NOTE: This platform uses free-tier AI APIs and works best with the Sample Manual due to usage limits. Newer AI models and production-grade APIs will improve performance and accuracy.
Try it yourself
Do not have a manual? Try the sample.
Download the synthetic 4-page battery manual, then upload or email it to explore the planned workflow.
This portfolio demo runs within free-tier Gemini API constraints. A compact synthetic manual makes the complete workflow practical while controlling quota and token usage.
See how the system works.
Switch between Overview, Web Upload, Email, and Ask Manual. Select any item for a short explanation, or open the Technical view for more detail.
Web and email use the same private workspace services. Four AI roles interpret information. The Calendar role uses fixed rules to turn eligible dates into ICS events.
01
Web upload → results
Gemini finds facts and source evidence. Fixed date rules calculate supported actions. Eligible dates then move to the Calendar role for release.
02
Ask your manual → cited answers
A separate question-and-answer branch searches retained documents when you ask.
03
Email → an ongoing conversation
The workflow can process attachments, ask for missing information, and continue after a reply. Calendar delivery requires clear approval.
Strategy
From discovery to measurable value.
About 15–20 informal conversations raised questions about how solar owners manage manuals, maintenance tasks, warranty information, and important dates. These early discussions helped shape the problem and MVP assumptions. Formal market research and demand validation are still needed.
About 15 to 20 informal conversations suggested that solar records may be scattered, maintenance tasks may be hard to find, and owners may be unsure which dates matter. Those concerns shaped a product concept centered on useful actions, timing, and source evidence from manuals people already have. Formal research is still needed to learn how often the problem occurs and what would earn a homeowner's trust.
The PRD defines private workspaces, web and email intake, optional installation dates, source-linked results, calendar rules, document Q&A, quality checks, and recovery from provider failures. The scope stays focused on understanding manuals and helping people act on them. AI is planned for evidence interpretation, while application rules control access, calculations, approvals, and delivery.
The web path gives people direct control over upload, results, calendar download, and manual Q&A. The email path can collect a manual and optional installation date, ask for missing information, return results, and deliver an approved calendar invite over several replies. Both paths use the same extraction, date, calendar, quality, storage, and access services. Saved state lets work pause and continue without relying on model memory.
Web or Email→Private evidence→Results→Calendar or Q&A
The measurement plan follows a user from manual submission to a useful result and then to a meaningful action. Activation Rate, Time to Value, Calendar Conversion, and Task Success cover different parts of that progression. Processing Failure and Validation Pass provide separate health checks, and targets should follow a measured baseline.
Planned result-set funnel
Each stage shows how a submitted manual moves toward a useful result and, when eligible dates exist, a calendar action.
Manual Submitted→
Processing Completed→
Validation Passed→
Actionable Results→
Calendar Action
Four core Product KPIs
Activation RateSubmitted workspaces that reach a QA-passed actionable result.
Median Time to ValueTime from submission to the first persisted actionable result.
Calendar ConversionEligible result sets that complete an ICS download or accepted email delivery.
Task Success RateTerminal result sets that produce QA-passed actionable results.
Product Health / Guardrails: Processing Failure Rate and Validation Pass Rate diagnose reliability. MVP acceptance separately checks safety, functional completion, evidence/date quality, calendar/validation quality, user value, and operational readiness.
The plan uses AI for reading and interpretation, with application rules for access, calculations, source tracking, approval, and external actions. Free-tier limits require saved progress and clear recovery states. Uncertainty stays visible, web and email share one document pipeline, and Manual Q&A must say when evidence is missing.
Each role has a narrow responsibility and passes structured, saved output to the next part of the workflow. This separation keeps language interpretation flexible while dates, access, approvals, and delivery remain predictable.
AI Agent 1: Document ExtractionReceives native PDF or image chunks plus page metadata. Gemini identifies equipment, maintenance rules, notes, and source evidence, then saves typed extraction output for date calculation and results.
Deterministic Agent 2: Calendar GenerationReceives validated results and an optional installation date. Application code calculates eligible dates and produces ICS events without using an LLM.
AI Agent 3: Email EvaluatorReceives one verified email case and its allowed state. Gemini interprets language or missing date information, while LangGraph and a case-scoped MCP boundary control the next permitted step.
AI Agent 4: Manual Q&A (RAG)Receives a question and retrieved chunks from retained workspace manuals. Gemini answers only from that evidence, and application code validates file and page citations.
AI Agent 5: QA/QCReceives extraction, results, email, or calendar artifacts. Gemini review and fixed checks produce a pass, bounded correction, or item-specific manual review state before release.
How they work togetherExtraction finds the evidence. Deterministic code calculates dates and calendar events. QA/QC checks important outputs. Email Evaluation guides verified conversations, and Manual Q&A answers later questions from retained manuals.
Documents enter through web upload or verified email and are stored privately in Supabase. Extraction turns page evidence into structured facts. Application code calculates dates, QA/QC checks important output, Calendar Generation creates eligible events, and Manual Q&A searches retained documents. Saved jobs let work pause and resume safely.
Web or Email→Private documents→Extraction and checks→Results, calendar, or Q&A
Gemini receives a focused task, a typed response format, and only the evidence needed for that step. LangGraph coordinates verified email turns, MCP limits available tools, and Supabase stores progress. Deterministic code remains responsible for security, calculations, approval, and side effects.
The application uses private storage, workspace isolation, file and submission limits, bounded retries, and free-only AI configuration. It does not use an LLM for access control, authoritative date arithmetic, approval, or calendar serialization. Workspaces and private files expire 15 days after creation.
Gemini reads document content, interprets email messages and installation dates, drafts answers from retrieved manual sections, and reviews selected outputs. Application code still controls access, calculations, approvals, and delivery.
LangGraph guides a verified email through the allowed steps. Supabase saves the conversation state so a reply, pause, or retry can continue later without relying on AI memory.
The Email Evaluator receives only the case-specific tools it needs. A separate Resend MCP connection lets trusted server code retrieve and send email without exposing provider credentials to Gemini.
The app breaks retained manuals into searchable sections. Supabase pgvector finds the sections most relevant to a question, then Gemini answers from that evidence and the app checks the file and page citations.
Supabase stores uploaded files, workspaces, jobs, results, email state, approvals, calendar records, RAG data, QA reviews, and product metrics. It also runs scheduled workers and keeps each workspace isolated.
Resend receives inbound messages and attachments. The application uses its verified email connection to send questions, results, private workspace links, and approved calendar invitations.
Vercel runs the Next.js website, server routes, provider callbacks, and protected worker endpoints using the production environment settings.
GitHub keeps the code, migrations, specifications, and documentation history together. Automated checks validate TypeScript, linting, tests, and the production build.
Spec Kit and OpenSpec were explored to make requirements and implementation planning more structured. They were useful experiments, but for this MVP the workflow added token usage and process overhead better suited to production-level development than the speed and scope needed here.