Selected Work

Work we can show you

We treat client engagements as confidential. Studies here are anonymized and published only after the client approves exactly what is shown. Named references are available on request once an engagement is underway.

Presenting client work

We show the shape of the work, not the client.

Every engagement runs under NDA, so what you see is the problem we were handed, the system we built, and how it was delivered. Client names, internal numbers, and anything the client has not approved stay off the page.

That is a deliberate trade. We would rather show our work in full, on the client's terms, than a wall of logos we cannot stand behind. If proof of a specific pattern matters to your decision, ask, and we will walk you through comparable work in a scoped conversation.

Selected engagements

One build in full, and the wider body of work.

21 engagements available.

PE-BACKEDNational Legal Services FirmLegal

Automating UCC Document Extraction and Classification

85%
reduction in manual processing time
The challenge
Paralegals spent 20+ hours per week manually extracting, reviewing, and classifying hundreds of UCC filings from a state database. Error rates were climbing as volume increased, and substantive legal work was being displaced by data entry.
What we built
Built an end-to-end pipeline that uses Python to retrieve and extract filing data, Claude Sonnet to classify standard records, and Claude Opus to handle complex edge cases. A custom web interface routes the results to paralegals for review, preserving human control over exceptions. The complete system runs within the firm's existing Azure infrastructure.

Supporting outcomes

96.2%
classification accuracy on standard filings
45 sec
per-filing processing time, down from 12 minutes
0
compliance incidents in first 3 months
Timeline
6 weeks
Phases
4

Built with

  • Python
  • Claude Sonnet
  • Claude Opus
  • Azure
  • Custom Web UI

Engagement archive

The wider body of work

20 engagements

VC-BACKEDMobile Auto Detailing OperatorAutomotive Services

Running a Mobile Detailing Business on One Platform

Customer booking, dispatch, communication, and payment collection were spread across manual processes and general-purpose tools. Operations lacked one source of truth for scheduling work, coordinating technicians, and keeping customers informed from booking through completion.

Full-Stack Web & SaaSWorkflow AutomationData EngineeringView engagement
The challenge
Customer booking, dispatch, communication, and payment collection were spread across manual processes and general-purpose tools. Operations lacked one source of truth for scheduling work, coordinating technicians, and keeping customers informed from booking through completion.
What we built
Vectrel designed and built a role-based operations platform connecting a mobile-first customer booking experience, an admin dispatch console, and a technician field workflow. The shared system handles service selection, scheduling, PostGIS-based assignment, Stripe deposits, status transitions, required photo proof, recurring service, Resend communications, and operational reporting. A versioned API and shared TypeScript packages preserve a path to native mobile clients without rebuilding the backend.
Timeline
12 working weeks (contracted)
Phases
4

Built with

  • Next.js
  • TypeScript
  • Supabase
  • PostgreSQL + PostGIS
  • Stripe
  • Resend
  • Vercel
  • Turborepo
VC-BACKEDEcommerce PlatformRetail

Recommendation Engine for Real-Time Product Relevance

34%
increase in average order value

A growing e-commerce platform relied on basic rule-based recommendations that failed to personalize at scale. Conversion rates on product pages had plateaued, and the existing system could not account for real-time browsing behavior or seasonal shifts in demand.

Custom AI DevelopmentFull-Stack Web & SaaSData EngineeringView engagement
The challenge
A growing e-commerce platform relied on basic rule-based recommendations that failed to personalize at scale. Conversion rates on product pages had plateaued, and the existing system could not account for real-time browsing behavior or seasonal shifts in demand.
What we built
Designed and deployed a recommendation engine that combines collaborative filtering with contextual language-model analysis. The system processes browsing patterns, purchase history, and product metadata in real time, then ranks relevant products for each customer context. Those recommendations are surfaced throughout the customer journey so the experience can respond to current behavior, product context, and changing demand.

Supporting outcomes

2.8x
improvement in recommendation click-through rate
18%
uplift in repeat purchase rate within 30 days
<200 ms
recommendation generation latency
Timeline
10 weeks
Phases
5

Built with

  • Python
  • TensorFlow
  • Redis
  • Next.js
  • AWS
ENTERPRISERegional Healthcare NetworkHealthcare

Unified Patient Data Pipeline and Intake Classification

60%
reduction in patient intake processing time

Patient data was fragmented across three separate EHR systems, two legacy databases, and manual spreadsheet processes. Clinical teams spent hours reconciling records, while inconsistent intake classification delayed routing and made it harder to direct each patient to the right care team.

Data EngineeringCustom AI DevelopmentWorkflow AutomationView engagement
The challenge
Patient data was fragmented across three separate EHR systems, two legacy databases, and manual spreadsheet processes. Clinical teams spent hours reconciling records, while inconsistent intake classification delayed routing and made it harder to direct each patient to the right care team.
What we built
Built a unified data pipeline that consolidates records from the EHR systems, legacy databases, and spreadsheet workflows into one queryable system. A classification layer evaluates incoming records for urgency, specialty, and required documentation, then routes each case to the appropriate care team in real time. Clinical operations can work from the consolidated record instead of reconciling each source manually.

Supporting outcomes

93%
accuracy in automated case classification
4 hrs
saved daily across the clinical operations team
100%
HIPAA-compliant architecture
Timeline
12 weeks
Phases
6

Built with

  • Python
  • Azure
  • PostgreSQL
  • Next.js
  • Claude
VC-BACKEDSaaS StartupTechnology

Unified SaaS Analytics and Decision Intelligence

3x
faster decision-making on key metrics

A growing SaaS company lacked visibility into key business metrics. Their data lived in Stripe, Intercom, and a custom PostgreSQL database with no unified reporting. The founding team was making decisions on gut feeling rather than data.

Full-Stack Web & SaaSCustom AI DevelopmentData EngineeringView engagement
The challenge
A growing SaaS company lacked visibility into key business metrics. Their data lived in Stripe, Intercom, and a custom PostgreSQL database with no unified reporting. The founding team was making decisions on gut feeling rather than data.
What we built
Designed and built a full-stack analytics dashboard that consolidates Stripe, Intercom, and PostgreSQL data into one current view of the business. An analysis layer detects anomalies, generates plain-language trend summaries, and alerts the team to churn risk and expansion opportunities. The resulting workflow gives founders a consistent place to review the underlying metrics before making operating decisions.

Supporting outcomes

22%
reduction in involuntary churn from early detection
$140K
in expansion revenue identified in first quarter
Real-time
unified view across all data sources
Timeline
8 weeks
Phases
4

Built with

  • Next.js
  • TypeScript
  • PostgreSQL
  • Stripe API
  • Claude
ENTERPRISENational Insurance ProviderInsurance

Claims Email Routing Automation

87%
reduction in manual effort

Incoming claims emails were manually triaged by a team of adjusters who spent 15+ hours per week reading, categorizing, and forwarding messages to the correct department. Misrouted claims caused delays, duplicate work, and frustrated policyholders.

Workflow AutomationCustom AI DevelopmentView engagement
The challenge
Incoming claims emails were manually triaged by a team of adjusters who spent 15+ hours per week reading, categorizing, and forwarding messages to the correct department. Misrouted claims caused delays, duplicate work, and frustrated policyholders.
What we built
Deployed a classification pipeline that reads incoming claims emails, extracts the policy number, claim type, and urgency, then routes each message to the appropriate claims team. The automated path replaces manual reading, categorization, and forwarding for standard messages, reducing the triage cycle from hours to minutes and sending each claim to the appropriate team.

Supporting outcomes

94%
routing accuracy on first pass
12 min
average time to assignment, down from 4 hours
30%
faster average claim resolution time
Timeline
5 weeks
Phases
3

Built with

  • Python
  • Azure
MID-MARKETMulti-Office Business Law FirmLegal

AI-Assisted Matter Intake and Conflict Review

63%
reduction in pre-conflict intake preparation time

New matter requests arrived through emails, PDFs, and web forms, forcing intake staff to normalize party names, summarize requests, and chase missing details before conflicts review could begin. Every incomplete submission added another handoff and delayed the point when reviewers received a usable matter file.

Custom AI DevelopmentWorkflow AutomationFull-Stack Web & SaaSView engagement
The challenge
New matter requests arrived through emails, PDFs, and web forms, forcing intake staff to normalize party names, summarize requests, and chase missing details before conflicts review could begin. Every incomplete submission added another handoff and delayed the point when reviewers received a usable matter file.
What we built
Built an intake workflow that processes requests from email, PDF, and web forms, extracts the relevant entities, and flags missing information before conflict review. The system drafts a consistent matter summary and routes complete files into a lightweight internal portal. Coordinators review the extracted details and resolve exceptions instead of rebuilding every intake from the original materials.

Supporting outcomes

88%
of incomplete submissions flagged before coordinator review
14 min
average time from intake receipt to review-ready file
29%
fewer internal handoffs per new matter
Timeline
6 weeks
Phases
4

Built with

  • Next.js
  • TypeScript
  • Python
  • PostgreSQL
  • Azure
  • OpenAI
MID-MARKETRegional Real Estate Law FirmLegal

Commercial Lease Abstracting Pipeline for Faster Diligence

58%
reduction in first-pass lease abstraction time

Paralegals manually pulled key dates, rent escalators, assignment clauses, and renewal terms from commercial lease packets into spreadsheets. During active deal periods, that repetitive abstraction work slowed diligence, limited how many packets the team could review, and made approved lease information difficult to reuse.

Custom AI DevelopmentData EngineeringWorkflow AutomationView engagement
The challenge
Paralegals manually pulled key dates, rent escalators, assignment clauses, and renewal terms from commercial lease packets into spreadsheets. During active deal periods, that repetitive abstraction work slowed diligence, limited how many packets the team could review, and made approved lease information difficult to reuse.
What we built
Delivered a document-processing pipeline that reads commercial lease packets and extracts the standard dates, rent terms, assignment clauses, and renewal provisions required for diligence. Low-confidence clauses are routed to a paralegal for review rather than accepted automatically. Once approved, each abstract is stored in a searchable internal dataset so the validated lease information can be reused across matters.

Supporting outcomes

92.4%
field extraction accuracy on standard commercial lease packets
3.1x
faster turnaround on renewal-review requests
150+
lease records consolidated into a searchable dataset during the pilot
Timeline
8 weeks
Phases
5

Built with

  • Python
  • AWS
  • Bedrock
  • Anthropic
  • Snowflake
  • Next.js
ENTERPRISEAm Law 200 Litigation PracticeLegal

Medical Record Chronology Automation for Litigation Prep

72%
reduction in first-pass chronology preparation time

Litigation teams were manually reviewing thousands of pages of medical records, billing statements, and deposition transcripts to build case chronologies for deadline-sensitive matters, creating duplicate review effort and slow handoffs from paralegals to attorneys.

Custom AI DevelopmentWorkflow AutomationFull-Stack Web & SaaSView engagement
The challenge
Litigation teams were manually reviewing thousands of pages of medical records, billing statements, and deposition transcripts to build case chronologies for deadline-sensitive matters, creating duplicate review effort and slow handoffs from paralegals to attorneys.
What we built
Built a secure chronology workspace that ingests medical records, billing statements, and deposition transcripts, then runs OCR and entity extraction across the case file. The system identifies dates, providers, treatments, and gaps in care and drafts source-linked timeline entries. Paralegals review each entry against its source before exporting the chronology for the litigation team.

Supporting outcomes

89%
of timeline entries accepted without material edits
2.5 days
faster from document receipt to attorney-ready chronology
31%
increase in active matters handled per litigation support specialist
Timeline
8 weeks
Phases
4

Built with

  • Azure
  • Python
  • PostgreSQL
  • Next.js
  • Anthropic
MID-MARKETRegional Specialty Retail ChainRetail

Unified Inventory Exception Dashboard for a Specialty Retail Chain

27%
reduction in stockout-related lost sales across pilot categories

Merchandising teams reconciled stockouts, overstocks, and transfer requests across POS, ERP, and ecommerce exports. The spreadsheet workflow was already stale by the time decisions were made, obscuring likely lost-sales risks and slowing replenishment or transfer decisions across stores.

Data EngineeringWorkflow AutomationFull-Stack Web & SaaSView engagement
The challenge
Merchandising teams reconciled stockouts, overstocks, and transfer requests across POS, ERP, and ecommerce exports. The spreadsheet workflow was already stale by the time decisions were made, obscuring likely lost-sales risks and slowing replenishment or transfer decisions across stores.
What we built
Built a central pipeline that ingests the POS, ERP, and ecommerce exports and normalizes them into one inventory view. An internal dashboard surfaces stockout, overstock, and transfer exceptions, prioritizes the items most likely to create lost sales, and issues replenishment alerts by store and SKU cluster. Merchandising teams can work from the current exception queue instead of rebuilding spreadsheets.

Supporting outcomes

41%
fewer manual spreadsheet hours for the inventory team
18%
faster inter-store transfer decisions
25 min
warehouse refresh time, down from roughly 6 hours
Timeline
10 weeks
Phases
5

Built with

  • Python
  • dbt
  • Snowflake
  • Google Cloud
  • Vertex AI
  • Next.js
MID-MARKETDirect-to-Consumer Home Goods BrandRetail

Automating Returns and Warranty Case Triage

72%
reduction in manual triage time

A growing direct-to-consumer retailer handled returns, warranty claims, and damaged-shipment requests through a shared inbox. Support and operations repeatedly read the same messages, looked up order details, and reassigned cases, creating duplicate work and slowing the first response to customers.

Workflow AutomationCustom AI DevelopmentFull-Stack Web & SaaSView engagement
The challenge
A growing direct-to-consumer retailer handled returns, warranty claims, and damaged-shipment requests through a shared inbox. Support and operations repeatedly read the same messages, looked up order details, and reassigned cases, creating duplicate work and slowing the first response to customers.
What we built
Implemented an assisted triage workflow that reads each incoming request, identifies whether it concerns a return, warranty claim, or damaged shipment, and extracts the relevant order details. The system places the case in the correct operations queue and proposes the next action. Staff retain the final decision while no longer having to classify and reconstruct every request manually.

Supporting outcomes

93%
first-pass routing accuracy
84%
same-day first response rate, up from 61%
8 min
average time to assignment, down from 47 minutes
Timeline
5 weeks
Phases
3

Built with

  • Next.js
  • Node.js
  • Python
  • Supabase
  • OpenAI
  • Vercel
MID-MARKETApparel RetailerRetail

AI-Assisted Catalog Enrichment for Faster Product Launches

46%
faster new-product launch turnaround

Ecommerce merchandisers manually wrote product copy, standardized attributes, and tagged seasonal launches across thousands of SKUs. The workload delayed new-product publication and produced inconsistent descriptions and attribute coverage between categories, especially when large launch batches arrived together.

Custom AI DevelopmentData EngineeringWorkflow AutomationView engagement
The challenge
Ecommerce merchandisers manually wrote product copy, standardized attributes, and tagged seasonal launches across thousands of SKUs. The workload delayed new-product publication and produced inconsistent descriptions and attribute coverage between categories, especially when large launch batches arrived together.
What we built
Built a catalog enrichment pipeline that processes each launch batch, drafts on-brand product descriptions, normalizes category attributes, and applies the required tags across SKU records. Merchandisers review the generated fields before publication rather than accepting them automatically. Approved records are then published into the retailer's commerce and search systems through one consistent workflow.

Supporting outcomes

67%
reduction in manual copywriting hours for launch batches
21%
improvement in on-site search click-through for enriched categories
95%
attribute completeness on launch-day SKU records
Timeline
7 weeks
Phases
4

Built with

  • Python
  • PostgreSQL
  • AWS
  • Bedrock
  • Anthropic
  • Hugging Face
MID-MARKETMulti-Site Specialty Clinic GroupHealthcare

Automating Specialty Referral Intake and Clinical Triage

52%
reduction in referral-to-review time

Referral coordinators manually reviewed faxed and emailed referrals, checked each packet for missing documentation, and routed cases across multiple specialty teams. Incomplete packets moved back and forth through the operation, creating avoidable review work and delaying when patients could be scheduled.

Workflow AutomationCustom AI DevelopmentFull-Stack Web & SaaSView engagement
The challenge
Referral coordinators manually reviewed faxed and emailed referrals, checked each packet for missing documentation, and routed cases across multiple specialty teams. Incomplete packets moved back and forth through the operation, creating avoidable review work and delaying when patients could be scheduled.
What we built
Built a document-ingestion and triage workflow for faxed and emailed referrals. The system extracts referral details, checks each packet for missing records, and flags incomplete submissions. Complete cases are routed to the correct specialty queue through a lightweight internal operations portal, replacing the manual sequence of reviewing documents and assigning referrals across specialty teams.

Supporting outcomes

91%
first-pass routing accuracy on complete referrals
38%
reduction in referrals returned for missing documentation
6.5 hrs
saved daily across referral operations staff
Timeline
8 weeks
Phases
4

Built with

  • Next.js
  • TypeScript
  • Python
  • PostgreSQL
  • Azure
  • OpenAI
  • Vercel
ENTERPRISERegional Outpatient Care NetworkHealthcare

Unifying Provider Roster Data Across EHR and Credentialing Systems

73%
reduction in manual roster reconciliation work

Provider roster data lived across the EHR, credentialing software, payer spreadsheets, and manual updates. Conflicting records created duplicate providers and mismatches in downstream systems, forcing operations staff to reconcile the same information before scheduling, reporting, and payer contracting work could continue.

Data EngineeringWorkflow AutomationView engagement
The challenge
Provider roster data lived across the EHR, credentialing software, payer spreadsheets, and manual updates. Conflicting records created duplicate providers and mismatches in downstream systems, forcing operations staff to reconcile the same information before scheduling, reporting, and payer contracting work could continue.
What we built
Designed a central data pipeline and warehouse layer that brings together provider records from the EHR, credentialing system, payer files, and manual updates. Nightly reconciliation identifies duplicates and mismatches, then surfaces exceptions for operations staff. The resulting dataset provides a single trusted roster view for downstream reporting, scheduling workflows, and recurring payer updates.

Supporting outcomes

96%
reduction in duplicate provider records
4x
faster turnaround on monthly payer roster updates
99.3%
successful nightly sync rate after stabilization
Timeline
10 weeks
Phases
5

Built with

  • Python
  • SQL
  • Snowflake
  • dbt
  • AWS
  • Supabase
MID-MARKETIndependent Revenue Cycle Management ProviderHealthcare

Medical Coding Quality Review Assistant

2.3x
increase in charts reviewed per QA analyst

Coding quality specialists manually compared charts with coded outputs to find documentation gaps and coding inconsistencies. The time required for each chart limited audit coverage, delayed feedback to coding teams, and made it difficult to focus reviewers on the records most likely to need attention.

Custom AI DevelopmentData EngineeringView engagement
The challenge
Coding quality specialists manually compared charts with coded outputs to find documentation gaps and coding inconsistencies. The time required for each chart limited audit coverage, delayed feedback to coding teams, and made it difficult to focus reviewers on the records most likely to need attention.
What we built
Built a coding review assistant that compares chart notes with coded outputs and highlights likely documentation or coding discrepancies. Confidence signals prioritize the charts most likely to require attention, while the original notes and coded record remain available to the reviewer. Human quality specialists make the final determination, allowing the team to expand coverage without adding headcount.

Supporting outcomes

89%
precision on high-confidence discrepancy flags
41%
reduction in average QA review time per chart
28%
faster feedback loop to coding teams
Timeline
12 weeks
Phases
5

Built with

  • Python
  • Hugging Face
  • Anthropic
  • AWS
  • PostgreSQL
MID-MARKETVertical SaaS ProviderTechnology

Automated Post-Sales Handoff for a Vertical SaaS Onboarding Team

46%
reduction in time from closed-won to implementation kickoff

After contracts closed, onboarding details still moved through Slack threads, spreadsheets, and email. Implementation leads had to reconstruct the customer requirements, chase missing fields, and manually coordinate downstream tasks, causing kickoff delays and incomplete handoff packets between sales and onboarding.

Workflow AutomationData EngineeringFull-Stack Web & SaaSView engagement
The challenge
After contracts closed, onboarding details still moved through Slack threads, spreadsheets, and email. Implementation leads had to reconstruct the customer requirements, chase missing fields, and manually coordinate downstream tasks, causing kickoff delays and incomplete handoff packets between sales and onboarding.
What we built
Automated the post-sales handoff flow by extracting implementation details from the existing intake inputs and organizing them into a consistent handoff record. The workflow validates required fields, creates the downstream onboarding tasks, and alerts the responsible teams when launch blockers or incomplete information appear, reducing dependence on Slack threads, spreadsheets, and email for the standard handoff.

Supporting outcomes

97%
handoff completeness on first pass
82%
fewer manual status-update messages between sales and onboarding
11 hrs
saved weekly across revenue operations and implementation leads
Timeline
6 weeks
Phases
3

Built with

  • Node.js
  • TypeScript
  • PostgreSQL
  • Azure
  • OpenAI
  • n8n
VC-BACKEDInfrastructure Software VendorTechnology

Source-Grounded Technical Answer Assistant for Enterprise Sales

43%
reduction in time to prepare technical questionnaire responses

Solutions engineers spent too much time assembling accurate answers for RFPs, security questionnaires, and late-stage prospect follow-ups. The source material was fragmented across documentation, release notes, and internal knowledge bases, so every response required another manual search and verification pass.

Custom AI DevelopmentFull-Stack Web & SaaSData EngineeringView engagement
The challenge
Solutions engineers spent too much time assembling accurate answers for RFPs, security questionnaires, and late-stage prospect follow-ups. The source material was fragmented across documentation, release notes, and internal knowledge bases, so every response required another manual search and verification pass.
What we built
Built a source-grounded assistant that searches the approved documentation, release notes, and internal knowledge sources for each technical question. It drafts response language with citations back to the supporting material and presents the result in a controlled sales-engineering interface. A solutions engineer reviews and edits every answer before it is copied into a questionnaire or sent to a prospect.

Supporting outcomes

91%
of generated answers returned with at least one source citation
28%
increase in questionnaire throughput per month
<5 min
median response preparation time for common security questions
Timeline
8 weeks
Phases
4

Built with

  • Python
  • PostgreSQL
  • Next.js
  • AWS
  • Bedrock
  • Anthropic
  • Vercel
MID-MARKETProduct-Led Cloud Software CompanyTechnology

Product-Led Trial Scoring and Routing for a Cloud Software Company

23%
increase in demo-booking rate from product-qualified accounts

The growth team had a high volume of free-trial signups but no reliable way to identify which accounts showed real purchase intent. Manual qualification consumed time across the team, while slow follow-up made it harder for sales to reach promising accounts when their product activity was strongest.

Custom AI DevelopmentData EngineeringWorkflow AutomationView engagement
The challenge
The growth team had a high volume of free-trial signups but no reliable way to identify which accounts showed real purchase intent. Manual qualification consumed time across the team, while slow follow-up made it harder for sales to reach promising accounts when their product activity was strongest.
What we built
Built a behavioral scoring pipeline that combines product usage, firmographic enrichment, and account activity signals for each trial account. The system produces a consistent score, ranks the accounts, and routes high-intent opportunities to sales automatically. The ranked output gives growth and sales a shared prioritization list for weekly pipeline reviews without rebuilding qualification by hand.

Supporting outcomes

88%
enrichment coverage across new self-serve signups
35%
reduction in manual qualification time for the growth team
2.4x
faster weekly account prioritization during pipeline reviews
Timeline
10 weeks
Phases
5

Built with

  • Python
  • BigQuery
  • Vertex AI
  • Google Cloud
  • Next.js
  • TypeScript
MID-MARKETRegional Property & Casualty CarrierInsurance

Automating Policy Servicing Requests Across Email and Forms

64%
reduction in manual triage time

Endorsement, cancellation, and billing-related service requests arrived through shared inboxes and web forms in inconsistent formats. Policy service staff had to identify the request, locate the policy details, and determine whether information was missing before assigning work, creating a recurring queue backlog.

Workflow AutomationCustom AI DevelopmentView engagement
The challenge
Endorsement, cancellation, and billing-related service requests arrived through shared inboxes and web forms in inconsistent formats. Policy service staff had to identify the request, locate the policy details, and determine whether information was missing before assigning work, creating a recurring queue backlog.
What we built
Implemented an intake workflow that processes requests from both email and web forms, classifies the servicing need, and extracts the relevant policy details. The system validates whether the request is complete and routes standard work into the correct servicing queue. Incomplete or uncertain cases become explicit staff exceptions, keeping human review focused on the requests that need it.

Supporting outcomes

93%
correct queue assignment on first pass
46%
faster response time on standard endorsement requests
31%
reduction in rework caused by incomplete submissions
Timeline
6 weeks
Phases
3

Built with

  • Node.js
  • Python
  • PostgreSQL
  • AWS
  • Bedrock
  • Anthropic
MID-MARKETSpecialty Commercial Insurance MGAInsurance

Submission Workbench for Small Commercial Underwriting

37%
faster quote turnaround for in-appetite submissions

Underwriters pieced together broker emails, loss runs, supplemental applications, and appetite guidelines from multiple systems. The fragmented intake made it difficult to see whether a submission was complete or in appetite before review began, slowing quote turnaround on smaller commercial accounts.

Full-Stack Web & SaaSCustom AI DevelopmentData EngineeringView engagement
The challenge
Underwriters pieced together broker emails, loss runs, supplemental applications, and appetite guidelines from multiple systems. The fragmented intake made it difficult to see whether a submission was complete or in appetite before review began, slowing quote turnaround on smaller commercial accounts.
What we built
Built a web-based underwriting workbench that consolidates data from broker emails, loss runs, supplemental applications, and appetite guidelines. The system generates a structured risk summary for each submission and surfaces missing items before the file reaches the underwriting queue, reducing the manual assembly required before an underwriter can begin reviewing an in-appetite submission.

Supporting outcomes

29%
increase in submissions processed per underwriter assistant
84%
of missing-document issues identified before underwriter handoff
22%
reduction in broker follow-up cycles per submission
Timeline
9 weeks
Phases
4

Built with

  • Next.js
  • TypeScript
  • Google Cloud
  • Vertex AI
  • Snowflake
  • Vercel
MID-MARKETSpecialty MGAInsurance

Automated Bordereaux Reconciliation for Carrier Reporting

78%
reduction in monthly reconciliation cycle time

The operations team received bordereaux, cancellation files, and commission statements from multiple broker and carrier partners in inconsistent spreadsheet formats. Every month, analysts remapped the files, reconciled totals, and chased inconsistent data, creating a recurring bottleneck in carrier reporting and month-end close.

Data EngineeringWorkflow AutomationFull-Stack Web & SaaSView engagement
The challenge
The operations team received bordereaux, cancellation files, and commission statements from multiple broker and carrier partners in inconsistent spreadsheet formats. Every month, analysts remapped the files, reconciled totals, and chased inconsistent data, creating a recurring bottleneck in carrier reporting and month-end close.
What we built
Implemented a reconciliation pipeline and internal review console that ingests each partner spreadsheet and maps it to a canonical policy schema. The system checks totals and exposure deltas against prior submissions, identifies missing or inconsistent data, and routes only low-confidence exceptions to analysts. The standardized records and analyst-reviewed exceptions support the recurring carrier-reporting workflow in one consistent format.

Supporting outcomes

97.4%
automatic field-mapping accuracy across recurring partner formats
2.1 days
faster month-end reporting close
43%
fewer manual partner follow-ups for missing or inconsistent data
Timeline
9 weeks
Phases
5

Built with

  • Python
  • Snowflake
  • PostgreSQL
  • Next.js
  • Azure

Method as proof

Anonymized does not mean vague.

A write-up here carries the real problem, the real architecture, the real delivery shape, and the outcomes approved for publication; only the client's identity and unapproved details stay private. We build the same way we present: agree on what a system has to do, prove it on your real cases before the full build, and hand it over with the ownership terms written down. You can see that method end to end, and how we handle your data, before you ever raise your hand.

Start a project

Bring us the workflow, product, or operating question.

We will tell you what to prove first, show you the delivery path, and be candid about whether Vectrel is the right fit. Then you decide.