Selected work

Real teams. Real production.

Engagements shipped to production with measurable outcomes. Names sanitized; the work is not.

RAG & ChatbotsCustom AI AgentsAI StrategyLLM Integration

RAG & ChatbotsB2B Software · Mid-market SaaS company

Cut support ticket volume 42% with a docs-grounded assistant

Problem

Support team was drowning in repetitive product questions. The help center had answers, but customers couldn't find them.

What we built

Built a Claude-powered chat assistant trained on the docs site, changelog, and resolved Zendesk tickets. Answers cite source pages and escalate to humans when confidence is low.

Result

Live across the docs site and in-app. Customers self-serve in seconds; support handles the hard cases.

42%

Ticket deflection

12 sec

Median resolution

4.6/5

CSAT on AI replies

Claude Sonnet 4.5PineconeNext.jsZendesk API

Custom AI AgentsIndustrial Services · Enterprise B2B sales org

Account research that used to take 90 min, now runs in 4

Problem

SDRs were spending half their day on pre-call research — pulling 10-Ks, news, LinkedIn, and CRM history into briefing docs.

What we built

Agent that takes a target company and produces a structured briefing: financial signals, recent news, key contacts, hypothesized pain points, and a suggested opening message.

Result

SDRs run it before every call. Cycle time crashed; reps spend the saved hours actually selling.

↓ 95%

Research time

↑ 38%

Calls/rep/week

8K+/mo

Briefings generated

Claude Opus 4TavilySalesforceLangGraph

AI StrategyRetail · Fortune 1000 retailer

From 47 AI ideas to a focused 12-month roadmap

Problem

Every department was pitching AI projects. Leadership had no framework to prioritize, no view of dependencies, no ROI model.

What we built

Six-week engagement: opportunity audit across 11 business units, build-vs-buy analysis, ROI modeling, and a sequenced roadmap with governance guardrails.

Result

Board-approved roadmap. Three flagship projects greenlit; 30+ others sequenced or sunset. Internal AI council operating from the framework we built.

Opportunities triaged

3 flagship

Greenlit projects

$28M+

Projected 3-yr value

WorkshopsROI modelingGovernance design

LLM IntegrationInsurance · Insurance carrier

Fine-tuned classifier replaces a brittle rules engine

Problem

Legacy keyword rules for claim routing missed nuance and required constant tuning. New claim types broke the system weekly.

What we built

Fine-tuned a small open-source model on 18 months of historical claims with human labels. Deployed behind a streaming API with eval gates on every release.

Result

Higher accuracy than the rules engine, lower cost than frontier-model prompting. Maintenance burden dropped to near-zero.

94.3%

Routing accuracy

+22 pts

vs rules engine

↓ 87%

Inference cost

Llama 3.1 8BvLLMModalPython

RAG & ChatbotsHealthcare · Multi-hospital health system

Clinical decision support grounded in 80K internal protocols

Problem

Clinicians wasted minutes per encounter searching the intranet for the right protocol, drug interaction, or formulary policy. Old SharePoint search was unusable.

What we built

HIPAA-compliant RAG over 80K internal documents (protocols, formularies, policies, recent literature). Citations on every answer, audit log per query, role-based access controls.

Result

Rolled out to 3,200 clinicians. Average lookup time dropped from minutes to seconds. Adoption hit 78% within 8 weeks.

80K

Documents indexed

↓ 91%

Avg lookup time

78%

Clinician adoption

Claude (Bedrock)OpenSearchAWS HIPAAOkta SSO

Custom AI AgentsFinancial Services · Top-20 US bank

AML investigation copilot triages 10× more alerts per analyst

Problem

AML team was buried under low-quality alerts. Each case took 45+ minutes of cross-system data gathering before an analyst could make a judgment.

What we built

Agent that ingests an alert, pulls relevant data from 7 internal systems, builds a transaction graph, drafts a risk narrative, and surfaces the top three hypotheses with citations. Analyst stays the decision-maker.

Result

Analysts now triage in minutes. False-positive escalations dropped sharply. Audit trail satisfies model risk governance.

↑ 10×

Cases/analyst/day

↓ 78%

Time per case

↓ 31%

False escalations

Claude Sonnet 4.5LangGraphSnowflakeNeo4j

RAG & ChatbotsE-commerce · DTC apparel brand

Personal shopper chat lifted AOV 23% on the storefront

Problem

Conversion was capped by a clunky filter UI. Customers couldn't articulate what they wanted, and the search bar surfaced wrong results.

What we built

Conversational shopping assistant with vision: customer describes occasion, style, fit, or uploads a reference photo. Agent grounds suggestions in live inventory, sizing data, and reviews. Handles size questions, returns FAQs, and gift cards too.

Result

Live on the homepage and product pages. 31% of shoppers engage; engaged shoppers convert 2.1× higher and spend more per order.

↑ 23%

AOV (engaged)

2.1×

Conversion lift

31%

Engagement rate

Claude Sonnet 4.5VisionShopifyPineconeVercel

LLM IntegrationManufacturing · Industrial manufacturer

Ops copilot turns sensor anomalies into prioritized work orders

Problem

Plant ops drowned in sensor alerts. Most were noise; the real signals slipped through. Diagnosing a flagged anomaly required pulling logs from 4 systems by hand.

What we built

Hybrid system: classical anomaly detection on the time-series, then an LLM agent that contextualizes the anomaly with historical incident data, manuals, and recent maintenance logs. Outputs a draft work order with severity, likely cause, and recommended action.

Result

Two pilot plants saw a step-change in mean time to acknowledge. Maintenance teams stopped chasing false alarms; real anomalies got addressed faster.

↓ 64%

Mean time to ack

↓ 73%

False alarm fatigue

↓ 19%

Unplanned downtime

Claude Opus 4AWS IoTTimestreamServiceNow

Custom AI AgentsLegal · Mid-market law firm

First-pass contract review in minutes, not days

Problem

Junior associates spent days on first-pass review of MSAs, NDAs, and SaaS agreements — tedious, error-prone, and expensive for clients.

What we built

Review agent fine-tuned on the firm's playbook. Flags deviations from preferred positions, suggests redlines, drafts negotiation rationale, and outputs a memo. Always reviewed by an associate before going to client.

Result

Cut first-pass turnaround from days to under two hours. Associates spend their time on judgment calls, not boilerplate. Firm now offers fixed-fee review for SMB clients.

↓ 89%

First-pass time

↑ 27%

Issues caught

Launched

New SMB pricing tier

Claude Opus 4Custom playbook RAGMicrosoft 365iManage

Custom AI AgentsEducation · K-12 EdTech platform

Socratic AI tutor that doesn't just give kids the answer

Problem

Existing chat-based homework helpers were essentially answer machines. Teachers and parents distrusted them; learning outcomes weren't improving.

What we built

Tutor agent designed around pedagogy, not Q&A. It asks scaffolded questions, gives partial hints, identifies misconceptions, and adapts to grade level. Teacher dashboard shows where each student got stuck.

Result

Pilot in 12 districts. Pre/post assessments showed measurable lift on novel problems — not just homework completion. Now in production curriculum-wide.

+18%

Skill-mastery lift

8.4/10

Teacher trust score

120K+

Daily active learners

Claude Sonnet 4.5Custom eval harnessPostgresVercel

AI StrategyPrivate Equity · PE-backed portfolio company

Stood up an AI org from zero in 90 days for a sponsor's portco

Problem

PE sponsor wanted AI in the value-creation plan but the portfolio company had no AI talent, no data infrastructure, and no governance. Investment committee wanted measurable progress in two quarters.

What we built

90-day engagement: hired a VP of AI on their behalf, set up data plumbing, designed the eval and governance frameworks, and shipped two pilot projects (pricing copilot + customer ops agent). Knowledge-transferred everything to the new VP.

Result

Both pilots in production by day 90. New VP fully onboarded. Sponsor's IC took the playbook to two more portcos.

47 days

Time to first AI in prod

2 of 2

Pilots shipped

2 portcos

Sponsor playbook reuse

StrategyHiringEval designGovernance

Custom AI AgentsHealthcare · Multi-location dental group

Voice agent answers 100% of after-hours calls

Problem

Front desks missed 30%+ of calls during peak hours and 100% after hours. Each missed call was a lost appointment and an unhappy patient.

What we built

Voice AI receptionist that books, reschedules, and answers FAQs in natural conversation. Handles English and Spanish, knows each location's hours and providers, and texts patients confirmations. Hands off to human staff for anything sensitive.

Result

Capture rate on inbound calls hit 98%. After-hours bookings became a meaningful new channel. ROI inside the first month.

98%

Call capture rate

+340/mo

After-hours bookings

<30 days

Payback period

Claude Sonnet 4.5VapiTwilioCustom EHR API

Ready to ship AI that actually works?

Tell us what you're building or what you're stuck on. We'll come back within 24 hours with a no-fluff plan.

Book a discovery call sales@globalsmartit.com