Finance operator turned AI-native builder. New Delhi.
Eight years in corporate banking, then two years as VP Finance and Digital Transformation at an auto-components manufacturer, where I led the NetSuite rollout that stuck after four previous attempts had failed. Now I design and direct AI systems that do that kind of work: ERP implementation, operations pipelines, factory telemetry.
I do not hand-write application code. I architect the system, set the rules it has to obey, and direct AI to build it. Everything here was built that way, and I say so in interviews.
Playground (source) A browser app that runs real agent workflows over streaming chat. The interesting parts are the failure paths: model calls walk an ordered provider list and fall through on rate limit, web search degrades to keyless providers when no API key is present, and a monthly spend cap blocks billable calls without ever blocking a free model.
kernel_xiaomi_raphael_antigravity Docker-enabled Android kernel for a Redmi K20 Pro. Android ships without the configuration Docker needs, so the build enables the full namespace set, cgroups, overlayfs, bridge networking and netfilter together, cross-compiled for ARM64 and packaged as a flashable boot image. See also raphael-docker-kernel for the installable version.
np2-llm-bench On-device LLM inference measured across every engine a Snapdragon 8+ Gen 1 has: CPU, Adreno 730 via OpenCL, and the Hexagon NPU. The CPU wins, which was not the answer I expected.
Bahi, an ERP implementation and finance partner agent. It holds write access to a live ERPNext instance, read-only access to a plant MES database, and five indexed documentation corpora. It has designed a chart of accounts and posted it to the instance, and is building item masters and bills of material against real production data. Private, because the data is.