01
Data
Find the signal nobody has wired up yet.
Vessel movements, Shopify order streams, POS tickets, HR records, drawings and BoQ documents. Ingestion is scrapers, APIs, webhooks and file parsers — with verification and confidence levels attached at the point of capture, not bolted on later.
- Python
- REST / webhooks
- Scrapers
- Postgres
- MongoDB
02
Processing
Make it trustworthy before it is clever.
Cleaning, entity resolution, deduplication and feature engineering. This is where confidence scoring, unit normalisation and leak-free train/serve parity get decided — most model failures I have debugged were really failures here.
- Pandas
- SQL
- Feature pipelines
- DuckDB
- Validation
03
Machine Learning
Predict the number someone will act on.
Gradient boosting for demand and churn, clustering for segmentation, regression for cost and CLV. Forecasts ship with intervals, because a planner needs the downside, not a single confident number.
- LightGBM
- scikit-learn
- XGBoost
- Probabilistic forecasting
04
LLM Reasoning
Language where rules are too brittle.
Retrieval over the business's own documents, structured extraction, tool calling and explanation. The model reasons; it does not invent facts — grounding and citation are part of the architecture, and evaluation runs on real examples.
- RAG
- pgvector
- LangChain
- Tool calling
- Structured output
05
Automation
Close the loop without a human in the middle.
Agents that act: send the outreach, recover the abandoned order, raise the reorder, book the appointment, escalate when confidence drops. Every action is logged, reversible and rate-limited.
- Agent orchestration
- Queues
- Redis
- WhatsApp / Twilio
- CRON
06
Real-World Application
A product someone opens on Monday morning.
Interfaces that make the intelligence usable: dashboards, queues, chat, voice, reports. Deployed in Docker on GCP or AWS with observability and cost ceilings, because an AI feature that quietly burns tokens is a liability.
- Next.js
- FastAPI
- Django
- Docker
- GCP / AWS
01
Data
Find the signal nobody has wired up yet.
Vessel movements, Shopify order streams, POS tickets, HR records, drawings and BoQ documents. Ingestion is scrapers, APIs, webhooks and file parsers — with verification and confidence levels attached at the point of capture, not bolted on later.
- Python
- REST / webhooks
- Scrapers
- Postgres
- MongoDB
02
Processing
Make it trustworthy before it is clever.
Cleaning, entity resolution, deduplication and feature engineering. This is where confidence scoring, unit normalisation and leak-free train/serve parity get decided — most model failures I have debugged were really failures here.
- Pandas
- SQL
- Feature pipelines
- DuckDB
- Validation
03
Machine Learning
Predict the number someone will act on.
Gradient boosting for demand and churn, clustering for segmentation, regression for cost and CLV. Forecasts ship with intervals, because a planner needs the downside, not a single confident number.
- LightGBM
- scikit-learn
- XGBoost
- Probabilistic forecasting
04
LLM Reasoning
Language where rules are too brittle.
Retrieval over the business's own documents, structured extraction, tool calling and explanation. The model reasons; it does not invent facts — grounding and citation are part of the architecture, and evaluation runs on real examples.
- RAG
- pgvector
- LangChain
- Tool calling
- Structured output
05
Automation
Close the loop without a human in the middle.
Agents that act: send the outreach, recover the abandoned order, raise the reorder, book the appointment, escalate when confidence drops. Every action is logged, reversible and rate-limited.
- Agent orchestration
- Queues
- Redis
- WhatsApp / Twilio
- CRON
06
Real-World Application
A product someone opens on Monday morning.
Interfaces that make the intelligence usable: dashboards, queues, chat, voice, reports. Deployed in Docker on GCP or AWS with observability and cost ceilings, because an AI feature that quietly burns tokens is a liability.
- Next.js
- FastAPI
- Django
- Docker
- GCP / AWS