Prove It Works
Fintech lending
Led by Stephen Thorpe
300M+ rows
Self-healing production data
The situation. A consumer and business lending marketplace operating across the US and
Latin America was processing up to 10,000 transactions a day on infrastructure that had outgrown its own
visibility. Data moved from an Oracle transactional database into the systems powering credit and risk
models, but nobody could see a failure until a client did.
What Stephen built. Stephen architected and ran the full data platform: custom CDC-style
loaders streaming Oracle changes into Kafka, feeding Cassandra, feeding Spark ML jobs that served live
inference through FastAPI. On top of that, he built reconciliation and monitoring jobs to catch ingest
failures automatically, and deployed autonomous agents — running on a multi-region, multi-cluster
Kubernetes environment (OCI, US and Brazil) — that read production logs, diagnose the failure, fix it,
and redeploy through the existing CI/CD pipeline (Bitbucket, Argo), all without a human in the loop.
The result. 300M+ rows in production. Real incidents caught and resolved by the agents
live, before clients noticed anything was wrong. A lending platform that could finally see its own
failures before they became someone else's problem.
Prove It Works
Healthcare benefits
Safety-guarded RAG
An assistant enterprises could deploy
The situation. An employee benefits provider wanted to offer its enterprise clients an
AI health assistant — but in a healthcare-adjacent context, "mostly accurate" isn't good enough. Answers
needed to be grounded in specific, vetted content, not general model knowledge, and the system needed
guardrails a generic chatbot doesn't have.
What we built. A custom RAG architecture using LangGraph for orchestration and a
purpose-built Milvus vector database, indexed specifically against the client's curated article library
rather than open-ended retrieval. We designed the guardrail and escalation logic to keep the assistant
inside its evidence base and hand off appropriately when a query fell outside it, then took the system
through to production deployment for the benefits provider's enterprise clients.
The result. A live, white-labeled health assistant that enterprise clients could deploy
to their own employees, with answers traceable back to approved source material — the difference between
a demo and something a regulated-adjacent business could actually stand behind.
Untapped Value
Commercial real estate
2B+ rows
Geospatial lease-price prediction
The situation. Commercial real estate pricing decisions — what a space should lease for,
which properties are underpriced — have historically relied on broker intuition and thin comparables. The
data to do better exists (location, foot traffic, demographics) but almost nobody has the pipeline or the
feature engineering to turn it into a working pricing model.
What we built. An ML platform trained on 2B+ rows, including anonymized mobile location
data, built on a Parquet and DuckDB pipeline for cost-efficient processing at scale. We engineered a set
of proprietary features that don't exist in off-the-shelf CRE tools — a retail gravity score, visitor
demographic profiles (age, income), and travel-radius/trade-area modeling — and layered an interactive
heat-map visualization on top so a broker or investor could see the trade radius, not just a number.
The result. A working lease-price prediction and property-comparison platform, currently
under review with an enterprise client — proof that CRE pricing data most firms already have access to
can be turned into a defensible, visual pricing tool, not just a spreadsheet of comps.
Untapped Value
Heavy-equipment logistics
Live dispatch
ML routing for a working fleet
The situation. An earth-moving and dump truck operator was scheduling and dispatching
largely by manual coordination — a workable approach at small scale, but one that breaks down as fleet
size, third-party leased equipment, and daily schedule changes multiply the coordination problem.
How we mapped it. Before building anything, Stephen interviewed dispatchers, operators
and management to document how jobs, trucks and leased equipment were actually being scheduled day to
day.
What we built. An ML/AI-powered dispatch system that dynamically re-optimizes schedules
and routes as conditions change through the day, built to handle both owned and third-party leased
equipment as first-class participants in the system rather than a manual add-on. The architecture was
deliberately built to support a future two-sided marketplace for leased equipment, not just solve
today's scheduling problem.
The result. A production dispatch system that turns a manual coordination process into
an automated one — and an architecture the client can extend into a leasing marketplace without a
rebuild, because that path was designed in from the start.
Untapped Value
E-commerce collectibles
Shopify at scale
Computer vision valuation
The situation. Brands with a secondary market for their physical goods — collectibles,
limited editions — have no efficient way to price resale value at scale. Manual appraisal doesn't scale
to a marketplace, and generic pricing tools don't understand what makes a specific item more or less
valuable.
What we built. A one-shot learning and CNN-based image recognition system that identifies
a collectible from a photo and estimates its resale value, then publishes that valuation directly back to
the brand's Shopify store — turning a manual appraisal process into an automated pipeline a brand can run
at marketplace scale.
The result. Brands gained a way to capture secondary-market revenue on their own products
automatically, without building any of the underlying computer vision or valuation infrastructure
themselves.