The work · By company, documented and ranked
The deep dives below were written against the actual code: module inventories, real routes, git history. Where something was a mockup, synthetic, or unvalidated, the page says so. Client names, countries, and figures are out. My engineering results are in.
Client work
Strategy&
· PwC
Government and enterprise engagements, anonymized under the rules on each page. Ranked deliberately: a portfolio that presents everything as equally impressive is less credible than one that ranks its own work.
Highest in portfolio - ~17k LOC, 6-stage ingestion, evidence-gated scoring, validated LLM-judge evaluation
High - closed-loop analytics, statistical anomaly detection, report generation, RBAC
High - multi-stage NLP pipeline, 4-tier LLM funnel, adaptive vector clustering, MinHash dedup, bilingual entity resolution
Medium-high - three-tier executive information architecture, ~20 bespoke chart components, dense comparative visualisation
Medium-high - schema inference from unstructured workbooks, real tool-calling agent loop, privacy-constrained handling
Medium - bilingual RTL data model, many-to-many objective/initiative mapping, generated data layer
Medium-high - offline-first GIS, packaged vector basemap, real geospatial analysis, test-enforced network invariant
Medium - vendor-delivered platform; my contribution was quality assurance and safety review
Medium-high - OCR and native PDF handling, table extraction, field mapping, accuracy benchmarking
High - multi-agent orchestration, MCP tool integration, always-on daemon, native macOS app
Medium - repeated and hardened across four engagements
Independent builds
My own
products
Built on my time, run on my money.
A system that maps what is about to trend: niche sources scanned, engagement clustered, ideas turned into scripts.
A full webinar launch for my education brand: AI wrote the scripts for three video ads, built the slides, the funnel, and the landing page, and ran Meta campaigns through a connector I built before an official one existed, monitoring every 12 to 24 hours. Cost per lead: 17 cents. 2,000+ registrations. Around 1,700 replay views on YouTube. Two to three weeks, next to a full-time job, first attempt.
A Telegram assessment chatbot with RAG and webhook shipment tracking (UK client). A real-time avatar demo built in days, deployed near the APIs for latency; the prospect called it magic and hired someone else. Both true.
Earlier · 2022 to 2025
Artefact
· Innovum
The data-science years: NLP topic classifiers and sentiment models for multinational clients, geographic clustering for charger infrastructure, transcription and summarization pipelines. Where the discipline of measuring before claiming was learned.
The first step
You leave the call knowing whether I can help, roughly what it would take, and what it would cost to find out for sure. If I am not the right person, I will say so on the call.
The 20 minutes, in order
Before you book