No case studies match those filters.
RAG-based support agent cut first-response time in half
How we built a retrieval-augmented support agent for a regulated lender: architecture, guardrails, and the tradeoffs that got it to 61% ticket deflection with zero compliance incidents.
AgentsFintech
Read the full case study →
Document pipeline replaced a manual triage queue
How we built a document classification and extraction pipeline for a freight brokerage that replaced a manual triage queue and quadrupled throughput.
CV & NLPLogistics
Read the full case study →
In-app shopping copilot, integrated with zero downtime
How we integrated a native conversational shopping assistant into an existing retail app's real inventory and cart flow, shipped in six weeks with zero downtime.
Mobile AIRetail
Read the full case study →
Fourteen AI ideas, scored down to three worth building
How we ranked fourteen AI ideas from a healthcare network down to three fundable builds, using structured feasibility, value, and compliance-risk scoring instead of a slide deck.
AI StrategyHealthcare
Read the full case study →
A build-vs-buy call that saved a fintech from a $400k platform contract
How a structured six-week evaluation saved a fintech company from a $400k vendor platform contract that couldn't actually do what the sales deck promised.
AI StrategyFintech
Read the full case study →
An ROI model picked the one AI project that paid back in seven months
How a four-week ROI modeling engagement ranked nine AI proposals for a third-party logistics operator by measured payback, and why the winner was the least exciting one on the list.
AI StrategyLogistics
Read the full case study →
A natural-language analytics product, built AI-native from zero
How we built a natural-language analytics SaaS product from zero, validating the core query pipeline before any UI work, and why a two-step text-to-query design kept the model from ever hallucinating a number.
AI-Native SaaSSaaS
Read the full case study →
An onboarding copilot cut new-account support tickets by 42%
How an in-product onboarding copilot for a B2B SaaS platform cut new-account support tickets by 42%, and why we built it to defer to human support rather than resolve everything itself.
AI-Native SaaSB2B SaaS
Read the full case study →
An AI-native proposal builder cut RFP drafting time by 68%
How we built an AI-native RFP response product for a professional services software company, with retrieval over past winning proposals, structured generation, and an evaluation harness that gated every prompt change.
AI-Native SaaSB2B SaaS
Read the full case study →
On-device AI cut latency and cost for a fitness app's core feature
How we built a real-time, on-device pose estimation feature for a fitness app, cutting cloud cost 76% and hitting 120ms latency by defaulting to on-device inference instead of the cloud.
Mobile AIFitness & Wellness
Read the full case study →
On-device personalization lifted repeat purchases without sending data off-phone
How on-device personalization for a retail app's product recommendations lifted repeat purchases 22% without sending any browsing or purchase data off the phone.
Mobile AIRetail
Read the full case study →
Offline on-device AI let utility inspectors complete 35% more inspections per shift
How we built an offline-first inspection app with on-device vision and speech-to-text for a regional utility, so field crews could capture structured defect reports with no cell signal.
Mobile AIEnergy
Read the full case study →
A multi-agent system cut fault triage time by 3.5 times
How we built a five-agent orchestrated system for manufacturing fault triage, and why we proved a single-agent design insufficient with real evaluation data before adding multi-agent complexity.
AgentsManufacturing
Read the full case study →
A contract-review agent cut legal's first-pass read time nine times over
How a contract-review agent for a mid-size legal team cut first-pass review time nine times over, by flagging deviations from a clause playbook instead of trying to fully automate review.
AgentsLegal
Read the full case study →
An IT service desk agent now resolves 64% of tier-1 requests end to end
How we built a tool-using AI agent for a 4,000-person company's IT service desk that resolves access requests, password resets, and software installs end to end, with scoped permissions and human approval for anything risky.
AgentsEnterprise IT
Read the full case study →
An internal MCP server gave every team's tools to one agent interface
How we built a single, centrally-owned MCP server exposing 14 internal tools across six engineering teams, replacing ad hoc integrations with a scoped, evaluated permission model.
MCPSaaS
Read the full case study →
One MCP server gave every sales agent the same five CRM tools
How one MCP server unified five separate CRM tools into a single, permission-scoped interface every sales agent uses, replacing five inconsistent internal scripts with one reviewed integration.
MCPB2B Sales Tooling
Read the full case study →
A governed MCP server cut ad-hoc data requests to the analytics team by 80%
How we built a Model Context Protocol server that gives business teams' AI assistants safe, read-only, semantically defined access to an e-commerce company's data warehouse.
MCPE-commerce
Read the full case study →
A fine-tuned model closed an 18-point accuracy gap prompting couldn't
How we fine-tuned a right-sized model for insurance claims classification, closing an 18-point accuracy gap over prompting alone while cutting inference cost 62%.
Custom ModelsInsurance
Read the full case study →
A custom fraud model cut false-positive declines by 61% without adding risk
How a custom fraud-detection model built for a payments company cut false-positive declines by 61% without increasing fraud loss rate, and the precision-recall tradeoffs that decision actually required.
Custom ModelsFintech
Read the full case study →
A custom demand forecast cut perishable waste by 23% across 38 stores
How we built a store-by-SKU demand forecasting model for a regional grocery chain's perishables, using gradient-boosted trees with weather, promotion, and holiday features, and cut spoilage by 23%.
Custom ModelsRetail
Read the full case study →
AI search added to an existing SaaS product without a rewrite
How we integrated semantic search into an existing project management SaaS product without a rewrite, dropping support tickets 31% by scoping it as better search, not a chatbot.
AI IntegrationSaaS
Read the full case study →
An AI tutor was added to a five-year-old LMS without touching its core
How we added an AI tutor and assistant into a five-year-old legacy LMS in five weeks without rewriting its core, working inside an aging codebase and a course-content model that predated any AI feature.
AI IntegrationEdtech
Read the full case study →
AI triage added to an existing property platform made work order dispatch 3.4 times faster
How we added AI maintenance request triage to an established property management platform through an isolated service and feature flags, with no changes to its core data model.
AI IntegrationPropTech
Read the full case study →
Cost and observability overhaul cut AI infrastructure spend 38%
How a cost-attribution audit found a silent retry-loop bug, an oversized model, and a missing cache, cutting an AI startup's model API spend 38% in six weeks.
MLOpsAI-Native Startup
Read the full case study →
Multi-model routing cut a media platform's inference bill by 47%
How multi-model routing, sending requests to smaller models when a large model wasn't needed, cut a media and publishing platform's inference spend by 47% without a measurable quality drop in user-facing output.
MLOpsMedia & Publishing
Read the full case study →
Self-hosted open-weight model serving cut per-request cost by 52%
How we moved a cybersecurity SaaS company's high-volume log summarization workload from a hosted API to a self-hosted, quantized open-weight model on Kubernetes with autoscaling, cutting cost per request by 52%.
MLOpsCybersecurity SaaS
Read the full case study →
A custom vision model ended a fatigue-driven accuracy swing
How a custom-trained vision model for manufacturing defect detection eliminated shift-fatigue-driven accuracy swings and let a production line run 3.1x faster.
CV & NLPManufacturing
Read the full case study →
NLP extraction turned six-week lease abstraction into two days
How NLP-based document extraction turned a six-week manual lease-abstraction process for a commercial real estate firm into a two-day process, and where we still required a human reviewer to sign off.
CV & NLPCommercial Real Estate
Read the full case study →
Edge vision for site safety cut false alarms by 71% on existing cameras
How we built a computer vision system for a general contractor that detects missing protective equipment and exclusion-zone breaches on existing site cameras, running on edge devices with privacy by design.
CV & NLPConstruction
Read the full case study →
One automated process, chosen by data instead of opinion
How we scoped one process out of a wish list of 'automate everything,' using real time-cost data instead of opinion, and built a human-in-the-loop onboarding review tool that saves four to five hours of coordinator time a week.
AutomationProfessional Services
Read the full case study →
An automated pipeline now processes 89% of invoices without a human touch
How an automated invoice-processing pipeline for a logistics company now handles 89% of incoming invoices without a human touch, and the exception-routing design that made the remaining 11% safe to leave manual.
AutomationLogistics & Freight
Read the full case study →
Automated KYC intake cut member account opening from five days to one
How we automated identity document intake, verification checks, and case routing for a credit union's new member onboarding, cutting account opening time from five business days to one while keeping compliance staff in control.
AutomationBanking
Read the full case study →
A real data foundation, built before any model work started
How we unified seven disconnected retail data sources into one reliable foundation before any forecasting model work began, and why we said no to building the model first.
Data EngineeringRetail
Read the full case study →
A rebuilt pipeline made twelve terabytes of daily sensor data usable
How a rebuilt data pipeline made twelve terabytes of daily IoT sensor data from a manufacturing plant reliably usable, after years of that data being collected but effectively unqueryable.
Data EngineeringManufacturing
Read the full case study →
Preparing 1.2 million regulated documents made retrieval accurate enough to trust
How we cleaned, deduplicated, versioned, and structured 1.2 million standard operating procedures and quality documents for a pharmaceutical manufacturer, taking retrieval accuracy from unusable to 96% on an evaluation set.
Data EngineeringLife Sciences
Read the full case study →
A voice agent that handles most calls without a human
How we built a low-latency voice agent for a physical therapy practice's phone line, handling 73% of calls end to end with a streaming pipeline built for phone-call latency from day one.
Voice AIHealthcare
Read the full case study →
A voice ordering agent lifted average order value by upselling correctly
How a voice ordering agent for a restaurant and hospitality chain lifted average order value 31% through correct, context-aware upselling, without making customers feel pushed.
Voice AIRestaurant & Hospitality
Read the full case study →
A voice agent booked 38% more service appointments by answering every call
How we built a voice agent for a dealership group's service departments that answers every call, books appointments directly in the dealer management system, and hands off to advisors with full context.
Voice AIAutomotive
Read the full case study →
A website copilot answered pre-sales questions a demo used to
How a curated, grounded website chatbot answered technical pre-sales questions for a B2B SaaS company, involved in 44% of qualified demo requests with zero fabricated pricing claims.
ChatbotsSaaS
Read the full case study →
A support chatbot deflected 58% of tier-1 tickets in its first quarter
How a support chatbot for a regional telecom provider deflected 58% of tier-1 support tickets in its first quarter live, by being strict about what it would answer rather than trying to cover everything.
ChatbotsTelecom
Read the full case study →
An order-aware support chatbot cut cost per contact by 71%
How we built an order-aware support chatbot for a direct-to-consumer apparel brand that resolves order status, returns, and exchanges through the brand's order management system, cutting cost per contact by 71%.
ChatbotsE-commerce
Read the full case study →
An adversarial security review found what internal QA missed
How an independent adversarial security review found 11 exploitable vulnerabilities, including a prompt injection through order-note data, in a customer-facing AI agent before its public launch.
GovernanceSaaS
Read the full case study →
An independent audit found four model risks an insurer's own team had missed
How an independent model-risk audit for an insurance company found four undisclosed model risks the insurer's own internal team had missed, and why the audit's value was in adversarial testing, not a compliance checklist.
GovernanceInsurance
Read the full case study →
An EU AI Act readiness program got a recruiting platform audit-ready in nine weeks
How we inventoried AI features, classified risk, and built the documentation, logging, and human-oversight controls a recruiting software company needed for AI systems the EU AI Act treats as high-risk.
GovernanceHR Tech
Read the full case study →