Fortune 500 tech leader · Enterprise Operations
Conversational AI
Operations Control-Tower AI Assistant
An enterprise-grade AI assistant layered onto a Fortune 500 technology leader's operations control-tower platform, there to strengthen the dashboards rather than replace them. It is a secure, role-aware conversational interface, grounded in live operational data, that understands the business rules behind the numbers (cycle time, SLA, managed deployment versus configuration services) and draws operational, predictive and qualitative insight together in answer to a single question.
60–70%
fewer ad-hoc analytics & reporting requests
8 months
from design to platform-wide rollout
Faster
partner performance & capacity decisions
Consistent
metric interpretation across teams & partners
RAG / retrievalRole-based accessLive operational dataBusiness-rule logic
Dental & medical-device manufacturer · Field Sales
AI platform · PWA
AI Field-Sales Platform — AI in every rep's pocket
A mobile-first platform that retires manual territory planning in favour of a daily, AI-generated route optimised for revenue, travel time and product opportunity. Six machine-learning models run as a closed-loop pipeline on Azure ML, and a GPT-4 layer reads each rep's visit notes to summarise the account and suggest the next move. Most field tools record what happened; this one tells the rep what to do next, and why.
25+
reps across the US & Canada
6
ML models, fully automated
0.83
precision on churn prediction (CatBoost)
GPT-4
in production on real sales notes
MILP route optimisationXGBoostCatBoostK-MeansGPT-4 (Azure OpenAI)Azure ML Studio
National government IP authority · GovTech / IP
Computer vision
National-scale trademark similarity AI
An AI-powered trademark conflict-detection system for a national government IP authority, one of the world's highest-volume trademark registries. Two front doors, a public applicant portal and an internal examiner tool, sit over a single similarity engine: image embeddings from VGG-16 and ResNet-50, text and phonetic matching, semantic meaning, and high-speed retrieval from an Oracle Vector DB, all returning ranked results with recommended actions. It was built to work with decades of legacy Oracle records.
~2,500
trademark applications received every day
National
scale — among the world's highest IP volumes
3-modal
visual, phonetic & semantic similarity
Live
in production, integrated with Oracle
VGG-16 · ResNet-50Image embeddingsOracle Vector DBPhonetic matchGIGW / WCAG
Dental & medical-device manufacturer · Healthcare CV
Visual search · PWA
Visual Product-Search App — point, shoot, identify
We replaced a slow catalogue lookup with visual search: a dentist photographs a diamond bur and the app identifies it on the spot. A multimodal LLM handles the primary identification, backed by a ResNet and YOLO-class hybrid for edge cases and new products, all wrapped in a production web app with a headless CMS, 3D product viewers and an ordering workflow. A focused twelve-week second phase trained custom vision models, among them YOLOv8, a 33-shape classifier and a fine-tuned OpenCLIP, to raise accuracy and coverage while cutting API calls and latency.
60% → 85%
correct bur in the top-5 image results
130 → 450
burs covered — the full product line
20,000+
augmented training images generated
PWA
works on any device, in the clinic, no install
Multimodal LLMYOLOv8OpenCLIPResNetHeadless CMS
US legal-services firm · Legal / Document AI
Agentic doc AI
Court-Fee & Invoicing AI — any document, one clean invoice
A high-volume monthly court-fee invoicing process in which files arrive in every format imaginable, from spreadsheets and PDFs to scanned mail, Word and plain text, each one previously extracted, cleaned and validated by hand. We built a queue of AI agents, one per file type, that read whatever the format and normalise it into a single clean structure, apply the firm's own rules for fee validation and docket matching, then generate invoices and statements into the accounting system without any re-keying. A second phase, now roughly seventy per cent complete, grows it into a full workflow portal, with Gemini improving extraction on scanned PDFs.
5 formats
Excel, PDF, scanned, Word & text — one pipeline
Agent / file
a dedicated AI agent per document type
Phase 1 live
the AI extraction core in production
End-to-end
extraction → validation → invoicing, no re-keying
Queue-based AI agentsMulti-format extractionGeminiGPT-4QuickBooks integration