CASE FILE № 016 — Typework

CLIENT OR PRODUCT

Typework is the AI business operations platform by Holomix. I began contributing in November 2025, working across backend, web, sandbox services, and Flutter alongside customer solutions and scenario delivery. My Holomix employment was July–October 2026; this case also covers earlier project contributions.

MY ROLE

Presales engineering · Full-stack development · FDE delivery, spanning solution design, platform implementation, and customer-scenario validation

THE PROBLEM

Long business processes depended on agent interpretation and execution, making outcomes inconsistent and difficult to inspect. Customers also needed reusable capabilities, integration with existing systems, web and mobile access, and tasks triggered by schedules or external events.

CONSTRAINTS

• •Deterministic execution, confirmation steps, and recovery across long processes
• •Reusing skills and workflows across agents, chat, pages, and scheduled tasks
• •Recovering guest uploads and asynchronous builds across sign-in and account changes
• •Tenant boundaries, external-application authorization, and communications-account isolation

THE INVESTIGATION

Mapped customer processes into inputs, actions, confirmations, and results, organizing suitable long-running tasks into agent-callable workflows. Used source code, run records, and multi-industry user journeys to check whether generated resources supported actual business operations, while addressing cross-platform and account-state handoffs.

THE STRATEGY

Build on Python/Django, reuse capabilities through skills, execute business steps with versioned workflows, and schedule tasks with Celery. Generate applications in Daytona sandboxes, linking pages and resources through manifests and bindings. Expose capabilities through web, Flutter, and External Application APIs with tenant, role, and authorization controls.

WHAT I BUILT

• •Platform foundation: designed the early architecture and core features using Python/Django, Celery, and Daytona. Developed scheduling and skills, and extended multi-provider LLM integration, RAG, question banks, and learning progress.
• •Workflow engine: implemented constrained execution steps, immutable versions, validation, dry runs, idempotent execution, interactive waits, and run auditing, with role/exposure controls and confirmation for high-risk agent actions.
• •Workflow product experience: connected agent routing, chat references, pages, and scheduled tasks. Delivered Workflow Center business/developer views, publish reviews, execution records, input-source metadata, and portable import/export.
• •AI onboarding: uploads, industry and blank-project paths, module selection, plan review, and build progress. Implemented module composition and isolated sandbox generation, with sign-in recovery and protection against stale asynchronous updates.
• •Pages V3: implemented the tenant-level backend, manifests and resource bindings, versioned R2 artifacts, source import/export, and sandbox builds. Added web dashboards, previews, and route editing while retaining legacy pages and role-visibility controls.
• •Engineering reliability: enforced tenant ownership of sandbox sessions, messages, and files; hardened web account changes, drafts, and debounced saves. Delivered agent debugging, knowledge/PDF previews, bilingual UI, regression coverage, and deployment configuration.
• •Communications integration: built Twilio voice/SMS, inbound/outbound signed webhooks, opt-outs, quotas, and scheduled actions. Added tenant subaccounts and credential lifecycle management to connect communications with business processes.
• •Enterprise integration: developed External Application APIs for document ingestion, member synchronization, chat sessions, and authorized workflow invocation, with application capabilities, tenant bindings, and request idempotency.
• •Flutter client: delivered Pages V3 discovery, categorization, WebView access, and reusable page authentication. Replaced persistent list WebViews with version-aware disk-cached thumbnails and added client regression tests.

THE HARD PART

Connecting flexible agent interpretation to inspectable business execution while preserving version, authorization, and idempotency boundaries. Maintaining consistent state across backend, sandboxes, web, and Flutter so delayed responses, old drafts, and saves cannot affect the wrong account.

THE VERDICT

• •Connected agents, skills, versioned workflows, and scheduling into reusable business automation with inspectable steps, explicit confirmations, and execution records.
• •Linked Pages V3 backend resources, sandbox builds, web editing, and Flutter access into an application delivery path for different customer scenarios on a shared platform.
• •Added Twilio and External Application integrations, connecting communications actions, enterprise documents, members, and authorized third-party workflow calls to platform processes.
• •Expanded regression coverage for tenancy, account changes, recovery, and cross-platform behavior, and used multi-industry user journeys to check generated artifacts against actual business execution.

TECHNICAL EVIDENCE

Django · Next.js · Python · React · TypeScript · Daytona · Celery · Cloudflare R2 · Flutter · Twilio · Zustand · Workflow automation · Vitest · GitHub Actions · i18n