Bridging the gap between customer interaction and backend data readiness — ending the era of manual data intake in insurance claims processing.
Customers call a contact center, wait for an agent, and answer the same repetitive questions every time — while the agent manually keys everything in.
Manual data entry introduces data quality issues, and every extra minute on the call compounds directly into higher operational cost.
See the SolutionCustomers call in and wait for an available contact center agent before intake even begins.
The same set of intake questions are asked manually, claim after claim.
Manual entry introduces inconsistent formatting and field-level errors.
Longer average handling time per claim adds up directly to cost.
ClaimFlow AI went from scoping to production inside AIG's claims environment in six weeks — integrated directly into their FNOL workflow, not bolted on top of it.
ClaimFlow connected to AIG's backend claims platform via secure API connectors — reading policy records in real time and pushing completed FNOL records straight into the queue.
Deployment followed a gated rollout — each phase unlocked only after quality metrics cleared, keeping AIG's live claims traffic fully protected throughout.
Measured against AIG's pre-deployment baseline across the same FNOL volume, results aligned with CodeSizzler's enterprise benchmarks.
ClaimFlow AI runs a state-machine-driven conversation that collects, validates, and extracts every required field — without a single form.
Agent: "Hello, please provide your 8-digit policy number."
Agent: "What is the date and time of the incident?"
Agent: "Where did the incident occur?"
Agent: "Can you briefly describe the damage?"
All four fields compiled into one structured claim record and submitted.
START → POLICY_NUMBER → INCIDENT_TIME → LOCATION → DAMAGE_DESCRIPTION → SUBMIT. The agent always knows exactly which question to ask next.
Both services are deployed and healthy, running independently on Azure.
FastAPI service running on Azure App Service, handling conversation orchestration and claim extraction.
Streamlit interface running on Azure App Service, deployed via a dedicated GitHub Actions workflow.
Full text-to-speech playback after every agent response, so the conversation feels truly hands-free — speak, listen, respond, repeat.
claimflow.codesizzler.in for the customer-facing UI, and api.claimflow.codesizzler.in for the backend service.
Voice input → Azure Speech-to-Text → ClaimFlow AI Agent → field extraction → live audit panel → Azure Text-to-Speech → voice response → claim submission.
A conversational ordering agent that greets, guides, and checks out your customer — by voice or text — without a single form or human operator.
Customers scroll through static menus, tap confusing UI flows, and never get to say "make it Jain" or "extra gravy, please" in plain language.
Operators lose upsell moments, dietary notes get missed, and every edge case — a nut allergy, a last-minute swap — becomes a support call. The gap between what a customer wants and what the kitchen gets is wide and expensive.
See the SolutionNo form field for "no onion, no garlic, extra gravy" — customizations fall through or get mis-keyed.
Nut allergies, Jain requirements, and vegan needs often go untagged until the food reaches the table.
A static UI never suggests the naan that pairs perfectly with butter chicken — Milo does, once, at the right moment.
Indian restaurants serve multilingual customers. A single-language UI leaves a significant segment underserved.
Milo runs a natural, memory-aware conversation that captures every detail — dining mode, items, spice, customizations, dietary flags — and closes with a real payment.
Milo greets the customer and asks: Dine-in, Takeaway, or Delivery?
Each item: name, spice level, customizations. All tracked across the full conversation.
Jain / allergy tags are set automatically. One complementary suggestion offered — never repeated if declined.
Full order read back grouped by category — starters, mains, drinks — with itemized ₹ totals including GST.
Dedicated checkout page: Cash / Counter / Razorpay online. Demo mode auto-activates when live keys are absent.
GREET → DINING_MODE → ITEM_LOOP → DIETARY_CHECK → UPSELL → CONFIRM → CHECKOUT → THANK_YOU. Milo always knows exactly where in the order it is.
Everything from voice input to live payment — shipped and working.
Core ordering, voice, and payment are all live. A few edge-case features are pending verification or are paused on external dependencies.
All 9 ordering capabilities, dietary tagging, upsell logic, order JSON output, and Razorpay checkout are live and tested end-to-end.
Nut allergy flag and vegan-option routing didn't appear in the last tested order JSON — needs explicit test coverage and a fix if missing.
Hindi and Tamil text responses are spoken through an English-accented voice engine. No true regional STT locale or matching native TTS voice is configured yet.
Phone/email OTP login and order-summary delivery via SMS or email are paused — blocked on a Twilio outage and not yet built.
Configure a dedicated Hindi and Tamil STT locale and matching native TTS voice (e.g. hi-IN-SwaraNeural, ta-IN-PallaviNeural) so the voice experience matches the customer's language from end to end.
Phone/email OTP login at the start of each session, followed by an SMS or email order summary after checkout — currently paused on Twilio outage resolution.
Explicit test suite to confirm allergy flags and vegan routing appear correctly in the output JSON for every edge case, with a fix shipped if missing.
Push confirmed order JSON directly to a kitchen display system or POS via a webhook, removing the final human relay between Milo's output and the kitchen.