Muhammad Zia · Senior Full Stack AI Engineer

Designed and shipped an agentic AI platform on Cloudflare (Workers, Workers AI, Durable Objects, Vectorize) that processes 4B+ LLM tokens a month. Novacom operations and marketing teams use it from a conversational interface; agents reason over internal data and take write actions across workflows, records, and downstream processes. Also shipped a React Chrome extension with a RAG pipeline over Product Data Sheets, and a Claude-powered ingestion pipeline that parses operations spreadsheets into structured fields, saving 2+ hours of manual work a day.

Least-privilege by design: agents hold no stored credentials, act only through the user's own authenticated session, and every write needs explicit human approval. Evals live in CI: the platform is split into independently testable modules (MDSD), so each pull request runs evals only for what it changes against a local Worker, and nightly and weekly LLM-as-judge runs against recorded response cassettes catch model drift against the previous baseline. Judge scores held above 87%; eval spend cut ~87%, from ~$30 to ~$4 a day.

Architect and ship core internal React applications at Novacom, including a schematics platform used by 200+ teams worldwide to model complex industrial diagrams: refineries, manufacturing plants, protection systems, tugboats, and more.

Owned end to end from system design through deployment, with scalability, clean abstractions, and long-term maintainability as the constraint.

Shipped production LLM and agentic features across Convert and Instantly: an MCP (Model Context Protocol) server exposing Convert's experimentation data and actions to LLM clients as first-class tools (query experiments, read results, trigger actions), an AI email editor at Instantly that detects spam-trigger language in real time and highlights it inline, and a natural-language lead search that replaces rigid filter UIs with vector search.

Brought agentic workflows to an enterprise experimentation platform via Anthropic's open MCP standard. At Instantly, directly improved outbound inbox deliverability and let users find leads by describing an ideal customer profile in plain English.

At Convert, owned the Shopify App end to end: the app merchants install from the Shopify App Store to run A/B tests on their storefronts, including the OAuth install flow, embedded admin UI, storefront script injection, and the compliance webhooks required for App Store approval. Led the frontend migration from Ractive.js to React, shipped Multi-Armed Bandit optimisation, and was a core contributor to the Visual Editor every customer uses to build A/B tests without engineering.

Multi-Armed Bandit work contributed to conversion lifts of up to 34% for enterprise clients. Kept backend schema integrity and cut technical debt through proactive refactoring. Full-time March 2021 to February 2024, then part-time contractor through April 2025.

At Instantly, architected a three-layer error-handling system (React Error Boundaries, Axios interceptors, and global runtime handlers) and built bulk tagging across campaigns, accounts, and emails.

99.8% failure capture and a 47% drop in support tickets. Bulk tagging was a major productivity win for high-volume users processing 10M+ records.

Led a full frontend redesign and design system at Pocketnote, an EdTech platform matching tutors with students, and upgraded the CI/CD pipeline to Heroku.

3x developer velocity on feature delivery. Deployments 68% faster and cloud operating cost down $2.4K/month.

Migrated the front-end from Vue.js to React.js at Cobija Industries, improving maintainability and performance.

Improved UI responsiveness and optimized backend interactions. Led a remote team of developers, providing technical direction and code reviews. Delivered stable, well-tested features.

Most MCP servers ship demo-grade: no per-tool auth policies, no write confirmation gates, no audit trail, and no PII redaction on tool results. Those are exactly the gaps that show up the moment you expose a production SaaS to LLM clients.

Published Express-style middleware for MCP servers: glob policies (allow / deny / requireConfirmation), token-bucket rate limits, JSONL audit logs with hashed args, and result redaction. Wrap an existing server in a handful of lines.

RAG pipelines usually ship on vibes: swap an embedder, glance at a chat demo, and miss retrieval regressions until users ask for a product code the vector index never hits.

Shipped a CLI + library that evaluates any pipeline against YAML golden questions: hit-rate@k, MRR, answer-contains, LLM-as-judge faithfulness, latency percentiles, and CI gates that fail on regressions.

React Error Boundaries catch render crashes; streaming LLM UIs fail differently: stalls, mid-stream drops, and tool-call errors that would otherwise wipe partial tokens the user already saw.

Published a headless React library with useGuardedStream and StreamBoundary: typed stream failure classes, partial-preserving fallbacks, exponential retry, and optional resume-from-checkpoint.

In production React apps, a single render error can take down the whole surface. Teams often rely on one generic boundary or none at all, and native boundaries never catch async failures, so users see a white screen and you lose usable error signal.

Published a layered pattern (global, route, and feature boundaries) with async-to-boundary bridging, error classification, recovery hooks, and pluggable reporters (including Sentry) so crashes stay contained, recoverable, and visible in observability tools.

LLM products accept arbitrary user text; without a gate, prompt injection, jailbreaks, and data-exfiltration phrases can reach the model and override instructions or leak context. Naive blocklists create false positives and miss obfuscated payloads.

Shipped a zero-dependency, TypeScript-first guard with weighted scoring across dozens of rules, verify/strip/analyze APIs, Express and Next.js middleware, a React hook for chat UIs, and an optional Claude adapter for a second verification pass before tokens hit your provider.

Tool-first design, auth for LLM clients, idempotent writes, and the schema/prompt failures that show up once MCP hits production.

The safety ladder from read-only to scoped autonomous writes, RBAC, audit trails, confirmation tokens, and a near-miss that forced the design.

Chunking for datasheets vs prose, hybrid search for product codes, golden-question eval before you touch a vector store, and extension latency budgets.

Cost, latency, and quality from a real ingestion pipeline, plus a confidence-gated routing pattern between edge models and frontier APIs.

A deep dive into architecting production-grade error boundaries that capture 99.8% of failures, refreshed for streaming LLM UIs.

How proper architectural decisions early on can 3x developer velocity and prevent technical debt.