Own performance, reliability, and ML-pipeline integration of the CMS that drives 3D-scan uploads through panorama stitching, sky replacement, depth-map generation, room-layout estimation, and virtual-tour delivery. ~90 merged MRs across backend, infra, and a cross-repo Tauri desktop tool. Main touchpoint between R&D and production.

• Brought Rust into a mature Python codebase. Shipped the first PyO3 / maturin extension for O(1) graph reads on large projects, then a critical-path ETA engine that walks the DAG of unfinished jobs using p50 duration percentiles, backed by an r2d2 Postgres pool that releases the GIL.
• Cut hot-path endpoint latency by 30–800×. work_started on 1,000 assets: 15.5 s → 19 ms (~815×). mark_artifacts_reviewed on 500 stills: 215 s → 6.9 s (~31×). mark_unreviewed for panos: 13.9 s → 0.08 s (~165×). Upload endpoint: 205 s → 128 s, with the rest pushed to async Celery.
• Indexed a 10M-row webhook table with zero downtime — partial B-tree on STATE_INITIAL, GIN jsonb_path_ops on payload, composite (state, created) — all shipped via non-atomic CREATE INDEX CONCURRENTLY migrations. Eliminated the sequential scans that were hammering deployment-resume and daily project-filter queries.
• Integrated 12+ ML processing clients into production: pano enhancement, sky replacement v2, monodepth, POI detection, pillar detection, house-number blur, image embeddings, TV replacement, and more. Each integration touches the processing-job schema, the graph engine, and the k8s deployment.
• Built the shadow-deployment / safe-rollout system. Quota-based experimental-features mechanism the team uses to ramp new ML models without risking the whole pipeline. Feature flags and dynamic settings govern 70+ namespaces and ~200 k8s artifact actions.
• Coplanner (workforce assignment). Real-time online tracking, role-based assignment permissions, and a dual-queue design (standard + student) backed by Redis Z-Sets with periodic rebuild and atomic pop.
• Platform-wide perf wins. stdlib json → orjson; toggleable global GZip via dynamic settings; rewrote a RAM-leaking metrics job from mocked zeros to efficient DB aggregation. Try to leave every endpoint in better shape for the next person.

12+ backend and automation contracts with a 100% Job Success Score. Embedded in client teams or delivered solo — the long-standing relationships come from clear communication and not surprising people.

Core engineering on TDN, Strange Logic's SaaS product. Streamlined L2/L3 support for the PHP application (500+ tickets resolved) and led the Stripe migration. On the second engagement, pushed for ClickHouse as a new back-end database and moved core APIs from PHP to Python. Kicked off the Expired Domain Search and Mass VAT Checker initiatives — both started as hallway conversations that grew into products.

Built a Python pipeline that scraped listings from the three largest UK real-estate directories, and designed the schema that held up to 4.3 TB of listings, images, and historical sales records.