AI Platform · Insurance

An AI Insurance Claims Automation Platform

A full-stack platform that automates First Notice of Loss — the moment someone reports a car accident. The claimant calls, chats, or taps a link, and an AI agent verifies the policy, gathers the details, collects GPS-verified photos, assesses the damage, and routes them to a repair or towing provider.

Built by
HashCode – FZE
Type
Full-stack AI customer-engagement & workflow-automation platform, built end to end
Industry
Insurance (motor / auto claims)

3

Intake channels

AR · EN · FR

Languages, auto-detected

100%

Stored PII encrypted

End to end

Coverage

Overview

When someone crashes their car, the first thing they do is report it — and that First Notice of Loss is exactly where insurers drown. We built a platform that automates it end to end. A claimant reaches it by phone call, chat, or a web link and talks to an AI agent. The agent checks their policy, walks them through what happened, collects location-verified damage photos, assesses the damage, points them to an approved repair or towing provider, and hands back a claim reference. What used to be a call-centre queue becomes a structured pipeline.

It's multi-channel and multilingual by design. Voice, chat, and web all feed the same claim and conversation-state model, so a claim behaves the same no matter how it started. The agent handles Arabic, English, and French — including regional dialect and Latin-typed Arabic — because that's how people in the region actually write and talk.

We own the whole stack — the conversational and voice AI, the backend and domain logic, the encrypted data layer, the operations dashboard, the geospatial routing, and the containerized deployment — built as one coherent codebase rather than assembled from off-the-shelf tools.

The Challenge

FNOL is high-volume, time-sensitive, and unforgiving. Accidents happen at any hour, and the person reporting one is usually stressed, on the roadside, using whatever channel is closest to hand. Done by hand, intake is slow and inconsistent — details get missed, evidence arrives unverified or not at all, dispatch is manual, and fraud is hard to catch at first contact. Language makes it worse: customers routinely switch between Arabic, English, and French mid-sentence, which most automated systems handle badly. The real engineering problem was making an LLM conversation trustworthy enough to run a regulated claims workflow — where steps can't be skipped, reordered, or hallucinated — while still feeling human to a shaken customer. Our answer was to:

  • Separate understanding from control, so the AI interprets people freely but never decides the workflow on its own.
  • Verify evidence at the point of capture — GPS-gated photos — instead of trusting whatever arrives.
  • Encrypt sensitive claimant data end to end, with every step auditable.
  • Handle voice, chat, and web through one shared claim model, in three languages.
  • Keep every AI provider swappable, so the platform is never locked to a single vendor.

The Solution

We built it in clean layers, keeping the business logic away from any one framework or vendor so nothing was painted into a corner. Here's how the pieces fit together.

Multi-channel conversational intake

One agent design serves real-time voice calls, chat, and web sessions, all writing into a single shared claim and conversation-state model. Verification, triage, routing, and feedback behave the same whichever way the claim started, and every conversation is isolated so concurrent claims never bleed into each other.

A conversation engine that can't go off the rails

Instead of letting the language model decide what happens next, we built a finite-state engine that owns every step as a function of the facts collected so far. The model gets two narrow jobs — understand (pull structured facts out of each message) and phrase (ask the next question naturally). Deterministic fallbacks guard the critical branches like injury and driveability, so the workflow stays correct even when the model is unsure. That's what makes an LLM safe to run a claims process.

Real-time voice AI

Inbound and outbound calls run through a real-time WebRTC voice layer on a speech-to-speech model, so the conversation is low-latency and natural, with automatic language detection. Full transcripts are captured against the claim for review and audit.

Evidence verified where it's taken

Damage photos come in through secure, single-use upload links that are GPS-gated: the platform records the coordinates at the moment of capture (and records why if it can't), then reverse-geocodes them into a real address, so an adjuster has proof of where the evidence was taken.

AI damage assessment and fraud signals

A vision-language model reads the uploaded photos and returns a structured assessment — affected areas, severity, repair-cost estimate — plus fraud-risk flags raised for human review past a confidence threshold. The model behind it is switchable, so we're not locked to one vendor.

Provider routing that defends itself

Approved experts, garages, and towing providers live in the platform's own directories with coordinates. Once the incident location is known, providers are ranked by real distance from it — turning “who do we send?” into an automatic, defensible decision.

A structured claim record, exportable

Every claim builds into a full structured accident report — insured party, third party, injuries, witnesses, experts, incident details — assembled from the conversation. Personal data is encrypted at rest with searchable hashing for lookups, and finished reports export to PDF and DOCX with the evidence photos embedded.

An operations dashboard for staff

A role-based internal dashboard gives the team a live view of claims, conversation sessions, voice transcripts, an evidence gallery, WhatsApp number management, and a geolocation map — all updating in real time.

Technology

Technology stack used to build the platform
LayerTechnology
FrontendNext.js 14 (App Router), React, TypeScript, Tailwind CSS, Zustand
Backend / APIPython, FastAPI (async), clean/hexagonal architecture, domain-driven design
Conversational AIStructured-output LLM agents (understand + phrase) over a deterministic finite-state engine
VoiceReal-time WebRTC voice platform running a speech-to-speech model
Vision AIVision-language model → structured damage assessment + fraud flags (vendor-switchable)
DataPostgreSQL 16, async SQLAlchemy, field-level PII encryption with searchable hashing
Messaging / asyncCelery on Redis, per-customer sharded FIFO queues
Realtime UIWebSockets with Redis pub/sub
GeospatialIn-app GPS capture, geocoding provider, distance-based routing
SecurityJWT auth, role-based access control, field-level encryption, HMAC webhook verification
InfrastructureDocker Compose, background voice & messaging workers, structured logging/observability

The architecture is layered and provider-agnostic: framework-free domain logic and use cases at the core, concrete adapters for AI, voice, messaging, storage, and geolocation at the edges, and the AI providers behind conversation, voice, and vision all configurable rather than hard-wired.

Results & Impact

3 channels

voice, chat and web, from one conversational core

3 languages

Arabic, English and French with automatic detection

GPS-verified

photo evidence captured at the point of upload

100%

of stored personal data encrypted at rest

Why It Matters

Putting a language model in front of insurance claimants is easy to show off and hard to trust. A claims process is regulated: steps can't be skipped, reordered, or invented, and any figure might be questioned months later. So we kept the model on a short brief — read the customer, phrase the next question — and let plain code own every decision. Evidence gets verified the moment it's captured, and claimant data stays encrypted the whole way through. What came out is something an insurer can actually put into production: real conversational and voice AI, computer vision, and geospatial routing, wrapped in the controls that make them safe to use on regulated work.

  • Applied AI
  • Insurance tech
  • Voice AI
  • Computer vision
  • FastAPI
  • Geospatial
  • PII encryption
  • FNOL automation

Build something like this with HashCode

AI platforms, automation, and custom software — designed, built, and operated end to end. Tell us about your project.