I build the systemsthat keep runningwhen nobody is watching.

Qtec Solution Limited

An AI learning platform for educators: a router that sends every query to whichever model actually fits, and a document pipeline that grounds its answers and generated content in a school's own curriculum.

• A query router reads intent and dispatches to whichever model fits — Gemini Flash for conversation, GPT-4.1 for reasoning, Claude for code, o3-mini for math, Sonar when the answer needs the live web — rather than routing everything through one model.
• Curriculum PDFs and slides go through a Mistral and Pixtral ingestion pipeline that handles text and images alike, get embedded into pgvector, and become the grounding for a RAG layer that answers from the school's own material instead of the model's memory.
• An editor agent locates the exact section of a generated worksheet or lesson plan that needs changing and rewrites only that part, so a small edit costs a few hundred tokens instead of regenerating the document.
• An internal admin panel — a Drive-like store with versioning — is where teachers upload course material, which feeds directly into the router and the generation tools built on top of it.

Fintech platform

A microservices platform serving about 80 live and historical Bitcoin datapoints to trading dashboards, plus news aggregation and a chatbot that answers questions against the same data.

• A Flask gateway fronts the whole platform and owns rate limiting and auth; a Node service pushes price and chain updates over WebSocket so dashboards never poll.
• Reads are absorbed by a Redis caching layer with a pre-warmer that fills hot keys before traffic arrives, which is what keeps the gateway's tail latency flat during market moves.
• Independent retriever services collect from each upstream source and write to MongoDB, so one failing exchange feed never takes the API down.
• Ships as containers to Cloud Run with health-check services per component.

n8n orchestration

An end-to-end pipeline that turns a page of property listings into qualified phone calls: scrape, find the real owner, scrub against the do-not-call registry, then let an AI voice agent make the call.

• A Chrome extension reads listing tables by column header rather than position, so the scrape survives the source site rearranging its layout, and POSTs each batch to an n8n webhook.
• n8n answers 202 immediately and fans the batch out row by row, so a large scrape never times out the browser. Qualification logic lives in the workflow: a listing passes only if its current status is newer than any expired or cancelled date on the record.
• Each row travels through FastAPI services — headless-browser owner lookup, OCR with LLM entity resolution, do-not-call scrubbing, then voice calling with an SMS fallback when nobody picks up.
• Flagged numbers are never auto-dialled but are still written out with full owner detail, so they can be worked by hand. Every row carries one id from scrape to report.

Cloud Run

A small service that lets self-hosted n8n workflows call Vertex AI in an organisation where downloadable service-account keys are blocked by policy.

• Code running inside Google Cloud picks up its service-account identity automatically, so the proxy authenticates without a key file and n8n never holds a Google credential.
• Workflows send an ordinary Gemini request with a shared-secret header; the response comes back untouched, so existing parsing nodes needed no changes.
• Access tokens are cached and refreshed only on expiry rather than re-read from the metadata server per call.
• Scales to zero between runs, which makes an always-available LLM endpoint cost close to nothing.