I lead engineering at JewelBench AI, a production platform that turns jewellery photographs and written briefs into manufacturable designs, costed bills of materials and 3D assets for manufacturers and retailers. I own the AI systems, the services they run in and the AWS deployment behind them. MSc Machine Learning, Royal Holloway, University of London.
|
Two to five reference photographs in; a costed bill of materials and a dimensioned spec sheet out. Vision-based analysis with self-consistency sampling and outlier-robust aggregation, paired with a deterministic dimension-based weight estimator. Measured against manufacturer ground truth: about 10% median error from printed dimensions plus finger size (106 records, cross-validated); about 15% from a photograph alone (110 records). In production and billing since September 2026. |
Rebuilt and shipped to every account in September 2026. Targeted image edits, a four-mode design wizard (variant creation, category transfer, sketch recreation, describe the change) capped at ten questions with a create-now option at every step, an automatic fidelity check with one retry, trade vocabulary including Indian terms, and 5,000-character briefs. The Bulk Generator plans and renders variations on the same engine. |
Ground-truth datasets mined from a manufacturer's CAD archive, an A/B accuracy harness that gates releases, and regression suites that lock behaviour in. Audited our own accuracy claims and replaced the three that did not hold with measured figures. Three services on AWS ECS (Java WebFlux API, Python 3D and estimator services), a React front end on CloudFront, rollback revisions, a security review with a risk register, authenticated service callbacks, and incident response. |
Spring WebFlux · AWS ECS · ECR · S3 · CloudFront · SQS
rhino3dm · trimesh · STL/OBJ/GLB
Explore the full stack
I designed the engineering workflow around coding agents for implementation under my review, an independent automated review before merge, and guard hooks that block secrets from sessions. The tooling I built for it is open source:
| Repo | What it does |
|---|---|
| secret-guard | Prompt and tool-call guard hooks that stop secrets leaking into AI coding sessions |
| ai-worktree | Run coding agents in isolated git worktrees with reviewable merge-back |
| cockpit | A live terminal dashboard: roster, token burn, active agents and machine gauges |
| hook-pack | A curated, individually selectable collection of production-quality agent hooks |
| ai-handoff | Hand a task between coding agents without losing context |
| notify-done | Desktop, phone and bell notifications when an agent run finishes or needs input |
| aispend | One per-day, per-tool token and cost table from local agent logs |
Jewellery tools: gia-charts-carat-calc, a live carat-weight calculator with the GIA estimation formulas, and stl-weight-calculator, STL weight for precious metals with an interactive 3D viewer.
| 2025 to present | Machine Learning Engineer, JewelBench AI, Mumbai. Leads engineering for the platform above. |
| 2022 | Business Data Analyst, Shree Laxmi Traders, Mumbai. Tableau and Power BI dashboards for sales and inventory. |
| 2021 | Data Science Intern, Zytex, Silvassa. SQL, ETL between plant data sources, batch-quality analysis. |
| MSc Machine Learning | Royal Holloway, University of London, 2022 to 2024 |
| B.E. Electronics and Telecommunications | St. Francis Institute of Technology, University of Mumbai, 2018 to 2022 |
| Also | UK National Quantum Computing Hackathon 2024 (NQCC), with CGI UK. Football and photography when the laptop is closed. |
