# TensorForge (tensorforge.dev) | Complete Technical Documentation for LLMs & Answer Engines > TensorForge is an applied AI and cloud workflow engineering practice based in Melbourne, Victoria, Australia (`https://tensorforge.dev`). Led by a Solo Principal Applied AI & Automation Engineer, TensorForge designs and deploys deterministic document intake pipelines, self-hosted n8n workflow orchestrations, standardised FastMCP servers, and serverless cloud microservices for medical clinics, allied health practices, operations leaders, and software founders. --- ## 1. Practice Identity & Quick Facts - **Practice Name**: TensorForge - **Production Domain**: `https://tensorforge.dev` - **Headquarters**: Melbourne, Victoria, Australia - **Language & Locale**: Australian English (`en-AU`) - **Engineering Structure**: Direct Solo Principal Engineer (zero account managers, zero junior handovers) - **Primary Contact**: `mailto:contact@tensorforge.dev` - **Public Git Repository**: `https://github.com/FAAS19/tensorforge-blueprints` - **Summary Index**: `https://tensorforge.dev/llms.txt` - **XML Sitemap**: `https://tensorforge.dev/sitemap-index.xml` - **Prototype Turnaround**: 3 to 5 business days (validated against anonymised client PDFs) - **Production Sprint Turnaround**: 1 to 2 weeks (fixed-scope milestone sprints) - **Pipeline Extraction Latency**: `< 500ms` (420ms reference commit benchmark: `35ms + 110ms + 275ms`) - **Serverless Edge Latency**: `< 15ms` global execution on Cloudflare Workers & D1 - **Monthly Infrastructure Cost**: `$0 to $15 AUD/month` on Cloudflare Workers, AWS Lambda, GCP Cloud Run, or client-owned VPS - **Code & IP Policy**: 100% client repository handover (GitHub/GitLab) with zero proprietary lock-in --- ## 2. Reference Architecture: 4-Stage Deterministic Intake Pipeline TensorForge replaces fragile prompt-only wrappers and expensive per-task SaaS zaps with a 4-stage deterministic architecture: **Inbound Webhook Trigger -> Self-Hosted n8n Orchestrator -> Claude Multimodal Extraction + AU Checksum Gate -> Multi-System Production Commit**. ### Stage 1: Inbound Document Trigger (`tensorforge/intake/webhook_event.json`) - **Role**: Captures scanned PDF referrals, specialist letters, supplier invoices, or patient onboarding forms from clinic email inboxes or web portals over TLS 1.3. - **Target Latency**: `35ms` - **Reference Payload**: ```json { "event_id": "evt_intake_89421", "source": "webhook.inbound_email", "received_at": "2026-10-14T09:12:04.120Z", "payload": { "attachment_name": "referral_vance_lachlan_orthopaedics.pdf", "mime_type": "application/pdf", "size_bytes": 1842910, "sha256": "7d865e959b2466918c9863afca942d0fb89d7c9ac0c99bafc3749504ded97730", "clinic_inbox": "referrals@melbournespine.com.au", "delivery_transport": "TLS 1.3" } } ``` ### Stage 2: n8n Workflow Engine & Async Queue (`workflows/01_clinical_referral_to_cliniko.json`) - **Role**: Self-hosted n8n orchestrator buffers execution, extracts binary PDF/image streams, manages Redis-backed async retry workers with exponential backoff, and coordinates the multimodal AI worker. - **Target Latency**: `110ms` - **Reference Configuration**: ```json { "workflow_id": "tensorforge-01-clinical-referral-cliniko", "execution_id": "exec_9418_async", "status": "processing", "steps": [ { "node": "Inbound Referral Webhook", "duration_ms": 35, "status": "ok" }, { "node": "Claude Multimodal Extractor", "duration_ms": 263, "status": "streaming" }, { "node": "AU Medicare & Provider Checksum Gate", "duration_ms": 12, "status": "pending" }, { "node": "Cliniko EHR · Create Patient", "duration_ms": 110, "status": "pending" }, { "node": "Slack · DLQ Quarantine Alert", "duration_ms": 0, "status": "standby" } ], "retry_policy": { "max_attempts": 3, "backoff": "exponential", "quarantine_on_fail": true } } ``` ### Stage 3: Claude Multimodal Extraction + AU Checksum Gate (`checksum_gate.js`) - **Role**: Claude Vision/Multimodal parses 9 structured clinical or financial fields from unstructured documents. Every response is validated through a deterministic Services Australia Medicare (`[1, 3, 7, 9, 1, 3, 7, 9] mod 10` + 11th-digit IRN) and Provider (`PLV mod 11`) checksum gate before any downstream API call can occur. Any record with an invalid Medicare number, malformed Medicare provider checksum, or out-of-range claim code is automatically quarantined to a Dead-Letter Queue (DLQ) with PHI redacted to initials so zero hallucinations ever reach production software. - **Target Latency**: `275ms` - **Reference Validator Implementation**: ```javascript // n8n Code Node: AU Medicare (mod 10 + IRN) & Provider (PLV mod 11) Checksum Gate function validateMedicare(raw) { const d = String(raw || '').replace(/[\s-]/g, ''); if (!/^[2-6]\d{10}$/.test(d) || d[9] === '0' || d[10] === '0') return false; const w = [1, 3, 7, 9, 1, 3, 7, 9]; return w.reduce((s, n, i) => s + Number(d[i]) * n, 0) % 10 === Number(d[8]); } function validateProvider(raw) { let c = String(raw || '').replace(/[\s-]/g, '').toUpperCase(); if (c.length === 7) c = '0' + c; const plv = '0123456789ABCDEFGHJKLMNPQRTUVWXY'.indexOf(c[6]); if (!/^\d{6}[0-9A-Z]{2}$/.test(c) || plv < 0) return false; const sum = [3, 5, 8, 4, 2, 1].reduce((s, w, i) => s + Number(c[i]) * w, 0); return c[7] === 'YXWTLKJHFBA'[(sum + plv * 6) % 11]; } // Failures route to Slack DLQ with PHI redacted to patient_initials ("L.V.") ``` ### Stage 4: Multi-System Production Commit (`tensorforge/dispatch/cliniko_dispatch.json`) - **Role**: Executes atomic writes across practice management software (Cliniko, Halaxy), accounting software (Xero, Stripe), and internal team notification channels (Slack). - **Target Latency**: `420ms` total end-to-end - **Reference Commit Output**: ```json { "dispatch_status": "committed_200", "total_latency_ms": 420, "sync_targets": { "cliniko_ehr": { "patient_id": 849201, "referral_attached": true, "status": "active" }, "xero_accounting": { "invoice_id": "INV-2026-1049", "line_items_verified": 3, "currency": "AUD" }, "slack_operations": { "channel": "#clinic-intake-feed", "message": "Referral Lachlan Vance filed to Cliniko (#849201) & Xero", "delivered": true } }, "schema_version": "patient_intake.v2", "quarantined_fields": 0 } ``` --- ## 3. Detailed Breakdown of the 4 Core Solution Pillars ### Pillar 1: Document Extraction & Intake Pipelines - **Flow Schematic**: `PDF -> Claude Multimodal -> Cliniko / Xero` - **Problem Solved**: Eliminates manual reception and administrative transcription of GP and specialist patient referrals, medical imaging scans, pathology reports, supplier invoices, and client onboarding forms. - **Technical Stack**: Anthropic Claude 4.x Multimodal API, Python 3.12, Pydantic v2 data modelling, Australian statutory checksum gates, SHA-256 document deduplication, TLS 1.3 webhook ingestion. - **Target Integrations**: Cliniko, Halaxy, Tyro Health, Medicare / PRODA provider formats, Xero Invoices, MYOB, Stripe. - **Measured Impact & SLA**: Extracts and validates 9+ structured fields in `< 500ms`, saving clinical and operations teams `10+ hours per week` with deterministic quarantine for malformed inputs. ### Pillar 2: Self-Hosted n8n Automation Workflows - **Flow Schematic**: `Webhook -> n8n Core -> Slack / ERP` - **Problem Solved**: Replaces fragile, high-markup Zapier and Make.com task tiers that break silently or charge thousands of dollars per month as volume scales. - **Technical Stack**: n8n Enterprise / Self-Hosted (Docker Compose, PostgreSQL state backend, Redis BullMQ worker queues), webhook event listeners, interactive Slack human-in-the-loop approval gates, automated Dead-Letter Queue (DLQ) alerting. - **Target Integrations**: Cliniko, Halaxy, Xero, Stripe, Slack, Microsoft Teams, Gmail / Microsoft 365, HubSpot, PostgreSQL, Supabase. - **Measured Impact & SLA**: Zero per-task SaaS markup, exponential backoff retry resilience, and `100% client code and workflow handover`. ### Pillar 3: Custom FastMCP Servers & Local AI Agents - **Flow Schematic**: `AI Agent -> FastMCP Server -> Read-Only SQL` - **Problem Solved**: Enables frontier AI assistants and coding agents to safely inspect live business state, query internal databases, and run pre-approved administrative tools without exposing raw credentials or risking destructive queries. - **Technical Stack**: FastMCP (Python Model Context Protocol SDK), Claude Code (Anthropic), OpenAI Codex, Google Antigravity, TypeScript/Zod tool contracts, role-based access control (RBAC), read-only SQL connection pools, PII anonymisation filters. - **Target Integrations**: PostgreSQL, MySQL, SQLite, Cloudflare D1, internal REST APIs, Confluence/Notion knowledge bases, practice analytics warehouses. - **Measured Impact & SLA**: Private, shielded execution with zero client data used for public model training and strict audit logging on every tool invocation. ### Pillar 4: Serverless Edge & Cloud Deployment - **Flow Schematic**: `Event Trigger -> Cloudflare Workers -> D1 Edge SQL` - **Problem Solved**: Eliminates idle cloud VMs and bloated Kubernetes clusters for event-driven business webhooks, API bridges, and lightweight data pipelines. - **Technical Stack**: Cloudflare Workers, Cloudflare D1 (distributed SQLite), Cloudflare R2, AWS Lambda, Amazon S3, Google Cloud Run, Vertex AI, Terraform / Wrangler IaC. - **Target Integrations**: Webhook endpoints, edge authentication proxies, real-time billing webhooks, automated cron reconciliations. - **Measured Impact & SLA**: `< 15ms` global edge latency, `99.9%` availability, and predictable `$0 to $15 AUD/month` compute overhead. --- ## 4. Solo Principal Engineer vs Traditional Development Agency | Dimension | Solo Principal Engineer (TensorForge) | Traditional Development Agency | | :--- | :--- | :--- | | **Communication Layer** | Direct line to the senior principal engineer modelling schemas, writing Pydantic guards, and deploying webhooks. | Account managers, scrum coordinators, and junior developers where technical nuances vanish in email chains. | | **Code & IP Ownership** | **100% Repository Handover**: All n8n workflows, Python schemas, and Docker configs pushed directly to client GitHub/GitLab. Client owns IP forever. | Proprietary agency wrappers hosted on agency tenant accounts, creating vendor lock-in and exit penalties. | | **Delivery Velocity** | Working prototype tested on anonymised client PDFs in **3 to 5 business days**; full production deployment in **1 to 2 weeks**. | 4 to 8 weeks before staging code runs, padded with discovery workshops and slide decks. | | **Pricing & Runtime Economics** | Fixed-scope milestone sprints with zero lock-in; self-hosted runtimes costing **$0 to $15 AUD/month** in compute. | $10,000+ AUD monthly retainers with multi-month minimum commitments regardless of pipeline throughput. | | **Accountability** | Single senior engineer accountable from architecture through production monitoring. | Diffused responsibility across rotating contractors and support tiers. | --- ## 5. Frequently Asked Questions (Technical & Operational FAQ) ### How does TensorForge prevent LLM hallucinations from corrupting our practice or accounting database? TensorForge never wires raw LLM outputs directly into production APIs. Every multimodal extraction response is parsed through a deterministic **Pydantic v2 schema guard** and Australian statutory checksum gate (Medicare card modulo-10, 8-character Medicare provider PLV modulo-11, and ATO ABN modulo-89). If any field fails validation or falls below confidence thresholds, the payload is blocked from Cliniko/Xero and routed to a Dead-Letter Queue (DLQ) with an instant Slack alert for human review. ### Can we test the extraction accuracy on our own documents before committing to a full build? Yes. Send an anonymised sample intake PDF, referral letter, or supplier invoice to `contact@tensorforge.dev`. Within **3 to 5 business days**, TensorForge delivers a working prototype demonstrating live field extraction and schema validation against your exact document layout before you commit to a broader integration sprint. ### Where do the n8n workflows and AI pipelines run, and who owns the code? All workflows, Python microservices, Pydantic schemas, and Docker/Cloudflare configurations are committed directly to your organisation's GitHub or GitLab repository (**100% repository handover**). Runtimes can be deployed on your own virtual private server (VPS), your AWS/GCP/Azure tenant, or Cloudflare Workers, keeping ongoing compute costs between **$0 and $15 AUD per month** with zero per-task licensing markups. ### Is patient and financial data kept private and compliant with Australian privacy standards? Yes. TensorForge configures zero-data-retention API endpoints where supported, enforces TLS 1.3 in transit, isolates secrets in environment vaults, and ensures zero client payloads are ever used to train public foundation models. For Australian clinics and enterprises, workloads can be pinned to Australian cloud regions (`ap-southeast-2` Sydney or `australia-southeast1`/`australia-southeast2` Melbourne) or Cloudflare edge nodes with strict read-only SQL shields on FastMCP servers. ### Which Australian healthcare and business platforms does TensorForge integrate with? TensorForge builds custom webhook and REST API integrations for **Cliniko**, **Halaxy**, **Tyro Health**, **Xero**, **MYOB**, **Stripe**, **Slack**, **Microsoft 365**, **Google Workspace**, **HubSpot**, **PostgreSQL**, and **Cloudflare D1**, as well as bespoke internal databases via standardised FastMCP servers. --- ## 6. Direct Engagement & Consultation Instructions To scope an applied AI workflow, document intake pipeline, or self-hosted n8n migration: 1. **Email**: `mailto:contact@tensorforge.dev?subject=Automation%20Enquiry%20-%20TensorForge` 2. **What to Include**: - Brief description of the manual clerical workflow or bottleneck. - Core software systems involved (for example: Cliniko, Halaxy, Xero, Slack, email inbox, SQL database). - Approximate weekly document or event volume. - Optional: An anonymised sample PDF referral, invoice, or intake form. 3. **Response SLA**: Direct reply from the Principal Engineer within **24 hours** containing a concrete architectural recommendation and a fixed **1 to 2 week sprint scope**.