation of every answer, spelled letter by letter or digit by digit, saved only after the hire says yes.

3. 03

   Basic details first, then a per-form gap check, so no question is asked twice.

4. 04

   Document upload, a full review screen, and a drawn or typed signature with date, time and IP.

5. 05

   Automatic filling of W-4, DE 4 and I-9 PDFs, stored in private S3 and read only through presigned URLs.

Key design decisions

1. 01

   The hire picks a language and reads the required-documents list. The app creates a session.

2. 02

   Each answer goes to the form service, which looks up the session step and forwards the message to the assistant service with that step.

3. 03

   The assistant service loads the prompt for the step and language, calls the OpenAI Assistants API, and streams text and ElevenLabs audio back as one multipart response.

4. 04

   The UI plays the audio, shows the transcript, then listens again. The AI reads back each answer and waits for a yes.

5. 05

   When a step is complete, a tool call saves the fields (basic details, DE 4, I-9, W-4). The form service validates them, for example a USCIS or A-Number for permanent residents, and moves the step forward.

6. 06

   At the upload step the UI opens the upload dialog. Files go to private S3 through multer.

7. 07

   At the review step the UI shows all fields. A "no" sends the hire back to correct them.

8. 08

   At the signature step the hire draws or types a signature. It is saved with date, time and IP.

9. 09

   The PDF engine fills each form and stamps the signature and audit data at fixed coordinates. Finished forms stay in S3 and are read only through presigned URLs.