AI & agents
Almost every system here has the same shape. Work out what can be decided by rule, decide it by rule, and hand the language model only the part that is genuinely about language.
That is not caution for its own sake. A model that does arithmetic produces confident wrong numbers, a model that resolves a patient or a product identity produces confident wrong matches, and neither is auditable after the fact. Scoping it narrowly is what makes the rest of the system reviewable.
Azlytics
Profit analytics for direct-to-consumer Shopify brands. Revenue is not what lands in the bank, so the whole dashboard is built on contribution margin instead.
WhatsApp order extraction
Turning free-form WhatsApp haggling into matched orders against a hand-kept catalogue with no clean keys, for an oil-blend vendor.
Tasks & calendar MCP server
One endpoint that lets an assistant read and write Google Tasks and Calendar, hardened against a specific client's probing behaviour.
Multi-provider LLM router
An OpenAI-compatible proxy over several providers that rotates a pool of keys and keeps working when the coordination layer does not.
Serverless file store
A two-way file store an AI assistant and a phone browser can both use, holding no server state at all.
Lead enrichment pipeline
A long-running enrichment and scoring pipeline that survives the third-party rate limits and timeouts it is built on.