AI Automation Case Studies

The Dental Trauma Center in Los Angeles documents detailed clinical exams, and every one becomes a long written report. The notes start as handwriting and hand-marked diagrams, so a person had to read and retype all of it before a doctor could review the result.

We built Digital Ink to take that step away. Clinicians capture the exam on a tablet. A vision model reads the handwritten notes and the markings on the body diagrams, and the system drafts the report in the practice's own voice. Nothing goes out until the doctor and staff have reviewed it.

What we built

• Tablet capture for exam documentation
• Vision-model reading of handwritten clinical notes and diagram markings
• Report drafts written in the practice's voice and queued for doctor review
• A workflow layer with an admin dashboard and an audit log
• Deployment on secured AWS infrastructure, integrated with the practice's systems and built for HIPAA workloads

Since launch

Running at the practice on secured AWS infrastructure, and the practice has since referred us to another healthcare client.

Where else this works

The same intake-to-draft pipeline fits any practice buried in forms: clinical notes, insurance and authorization paperwork, consent forms and lab prescriptions. See AI document processing.

AST Defeasance advises on commercial loan defeasance and interest-rate caps. Every closing starts with borrower onboarding: collecting W-9s, formation documents, EIN letters, loan agreements and org charts, working out who the beneficial owners are, and filling out each partner bank's own forms.

We built an onboarding portal around that work. Borrowers upload through a protected, deal-specific link with no account to create. AI classifies each file and extracts its fields, and it is built to return a blank rather than guess. A rule engine, not the AI, decides beneficial ownership and records a plain-English reason for every row. The portal fills each bank's own forms, an analyst reviews and approves, and the signing package goes out for e-signature with every file saved to the firm's OneDrive.

Before that, our first project for AST turned years of archived deal documents into structured, CRM-ready records enriched with the deal contacts.

What we built

• Password-protected, deal-specific upload links with no borrower accounts
• AI document classification and field extraction across W-9s, formation documents, EIN letters and loan agreements
• Org-chart reading plus a deterministic beneficial-ownership rule engine with a written reason for every decision
• Automatic filling of each partner bank's own Word and PDF forms, across six bank document sets
• An analyst review gate, e-signature with progress tracking, and automatic filing to OneDrive
• An analyst workspace with a deal pipeline, an activity trail and year-over-year reporting

Results

• AST came back for a second project after the archive pipeline
• Six bank-specific document sets released to production after an audit of 33 documents and 216 pages, with all 126 text checks and 115 backend tests passing
• Nearly 1,400 historical deals converted into structured 17-field records with counterparty contacts, ready for CRM import

Where else this works

The pattern of protected upload, AI extraction, explicit rules, human sign-off and e-signature fits personal injury client onboarding and any firm that fills the same forms for every new client.

100 Boulevard handles Medicaid applications and recertifications for managed long-term care plans. Each application lists the member's income and assets, and each one has to be backed by a document: a bank statement, a pay stub, a Social Security letter, a 1099, a pension or annuity statement.

We built a new AI validation process inside their existing document-automation system. It picks up the cases flagged for validation in priority order, splits uploaded files into typed documents, chooses the best proof for each entry the intake rep made, checks the amount against a tolerance, updates the record and links the document it used. When it can't verify something, it writes a plain-language reason the team can act on, then moves the case along to the next stage.

What we built

• A priority-ordered case selector that runs alongside the existing process
• Document intake that splits multi-document PDFs, ZIPs and phone photos into typed records
• A matching engine that ranks proof documents and checks amounts against tolerance bands
• Plain-language reasons on every record it can't verify, plus an explain mode for staff
• Specialized readers for statements with several account holders, annual tax forms and proof of address
• A consent and drawn e-signature step for the member upload portal (in rollout)

Results

• Zero extraction errors across the first 30 production cases, per the client's reviewer
• 576 cases completed in one production day, enough to clear the queue on its own
• 20 cases selected in priority order and completed in 68 minutes
• Backed by an automated suite of 452 tests

Where else this works

Any team that checks documents against what a customer told them, such as insurance verification, loan files or records review, can run the same read, match and explain loop. See AI document processing.

Groovit turns quizzes into a 3D runner game that students control with their bodies through a webcam, while teachers, parents and school admins manage classes, assignments and rewards. The founder brought us in to make the platform ready for real schools.

We consolidated the codebase, rebuilt the production stack on the client's own AWS, Vercel and MongoDB Atlas accounts, and hardened per-school data isolation and server-side scoring. We moved pose detection off the main thread so the game runs smoothly on low-end school Chromebooks, added testing, guarded deployments and privacy-first monitoring, and moved the app onto groovit.com without breaking existing links. Movement detection runs entirely in the browser, so no video leaves the device.

What we built

• Production infrastructure on client-owned accounts, defined as code, with automatic rollback
• Per-school data isolation, server-side scoring and brute-force lockouts
• In-browser pose detection moved to a background worker with adaptive resolution
• Test plans, browser end-to-end tests, a load-test harness and monitoring
• New features including a market and missions redesign, a practice world and school administration

Results

• Classroom Chromebook frame rate up from about 10 to about 50 frames per second
• Production live on the client's own accounts about one week after work began
• 53 defects closed and verified in the first three and a half weeks
• 35 concurrent simulated students with zero errors in load testing
• Automated backend checks grew from 116 to 478, plus 170 frontend checks

Where else this works

Teams with a working product that isn't ready for real customers get a launch-readiness rebuild covering security, testing, deployments and performance without starting over. See custom software development.

JTravel helps Jewish travelers find each other and find a Shabbat meal wherever they land. The app was already live on the App Store, built by earlier developers, when the founders brought us in for a pilot.

Working on top of the existing codebase, we delivered a full dark-theme redesign, native email, Apple and Facebook login, six-photo traveler profiles, a home-base privacy radius, dated trips with matching to other travelers at the same destination, and a complete Shabbat meal marketplace. Hosts apply through the app, admins approve them in a separate web dashboard, and guests can RSVP through a public invite link without installing anything.

What we built

• An app-wide redesign with a new design system and navigation
• Native email, Sign in with Apple and Facebook login
• Expanded traveler profiles and a home-base privacy radius enforced by the API
• Dated trips with destination matching between residents and visitors
• A host-and-guest Shabbat meal marketplace with public RSVP links and an admin approval dashboard
• App Store hardening: automated image moderation, reporting and full account deletion

Results

• All six signed pilot deliverables delivered or materially addressed
• Work beyond the signed scope, including trip matching and a host-approval admin dashboard
• API test suite passing at the release audit: 30 suites and 391 tests

Where else this works

Existing apps that need a redesign, new features and App Store readiness without a rewrite. See custom software development.

Nick Matau goes live for hours at a time, most nights. The best moments were buried somewhere in hours of footage, and finding them by hand took longer than the stream itself.

We built a clipping agent he can drive from a simple form. He pastes a stream link and can add a plain-English request, like the moments where he fact-checks someone live. The agent downloads and transcribes the stream with word-level timestamps, identifies the speakers, splits it into conversations and maps the topics. Claude then proposes, filters and trims candidate clips against a profile tuned to his channel, stitching related moments together when that makes a better clip. The finished clips land in a shared Drive folder with a report explaining why each one was chosen.

What we built

• Form-based intake for a stream link or a Drive file, with optional plain-English requests
• Transcription with word-level timestamps and speaker identification
• Conversation segmentation and topic mapping across multi-hour streams
• A three-pass Claude clip finder that scores moments against a creator-specific profile
• Automatic cutting, stitched composite clips and delivery to Google Drive
• An always-on cloud worker, plus a review interface prototype

Results

• A full 4-hour-52-minute stream processed end to end, with 10 clips cut in about a minute
• One run produced 10 clips, 6 continuous and 4 stitched composites, each with a hook, a key quote and a score
• Running as an always-on cloud worker and maintained for about six months

Where else this works

Podcasts, webinars, interviews, sermons and other long-form video all have the same problem: hours of footage and a few minutes worth posting. See the AI video editing agent.

A New Jersey firm that packages utility-incentive HVAC upgrades for commercial and multifamily buildings needed to be found by property owners and managers, on Google and in AI answers, starting from a brand-new domain.

We built the site and run the search program around it: a technical audit and fixes, keyword research checked against the live results pages, focused service pages written to one search each, evergreen guides built on primary sources so AI tools can cite them, structured data on every page, and a robots policy that lets AI search read and cite the site while declining model training. Progress is reported monthly from Search Console, GA4 and Bing.

What we built

• A 70+ page static website with structured data on every page
• A technical audit and remediation pass, plus a full crawl review
• Bottom-of-funnel service pages, each written to a single search
• Answer-first openings and question-shaped sections built for AI Overviews
• Search Console, GA4, Bing Webmaster Tools and IndexNow measurement

Results

• Technical audit issues taken from 99 to 0, including canonical tags that would have kept the whole site out of Google
• First Google impressions three days after the domain went live
• A new service page cited as the #1 source in Google's AI Overview for its target search about two weeks after launch
• Searches with the site in Google's top 10 grew from 6 to 18 in a 28-day comparison, and average position on comparable commercial searches improved from 67.7 to 47.2

Where else this works

The same SEO, AEO and GEO program works for any local or regional service business, including dental practices, med spas and law firms. See AI SEO services.

A full-service dental laboratory receives prescription forms from dentists, mostly handwritten, on several different printed templates. Every one has to be read and keyed in before the case can start.

We built a pilot that turns each scan into a structured record with a confidence score on every field. Two independent vision models read each form, the readings are merged and checked against a dental vocabulary, and any field below the confidence threshold goes to a review panel or a person instead of being guessed. Every model call is logged with its cost and time.

What we built

• Scan normalization and automatic template matching
• Double reading by two independent vision models
• Validation against a dental controlled vocabulary
• Confidence scores on every field, with uncertain fields routed to review
• A cost and time ledger for every model call

Results

• Run on 6 real handwritten prescription forms
• About four minutes and under $1 of AI cost per prescription
• Scans grouped automatically into 3 printed-template families

Where else this works

Any team that keys handwritten or scanned forms into a system, from lab slips to intake forms, can use the same double-read and review pattern.

A private client needed a long, multi-step web process run reliably at volume, with results logged to a spreadsheet the whole team uses.

We built a concurrent automation engine with a self-recovering job queue: failures come back on an escalating schedule, interrupted runs resume where they stopped, and nothing waits on a stuck row. We then moved it from a server fleet into a self-contained desktop app with a local database and a live activity view, so each operator machine runs on its own. Alongside it, monitoring services check a list of pages on a schedule, compare each result with the last snapshot and send de-duplicated alerts to messaging apps.

What we built

• An asynchronous engine running several browser sessions at once through a multi-step process
• A self-recovering queue with escalating retries and crash recovery
• A native desktop app with a local database, live activity log and spreadsheet sync
• Page-monitoring and screenshot-to-data services with threshold and cooldown alerts
• Operator and developer documentation for the handover

Results

• Moved from a multi-server setup into one self-contained desktop app
• Handed over through five builds with operator and developer guides

Where else this works

Repetitive web tasks with no API, monitoring that has to catch changes fast, and reporting that lives in a spreadsheet. See workflow automation.