Real projects.
Real results.

A mid-sized freight forwarding company had a team of three full-time operators whose entire job was opening PDFs from carriers, extracting line-item charges, validating them against purchase orders, flagging discrepancies, and entering approved invoices into their ERP system. The PDFs came from 60+ carriers in different formats, languages (English, German, Polish, Dutch), and layouts. Every exception required a human judgement call.

We built a multi-stage AI agent pipeline: a document classification layer identifies the carrier and invoice type; a structured extraction layer (fine-tuned on their historical invoices) pulls line items, charges, and reference numbers into a validated JSON schema; a rules engine compares extracted data against PO records in their ERP; confident matches are auto-approved and written back; discrepancies above a threshold are routed to a human review queue with a pre-filled context card. The entire pipeline runs on AWS with a Slack-based review interface.

A fast-growing D2C marketplace was onboarding 800–1,200 new products per week from brand partners. Each product needed to be classified into their 6-level category taxonomy (Category → Subcategory → Type → Material → Gender → Age Group) before it could go live. Manual tagging by a content team took 3–4 days per batch. Incorrect tags caused products to surface in wrong search results, damaging conversion rates and brand trust. Their existing attempt at rule-based classification had 61% accuracy and required constant maintenance.

We built a multi-label text classification model fine-tuned on 140,000 historical product records with verified tags. Input features: product title, description, brand name, and key attributes. Output: confidence-scored predictions for all 6 taxonomy levels simultaneously. Products above a configurable confidence threshold are auto-tagged and sent to live; below-threshold products are queued for human review with the model's top-3 predictions pre-populated. We built a lightweight review interface where the content team accepts, corrects, or overrides — corrections feed back into a monthly retraining cycle.

A UK-based B2B SaaS company had a 4-year-old React SPA that had accumulated significant technical debt: 11-second initial load times on standard connections, no server-side rendering (catastrophic for SEO), a deployment process that required manual coordination between two engineers, no test suite, and a frontend codebase that their CTO described as 'impossible to onboard new developers into.' They were losing deals to competitors with faster, more polished products and had just hired a new head of engineering who mandated a rebuild.

We rebuilt the product in Next.js 15 App Router with React Server Components, co-developing with their in-house team over 14 weeks. Key decisions: component library built from a custom design system (not shadcn or MUI) for full visual control; server components for all data-fetching-heavy pages; client components restricted to interactive UI; Prisma + PostgreSQL on Supabase for the database layer; Auth.js for authentication; GitHub Actions pipeline with type checking, unit tests, and zero-downtime ECS deployment on every merge to main. We ran the old and new systems in parallel for 6 weeks before full cutover.

A facilities management company deployed 120 technicians across remote sites in regional Australia. Job assignments were printed paper sheets. Completion photos were emailed from personal phones. Client sign-offs were wet signatures on physical forms. The back-office team spent 6–8 hours every day manually keying technician reports into their job management system. Data entry errors caused invoice disputes. Photos were frequently lost or untagged. There was no way to know job status until technicians returned to base at the end of the day.

We built an offline-first React Native (Expo) application using SQLite (via Expo SQLite + Drizzle ORM) as the local database. Every job, client record, and form template syncs to the device when connected. Technicians can work entirely offline — complete jobs, capture photos, collect digital signatures on the touchscreen, add notes — and all data syncs automatically when connectivity returns. The sync engine handles conflict resolution explicitly: the server timestamp wins for status updates; photo uploads are idempotent; signatures are append-only. A companion web dashboard gives back-office staff real-time job status and auto-generates invoices from completed job data.

A Series A healthtech company was storing patient intake data, appointment records, and clinical notes on a single EC2 instance with no encryption at rest, no automated backups, no access logging, and root account credentials shared among five engineers. They had just engaged a HIPAA compliance consultant who flagged 23 technical controls as failing or absent. Their Series B lead investor required a clean security posture before term sheet. They had 8 weeks.

We conducted a 5-day audit producing a prioritised findings report, then executed a phased migration. Critical (week 1–2): enable CloudTrail logging, remove root account usage, create individual IAM users with MFA, encrypt the RDS instance at rest, enable VPC Flow Logs. High priority (week 3–5): migrate to Multi-AZ RDS, implement automated backup with 30-day retention, move secrets to Secrets Manager, restrict security groups to minimum required access, deploy WAF in front of the application load balancer. Infrastructure as code (week 6–8): Terraform the entire environment, implement the CI/CD pipeline with security scanning, set up CloudWatch monitoring with PagerDuty alerting, produce audit artefact documentation.