Osaid Shah

SIV.AG has built software for German utilities for forty years. I work on the phased migration of that legacy monolith to event-driven microservices, with the original still in production.

I delivered two services greenfield to production: a master-data platform consolidating millions of address records across tenants, and the Feed-in MVP for grid registration, solar onboarding and metering. Both required zero-downtime migration and domain modelling from undocumented legacy rules. I defined the architecture standards through ADRs and introduced AI-assisted development across the team.

Formulatrix builds automated liquid-handling instruments for laboratories. I built the instrument control software.

I shipped the React and Redux control interfaces for the aµtoPulse v4 and Reagent Exchanger, integrated real-time camera streaming with the embedded controls, and built an MQTT ingestion pipeline into PostgreSQL that surfaced production telemetry as OEE dashboards.

I built and led delivery of the multi-tenant automotive operations platform behind these marketplaces: CRM, auctions, inspections, lead assignment, ROTA scheduling, reporting and role-based access control, from a single codebase.

Most of the work replaced manual process with workflow automation — QR-based warehouse attendance, rule-based lead routing and scheduled field-sales allocation.

A car-parts sourcing platform in daily production use in Dubai, built on the WhatsApp Cloud API. Customers send photos, video and voice notes; the system transcribes them, resolves vehicle identity through OCR, function calling and NHTSA vPIC lookups, routes RFQs to suppliers and generates quotations.

Python throughout: multimodal and multilingual intake, structured outputs, tiered model routing with escalation on low confidence, and deterministic business rules between stages. The pipeline runs five stages — SEE, EXTRACT, CHECK, DECIDE, BUILD — each independently executable, with human approval gating every outbound message.

10 → 100+

inquiries handled per operator per day

~90%

of inquiries completing without a human handoff

−82%

inference cost per inquiry, after tiered model routing

Coding subscriptions are flat-fee; API tokens are metered, so billing decides the role. Planning and review stay on the subscription you already pay for — Claude Code or Codex — and bounded implementation goes to cheaper models over OpenRouter, chosen per task.

Each worker takes its own git worktree and merges onto a single integration branch. A check is a command that passes or fails, so nothing is done because a model said it was. Profiles bound the delegation tree by width, depth and total, and one engine runs three modes: conversation, a single-task retry loop, or a dependency graph.

A private second brain covering tasks, finances, notes, journaling, books and career records, with agent access over the same data.

FastAPI and PostgreSQL, with SvelteKit and Android clients. Local ETL pipelines normalise WhatsApp, Gmail and chat exports into an append-only event store with searchable memory and a generated chronology. MCP servers expose scoped tools to agents behind confirmation policies, audit logging and Cloudflare Access.