Address-directory crawling estate
2020–2021
Crawling and address-extraction system that builds a directory of the physical locations of internet businesses from public web pages.
Engineering
Java
I built web-scale crawling and address-extraction systems in Java.
See its threads in the worldUsed on
Payment-proof review for a P2P crypto marketplace
2023–2024
A service that classifies payment-proof images on a peer-to-peer bitcoin marketplace and returns the key text line by line.
Engineering recruitment pipeline
2021–2026
Internal engineering hiring pipeline at Smarter.Codes, run as one thread per open position with async tests before live interviews.
NLX entity-extraction API
2022–2023
Natural-language API that returns entities, facts and a parse tree from domain-specific text such as ingredient declarations and medical supply descriptions.
Medical-device monitoring for hospitals
2019–2021
Monitoring and remote-access system for hospital medical devices: live dashboards and alerts for staff, browser-based remote access for service engineers.
Kray store rollouts
2023–2026
Work that puts Kray, an AI e-commerce search product, into real stores: a store cloner, demo store and grocery rollouts.
Kray.AI search and document assistant
2021–2024
An AI search and document-assistant platform sold as Kray.AI, with a WooCommerce plugin and a demo store.
Company-location extraction at web scale
2021–2023
A web-scale crawler that finds a company's office address from its hostname, built for an external client.
Stories where it mattered
2024 · Kray.AI search and document assistant
One model for everyone, or one per customer
- Situation
- A document assistant for a fleet operator needed an architecture, and one model trained on every vessel's documents was the assumed default.
- Task
- Lay out the real trade-offs of cost, latency and access control instead of just picking one.
- Action
- I wrote both options out for client and team: one shared model versus one model per vessel, with pros, cons and the replica arithmetic.
- Result
- The partitioned design went forward, with the loss of cross-fleet questions accepted openly.
2026 · Engineering recruitment pipeline
A hiring process that does not depend on me being in the room
- Situation
- Technical hiring for a research engineering role ran through me, a bottleneck and a single point of judgement.
- Task
- Build an interview panel that keeps the bar without depending on me.
- Action
- I proposed one lead and two shadows, shared candidate review beforehand, independent written feedback before discussion, and shadows who graduate into leads.
- Result
- The format became how the round runs, and the shadow-to-lead path grew the pool of interviewers.
2026 · Engineering recruitment pipeline
A one out of five, written so the candidate could act on it
- Situation
- A research engineering candidate reached a working coding-round solution with AI assistance but could not explain why the Python worked.
- Task
- Score the round and write feedback that the candidate, recruiter and panel could all act on.
- Action
- I scored it low and wrote out the reasoning behind each mark, recommended not proceeding, said what to strengthen, and gave a reapplication window.
- Result
- The rejection was reviewable rather than a bare verdict, and the panel's differing views were visible to decision makers.
Goes with
Maven
Python
Distributed Systems
Docker
Architecture Decision RecordsThrow me a paper plane
Write a line; I’ll find it in my inbox.