A lot was done over the weekend. But the processing of audio and image files is taking too long. And the UI doesn't look good. The business owner so far is not impressed.

Also the code got really messy. Deepseek and me will start doing some refactoring and simplification, and figure out a way to gradually populate the dashboard as processing goes.

What's shipped so far (like building a house)

Phase 0 was the foundation (rooms 0.1 through 0.7)

• The scaffold (pyproject.toml, Makefile, directory layers, make check)
• Domain objects: Money, Customer, Trip, Payment, Message, etc. (the brick and mortar)
• Ports: VisionModel, TextModel, Transcriber, Repository, Clock (the door frames)
• Mock adapters so every test runs offline, no GPU needed
• WhatsApp _chat.txt parser: handles iPhone and Android formats, media markers, multiline messages
• Gradio dashboard shell with Portuguese UI and custom CSS
  Phase 1 & 2 were the interior rooms (T1 through T8)
• Timeline (UC6): voice notes transcribed, chronological events
• Payments (UC2): reconcile receipts against quotes, show outstanding
• Trips (UC4): structured trip data, summary messages, .ics calendar
• Documents (UC7): track RG/CPF submissions per passenger
• CRM card (UC1): orchestrator that ties everything together from a single upload
• Pipeline (UC3): infer stage (QUOTE > DOCUMENTS > PAYMENT > TICKETING > TRAVEL > DONE) and next action
• Reply drafting (UC5): prompt assembly for her-voice replies
• Dashboard (UC8): aggregate across customers, search "who still owes me?"

Today's four commits were plumbing fixes I flagged

• .gitignore now covers *.zip
• Timeline now has a passing Gherkin scenario (it was the only missing feature test)
• ZIP upload: load_zip() extracts WhatsApp exports to a temp dir, parses _chat.txt, routes media files — so those 300MB ZIPs work
• Continuous accumulation: uploading a second export for the same customer merges payments, trips, messages, contacts instead of overwriting: the dashboard is a growing file, not a fresh one each time

What's still missing (next rooms to build)

Phase 3 — Real model adapters (T9, T10, T11)

• The live llama.cpp MiniCPM-V adapter so it actually OCRs receipts and tickets instead of returning canned data
• The whisper transcriber so audio notes are really transcribed
• Modal dev bench for offline data curation
  Phase 4 — Fine-tuning (T12, T13)
• The LoRA pipeline so replies actually sound like her
• The evals harness to prove the model is good enough
  Phase 0 leftover — SQLite repository (T0.6)
• Only in-memory exists right now. Data vanishes on restart. SQLite would make it persistent

Switched to opencode and added agent browser, which greatly improved the feedback/debugging loop for Gradio UI. Deepseek was able to fix errors and we have something working now!

Published our space: https://huggingface.co/spaces/build-small-hackathon/vivamais

• Phase 0 built the project skeleton with uv, ruff, mypy, and pytest, plus the framework free domain core (Money, Customer, Trip, Payment, Message, Document entities and PipelineStage/MessageKind enums). We added application ports as typing.Protocol interfaces, deterministic mock adapters for offline testing, a WhatsApp chat parser handling multiline messages and media placeholders, and a config driven adapter factory wired through a composition root in app.py.

• Phase 1 delivered four core use case interactors: the CrmCardInteractor for structured customer cards, PaymentReconciliationInteractor for total paid and outstanding balance, PipelineInteractor for pipeline stages and next actions, and TripSummaryInteractor for Portuguese summaries and .ics calendar files.

• Phase 2 tied everything together with a BuildCrmCard orchestrator and CRM presenter, then added DocumentTrackerInteractor for passenger document readiness, DraftReply for in voice reply generation, and DashboardInteractor for cross customer aggregation and search. All 107 tests pass across 8 passing Gherkin scenarios.

• I'm sharing quick, short updates in this X thread as well: https://x.com/amphetamarina/status/2062864424403112188

What have I done so far:

• Checked Modal pricing since we've received credits for the hackathon. Initially I wanted to use it for inference servers for my coding agents, but It could be prohibitive in terms of costs. I will use it primarly to host @OpenBMB-LLM MiniCPM-o 4.6 (1.3B) model for testing, dataset curation and possibly fine-tuning tasks.

• The business owner provided me 10 raw WhatsApp exports from her most valuable (ROI) and frequent customers. I already have enough data to start.

• For my main coding agent, I chose Deepseek-V4-Flash model on OpenRouter because I am already familiar with Deepseek-V4 models, they are good at long running tasks and they are very cheap. I'm also going to use it with OpenAI's Codex harness.

• I interviewed her to narrow down the most valuable use cases for her business. Came up with these, in order of priority/deliverable value:

  1. Auto-built customer CRM card
  • Given a raw WhatsApp conversation export with its media,
  • When the user uploads it to the Space,
  • Then the system extracts the text, the receipt and ticket images, and the whisper-transcribed voice notes, and assembles a structured card per customer (who they are, who is traveling with them, contacts, and the trips discussed).
  2. Payment tracking and reconciliation
  • Given a customer thread that contains payment receipts (PIX/TED) and quotes,
  • When the user opens that customer's card,
  • Then the system shows the total paid, the outstanding balance, and the profit on the sale.
  3. Sales-pipeline status and next action
  • Given a conversation history that reveals where the sale stands,
  • When the user opens a customer,
  • Then the system places them on the funnel (quote, documents, payment, ticketing, travel), lists the remaining steps, and surfaces the single most important action to take next.

  4. Trip summary and reminders

     • Given an uploaded e-ticket or itinerary with a locator, passengers, and flights,
     • When the user processes it,
     • Then the system produces a structured trip, the trip-summary message in her style, check-in reminders, and a calendar (.ics) file.

  5. Reply drafting in her voice

     • Given a LoRA fine-tuned on her real chats,
     • When a customer message comes in,
     • Then the system drafts the next reply in her warm, personal style for the user to review and send.

  6. Voice-note transcription and timeline

     • Given a thread that contains voice notes,
     • When the user opens it,
     • Then the system transcribes each note with whisper and builds a searchable, summarized timeline of the relationship.

  7. Document-collection tracking

     • Given a trip whose passengers must each submit RG and CPF,
     • When the user prepares ticketing,
     • Then the system shows which documents have arrived and which are still missing.

  8. Cross-customer dashboard and search

     • Given several customers already processed,
     • When the user opens the overview,
     • Then the system aggregates money, time, and profit and answers plain-language questions such as "who still owes me?" or "which trips depart this week?".

Next Steps:

• Create the space with a skeleton structure.
• Set up Codex with OpenRouter + Deepseek V4 Flash
• Set up MiniCPM-o 4.6 (1.3B) on Modal servers.
• Write an implementation spec, Gherkin test cases, and AGENTS.md instructions for the actual coding (BDD).