Local AI and automation

A letter arrives — the rest runs on its own

We don't build chatbots — we build systems that process documents, prepare decisions and carry workflows end to end, locally, inside the company's walls. And we only use AI where it's genuinely the best tool.

Project 01

Processing incoming requests

E-mail, attachments, handwriting — classified by intent, with the data needed to act and reply extracted from the e-mail and from internal systems. Several models work together, and AI only where it speeds up and supports the process.

Incoming letter№ 2026-0614/1
Fromkiss.marta@pelda.hu SubjectClaim report — water damage in the bathroom

Dear Insurer, last night our upstairs neighbour's burst pipe flooded our bathroom ceiling. The estimated damage is 350,000 Ft. My policy number: HB-2024-1182. I'm attaching the completed claim form and the photos.

Attachments · 2 items

Photo × 4

bathroom ceiling

Hand-filled claim form

date of damage: June 9, evening · cause: upstairs neighbour, burst pipe

✓ handwriting recognized — data extracted

The process

1.Finding and interpreting case historysystem

2.Processing the photos and the claim form — OCR ✓AI · vision

3.Classification based on the e-mail text and attachmentsAI · encoder

4.Extracting intent and data: property damage claimAI · LLM

5.Customer identification by policy numbersystem

6.Executing the required actions based on internal rules and training materialsAI

classificationclaims · property
policy numberHB-2024-1182
estimated damage350,000 Ft
date of damage · from handwritingJune 9, evening
forwarded to the claims handler
Incoming letter№ 2026-0614/2
Fromnagy.eszter@pelda.hu Subject(no subject)

Dear Sir or Madam, please find attached. Kind regards, Eszter Nagy

Attachments · 1 item

Photo × 1 — handwritten letter

…after the storm the roof tiles slid off and the ceiling is leaking… please help as soon as you can · June 10. 」

✓ handwriting recognized — content extracted

The process

1.Finding and interpreting case historysystem

2.Processing the photo of the handwritten letter — OCR ✓AI · vision

3.Classification: the e-mail body is uninformative — the attachment is the sourceAI · encoder

4.Extracting intent and data: damage claim (storm, roof)AI · LLM

5.Customer identification by name and address — policy number missingsystem

6.Preparing a follow-up request based on the rulesAI

classificationclaims · property
eventstorm · roof damage · leak
policy numbermissing → follow-up
customeridentified by name + address
forwarded to a case handler with a draft follow-up request
Incoming letter№ 2026-0614/3
Fromtoth.gabor@pelda.hu SubjectBaggage claim — our suitcase was two days late

Hello! On our flight to Paris our baggage was two days late; I'm attaching three receipts and a cost list for the necessary purchases, 84,600 Ft in total. Policy number: UT-2026-0457.

Attachments · 4 items

Receipts × 3

photographed by phone

Handwritten cost list

towel 9,900 · pyjamas 12,400 … total: 84,600 Ft

✓ handwriting recognized — items extracted

The process

1.Finding and interpreting case historysystem

2.Processing the receipts and the handwritten list — OCR ✓AI · vision

3.Classification based on the e-mail text and attachmentsAI · encoder

4.Extracting intent and itemsAI · LLM

5.Customer identification by policy numbersystem

6.Reconciling totals and forwarding based on the rulesAI

classificationtravel insurance · baggage
policy numberUT-2026-0457
receipts · reconciled3 receipts · 84,600 Ft
eventbaggage delay · 2 days
forwarded to the claims handler

Project 01

Sorting and processing incoming requests

No custom process definitions or months of process mapping needed: we start from the training materials, policies and reply templates you already have. Requests are processed in multiple steps, measured at decision points in an auditable way — speeding up human work.

Project 02

Internal decision-support system for case handlers

It supports case handlers in their daily work using materials the company already has — with a language model running on the company's own server, without internet access. It searches contracts and policy terms; answers come with source references, and each domain's experts can improve the knowledge base themselves.

Contact

Let's talk.

If you'd like to talk, just reach out.

hello@how.hu →