Case Study / Insurance · Applied AILive

The AI automation layer for insurance.

Streaver took Apeironix from a proof of concept that timed out on a single document to a multi-tenant, self-service platform that cuts agency busywork by more than 80% and now processes live carrier data in production.

Client
Triforta
Product
Apeironix
Industry
Insurance technology
Location
Reno, Nevada · US
Engagement
Q1 2025 → Ongoing
Scope
Next.js · TypeScript · AWS · GPT 5.5 · Stagehand
80%+Time saved for agencies on policy data extraction & AMS hydration
1 → +500From single-file processing to concurrent, multi-tenant throughput
Nov 2025Self-service portal launched to market after investor expansion
5.0/5Verified Clutch rating: quality, schedule, cost, willingness to refer
01The Client

Triforta turned its own AI tool into a product.

Headquartered in
Reno, Nevada · US
Founder & CEO
Rodney Mattos
Agency
Triforta
Stage at engagement
Proof of concept → product

Apeironix is an insurance-automation company that turns the unstructured chaos of the insurance back office into clean, actionable data. It extracts information from disparate documents, PDFs, and carrier emails, uses it to hydrate agency management systems and CRMs, and layers automation on top to eliminate the mundane work that drains an agent's day.

The company was born inside Triforta, a tech-forward insurance agency where founder Rodney Mattos had been training his own language models to understand insurance terminology and solve his own team's problems. Triforta became the first proving ground, and the demand from peers who saw it working became the reason to turn an internal tool into a product.

02The Challenge

A POC that timed out on a single document.

Rodney came to Streaver through a referral from a longtime collaborator after a US consulting firm vouched for the team. The first ask was sharp and unglamorous: the existing proof of concept didn't scale.

01

Synchronous, blocking API

One file at a time. The moment two documents (or two users) hit the system at once, it timed out.

02

No retries, no progress

Zero job tracking. No way to give a user feedback while extraction was running.

03

Trapped inside Salesforce

Meant to live inside Salesforce and serve a whole agency, but it couldn't survive concurrent requests.

04

Engineer-in-the-loop onboarding

Every new agency required Apeironix engineers. The tool was still a service.

The AI document extractor ran on a synchronous, blocking API. It could process exactly one file at a time. The moment two documents (or two users) hit the system at once, it timed out. There were no retries, no job tracking, and no way to give a user feedback on progress. For a tool meant to live inside Salesforce and serve a whole agency, that was a hard ceiling.

The deeper challenge was strategic. Apeironix needed to become a product other agencies could buy and run themselves, without Apeironix's engineers in the loop for every onboarding.

03The Solution

First reliable. Then a platform.

Streaver embedded a dedicated pod that worked in synchronized cycles with Apeironix's leadership and delivered in two deliberate phases: first making the engine reliable, then turning it into a platform.

Phase 01 · The async backbone

From timeouts to throughput

Streaver designed an AWS-native serverless middleware to sit between Salesforce (where users trigger extraction) and the AI extractor. It introduced asynchronous processing, concurrent handling of many documents from many users, automatic retries, queue-based job scheduling, and clear progress feedback, all under a multi-tenant model with API-key authentication, file encryption, and data-lifecycle policies built for insurance compliance.

Phase 02 · The platform

From engine to product

As the system matured and outside investors came on board, scope accelerated. Streaver architected the full integration backbone linking Salesforce CRM, Apeironix's own small and large language models, and AWS infrastructure into a single coordinated engine, then built the self-service portal: onboarding, role-based access, subscriptions, and a modular tool catalog any agency can adopt on day one.

04Under the Hood

Async, multi-tenant, carrier-aware.

A serverless, multi-tenant architecture that coordinates job dispatch, document intelligence, carrier automation, and human review: the engine behind every tool in the catalog.

Triforta platform architecture diagramTwo chained pipelines behind one Next.js application: document extraction (Textract OCR plus a direct raw-PDF AI pass, reconciled by a comparison gate) and carrier quoting (SQS FIFO into ECS Fargate browser workers driving carrier portals). The extraction pipeline shapes each document into the carrier's own form schema, which the Portal then hands to the quoting pipeline. Both report back by signed webhook for human review.USERAPPLICATIONPIPELINE 01 · DOCUMENT EXTRACTIONPIPELINE 02 · CARRIER QUOTINGPERSISTENCEDOWNSTREAMAGENCY TEAMBroker · CSRUploads · verifiesNEXT.JS 15 · ECS FARGATEPortalReact 19 · tRPC · DrizzleCognito SSO · StripeDISPATCHPresign → direct S3Normalise · merge PDFsPOST job + webhook URLHAND-OFFRPA schema → extractorExtraction → quote jobREVIEWField-level verifyBounding-box overlayTENANCYCompanies · seats · plansUsage limits · overagesSERVICE 01 · ASYNCExtractorHono · API GatewaySQS-chained LambdasCached OCR · idempotentPOST /extractAWS TEXTRACTOCR passText · tables · geometryAI · DIRECT PASSRaw PDFGPT-5.2 · Claude OpusRECONCILIATIONComparison gateTwo passes · field diffAGREE → acceptDIFFER → flagConfidence + bboxSERVICE 02 · ASYNCRPASQS FIFO per carrierDynamoDB semaphoreEventBridge finaliserPOST /jobs · /retryECS FARGATEBrowser workerPlaywright + StagehandSES INBOUND · TOTPMFA relayPer carrier · 5 min TTLTARGETSCarrier portalsQUOTED → premiumDECLINED → reasonQuote PDF · capturesSHARED STATEPostgres · S3 · DynamoDBCompanies · policies · quotes · billing · documents & results · job state · semaphoresINTEGRATIONSAMS · ZapierSalesforce · HMACWEBHOOK · signed callback · results read from S3PRESIGNED PUTbrowser → S3Every queue has a dead-letter queue · structured logs to CloudWatch + Sentry on every hop
Figure 1 · System architecture
05What We Shipped

A catalog agencies enable themselves.

A modular catalog of AI tools that agencies enable through a self-service portal, with onboarding that recommends the right modules, role-based administration, and subscription management built in.

Self-Service Portal

API keys and data mapping let agencies onboard with zero dev work.

Policy Data Extractor

Pulls policy data from PDFs documents, ending manual rekeying.

Commissions Extractor

Turns carrier commission statements into clean, reconciled revenue data.

Quoting

Retrieves bind-ready premiums across carriers via AI parsing and APIs.

Proposal Builder

Auto-generates white-labeled proposals in seconds.

Policy Comparison

Side-by-side AI comparison that surfaces coverage gaps instantly.

06Extractor Tool

Turning insurance policies into structured, auditable data.

Insurance policies arrive as PDFs — dense tables, forms, and fine print that someone has to read and retype into a system. We built a serverless extraction service that does it automatically and shows its work: every field it returns is traceable to the exact spot on the page it came from.

How it works

  1. 01

    Upload

    A policy PDF lands in storage and the platform starts an extraction job. The API responds immediately and the rest runs asynchronously, so long documents never hold up the user's screen.

  2. 02

    Read the document

    AWS Textract runs OCR, capturing text, form fields, and tables. Tables are rebuilt in place so the model reads them as tables rather than as scrambled lines of text.

  3. 03

    Extract the fields

    A language model fills in a versioned schema defined per line of business. For every value it returns, it must also return the verbatim text and page number it took that value from.

  4. 04

    Anchor it to the page

    Each citation is matched back against the OCR layer to recover the field's exact coordinates on the page. That's what lets a reviewer click any value and see it highlighted in the original PDF.

  5. 05

    Deliver

    Results are stored and pushed back to the platform by webhook, together with coverage metrics, confidence signals, and the AI and OCR cost of the job.

Two readings, one answer

For high-stakes documents, the service reads each policy twice in parallel — once from the OCR text, once from the PDF itself — and compares the results. Fields where both readings agree pass through. Where they disagree, a second AI pass weighs the evidence and explains which value it trusts and why. Reviewers get a short, ranked list of what actually needs a human instead of a wall of data with no signal.

What makes it hold up in production

Auditable by design

No value ships without its source text, page, coordinates, and confidence.

Model-agnostic

Routing sits behind a gateway, so switching between OpenAI, Anthropic, and Google models is a request parameter — with cost tracked per model.

Schemas are data, not code

Adding a field or a new line of business is a new schema version, not a deploy. Every extraction records the exact version that produced it.

Built to fail gracefully

Jobs are idempotent and safe to retry, OCR results are cached so reprocessing with a different model doesn't pay for OCR twice, and every stage has its own dead-letter queue.

Document extraction service architectureA serverless extraction pipeline: a validating API, a queued AWS Textract OCR stage, and a processor that runs one authoritative model pass over the OCR text plus an optional cross-check pass over the raw PDF, reconciles them deterministically, and anchors every value to its position on the page. Job state lives in object storage rather than a database.CALLERINGESTOCREXTRACTIONRESOLUTIONSTATEDELIVERYCALLERPortalSubmits a jobReferences a fileEXTRACTION MODEStandardCross-checkedDirect readSCHEMA OWNERAuthors + versionsServes them at runCALLBACKJob + statusThen reads resultAPI GATEWAY · HONOSubmitAuthenticatedValidate the requestSafe to retryRecord the requestACCEPTED · QUEUEDNever blocksSQS → TEXTRACT → SNSOCR acquisitionFair-share limitsAsync analysisTABLES + FORMSTables kept intactCACHED RESULTSRe-runs are freeQUEUE · LONG-RUNNINGProcessorLoads the schemaPass 1 · OCR textGPT-5.5 · citationsCited to the sourceTHREE MODESText · both · directPASS 2 · SHADOW CHECKOver the raw PDFSame schema, no OCRCross-check mode onlyDETERMINISTIC DIFFReconciliationNormalise · compare each fieldDISAGREEMENT→ flagged, never mergedOPTIONAL · LLM JUDGEPicks a side and says whyBOUNDING BOXESCitations → page coordsMatched back to the pageExact field positionS3 · JOBS BUCKETJob state, no databaseRequest · OCR · result · errors · delivery statusStatus follows from what has been writtenSQS · WEBHOOK SENDERDeliveryRetries with backoff until the caller acknowledgesPermanent failures stop early · dead-letter queue with alertingSCHEMA FETCHWEBHOOK · job + status — the caller then reads the resultPer-tenant fair queueing · dead-letter queue and alarm on every hop
Figure 2 · Document extraction service
07Quoting Tool

One API call instead of a fifteen-screen portal wizard.

Commercial insurance carriers rarely expose quoting APIs. The one channel that always exists is their agent portal — a 10-to-15-screen wizard that a human fills out by hand, one carrier at a time. This service turns that manual process into a single API call.

How it works

  1. 01

    A quote request comes in

    One API call carries the business details — company, vehicles, drivers, coverages — along with the carrier and line of business. The payload is validated against a carrier-specific schema before anything else happens, so bad data fails in milliseconds instead of halfway through a portal session.

  2. 02

    The job is queued fairly

    Requests land in a FIFO queue partitioned by client and carrier. No single high-volume account can starve the others, and no carrier gets hit with more concurrent sessions than it tolerates.

  3. 03

    A dedicated browser spins up

    Each job runs in its own ephemeral cloud container with a real browser. It logs in, clears MFA, and walks the carrier's wizard screen by screen — filling forms, adding vehicles and drivers, selecting coverages — exactly as an agent would.

  4. 04

    AI covers the gaps

    The primary path is a deterministic script: fast, cheap, predictable. But portals change without notice. When a script hits a screen it doesn't recognize — a new required field, a question added overnight — an AI layer reads the page, answers it from the business data, records what it assumed, and hands control back. The run adapts instead of failing.

  5. 05

    The result comes back

    Bound premium, decline, or referral to underwriting — the outcome is written to durable storage and delivered to the caller via webhook. Every run keeps its own audit trail: screenshots, page snapshots, structured logs, and per-run AI cost.

  6. 06

    Retries resume, they don't restart

    If a run breaks mid-wizard, the retry reopens the same submission in the carrier's portal, reconciles what was already entered, and continues from where it stopped. No orphaned quotes cluttering the carrier's system, and no repeated work.

Built to add carriers fast

The hard part of this kind of system isn't the first carrier — it's the tenth. Each new integration is built by an AI-assisted pipeline: one stage explores the live portal page by page and writes the implementation plan from what's actually on screen, the next generates the code, and the last runs the integration against a real browser, diagnoses its own failures, and fixes until it passes. Production issues are reproduced locally from a single quote ID.

The result

13 integrations across 8 carriers and 2 lines of business, maintained by a small team.

Carrier quoting automation architectureAn event-driven browser-automation service: FIFO queueing per client and carrier, layered concurrency limits, ECS Fargate workers driving carrier portals with Playwright and a model fallback, a relayed one-time-code channel for carrier logins, and a finaliser that releases the slot and delivers the outcome by retrying webhook.CALLERINGESTSCHEDULINGEXECUTIONTARGETSSTATECOMPLETIONCALLERPortalSubmits a quoteRetries a quoteSENDSLine + carrierBusiness detailsCallback detailsRESUMEReopens the samequote at the carrierCALLBACKQuote · declineSteps · assumptionsAPI GATEWAY · HONOSubmitCarrier is supportedValidate the payloadConfirm credentialsRecord the requestACCEPTED · QUEUEDFIFO per client+carrierSQS FIFO CONSUMERDispatcherLayered concurrencyGlobal capPer-client capOne per carrier loginNO SLOT → REQUEUEWaits for a slotECS FARGATEBrowser workerPlaywright scriptScripted, then adaptiveAI fills blockers andrecords assumptionsTIME-BOXED RUNSManaged credentialsSES INBOUND · OTPMFA relayDedicated mailboxParsed · short-livedHanded to the workerTARGETS · NOT APISCarrier portalsQUOTED · SUBMITTED→ premium, quote number, PDFDECLINED · PENDING ACTION→ the reason is surfacedS3 · PER RUNEvidenceQuote PDF · confirmation shotFailure capture · recordingDYNAMODB · ONE TABLEJobs · semaphores · dedupJob records · concurrency slots · de-duplicationThe finalizer is the only thing that releases a slotEVENTBRIDGE · TASK STOPPEDFinalise and deliverUnfinished runs are treated as failures · retries with backoffFailures raise a de-duplicated alert for the teamWEBHOOK · quote · steps · assumptionsFIFO order per client and carrier · one carrier login is never driven twice
Figure 3 · Carrier quoting automation
08The Results

Prototype to market-ready, on time, on budget.

Streaver carried Apeironix from prototype to a market-ready product, on time and on budget, while absorbing waves of new ideas that would have stalled most teams. Work that could have taken an average shop 6 to 12 months of change requests was turned around fast enough that Apeironix chose to move its launch later simply to fit more value in before going to market.

Today the platform is live in production, processing real commission statements across national carriers, with agencies cutting back-office time by more than 80%. The relationship is ongoing and expanding into new modules and a fully embedded, portal-less integration path.

  1. DEC 2024Referral & discoveryIntroduced through a trusted referral; the async-middleware problem is scoped.
  2. Q1 2025Kickoff & the backboneAWS serverless middleware replaces the single-file POC with real throughput.
  3. MID 2025Prototype → MVP → betaPilot agencies get early access; ROI and time savings are proven.
  4. NOV 2025Self-service launchThe portal goes to market, deliberately timed to ship more value.
  5. 2026Live & expandingIn production across national carriers; new modules in active development.
I wish I could give Streaver a six.
Rodney Mattos
Rodney MattosFounder & CEO, Apeironix, Inc.

"Streaver has handled the extra work and additional requests amazingly." Verified Clutch review, 5.0 across quality, schedule, cost, and willingness to refer.

09Why It Worked

A vendor that learned the business.

Apeironix gained a partner that thought like an owner. Streaver pushed back when it saw a better path, kept communication right-sized with weekly working syncs, and took a seat in Apeironix's own producer-training program to learn the language of insurance and shorten every future build.

Tala
Tala
Team Leader

Led development of an insurance automation platform on AWS: Next.js portal, serverless RPA engine with AI browser automation, and OCR-based document extraction.

Anush
Anush
Senior Product Designer

Managed the product design for Triforta from concept to handoff. Acted as the link with the client to align expectations and translate complex requirements into an intuitive UI.

JP
JP
Senior Full-Stack

Co-architected the serverless RPA engine and its AI browser-automation flows alongside the team lead. Contributed to the new OCR-based document extraction pipeline.

Santi
Santi
Senior Full-Stack

Built a full-stack insurance broker portal from the ground up, delivering AI-powered automation tools, document workflows, and a complete billing system into production.

Axel
Axel
Junior Full-Stack

Contributed to building and maintaining an RPA that automates insurance agent workflows on real carrier and brokerage websites.

An idea outgrowing its prototype?

We turn POCs into platforms without burning the foundation.

From a single-file proof of concept to a multi-tenant, self-service AI platform in production. If your AI demo works but won't scale, you're where Apeironix was a year ago.

Start a conversationSee more case studies

10Continue reading

Related reading

Back to all case studies