A/IAALQUIM IASELECTED WORK · CASE 10
← Back to portfolio

CASE 10 · TECHNICAL PILOT MULTIMODAL ASSISTANT · DETERMINISTIC CORE

Lía

The LLM interprets; the application decides.

Lía is a WhatsApp-first personal assistant that turns voice notes, documents and photos into structured memory with provenance: decisions, questions and commitments. A deterministic engine governs states, permissions and billing, while two models —an interpreter and an evaluator— work under explicit authority limits.

ROLEFOUNDER · PRODUCT · AGENTIC ENGINEERING2026
Watch the case video Connect on LinkedIn
134/134user journeys validated
7/7adversarial GO, zero critical
39/39real JEV evaluations
6containers in staging

CASE VIDEO

Narrated walkthrough of the Lía case: approach, challenges and outcome (3:33).

PROJECT CARD / RECRUITER VIEW

The project's technical dossier.

EXPLICIT STATE TECHNICAL PILOT
TARGET USER

People and teams who receive commitments, dates and documents over WhatsApp and need dependable memory with human confirmation.

WHAT I BUILT

I designed and built the whole system: the deterministic engine, the resumable multimodal pipeline (Whisper + OCR with evidence regions), the JEV evaluation layer, multi-tenant RLS, the magic-link portal and the sandbox billing loop.

AI ENGINEERING SIGNAL / RECRUITER VIEW

Backend, APIs, deployment, testing and controls.

DEPLOYMENT
Staging on Oracle Cloud: api, worker, web, media, nginx and postgres; health gates, offsite backups and rollback drills.
BACKEND / APIS
Meta WhatsApp Cloud API with signed (HMAC) webhooks, magic-link portal and optional channels (Telegram, Google).
AI LAYER
DeepSeek interprets; JEV audits with closed options and confidence, with per-flow off/shadow/enforce modes and its own budget ledger.
DATA / SECURITY
Account- and workspace-level isolation with RLS, five roles with explicit scopes, and identity linking via a one-time challenge (HMAC fingerprint only).
BILLING
Versioned COP/USD pricebook, immutable quotes, idempotent webhooks verified with HMAC over the raw bytes, replay deduplication, a 14-day cardless trial and reconciled entitlements.
TESTING
134/134 user journeys, a 7/7 adversarial go-live review with zero critical or high findings, and 434/434 local tests in the JEV integration.
ARCHITECTURE OVERVIEW
  1. 01WhatsApp Cloud API + signed webhooks
  2. 02Worker: resumable audio + Whisper + OCR with evidence
  3. 03Memory with provenance · retrieval isolated by account/workspace/owner
  4. 04Multi-tenant PostgreSQL with RLS · five roles
  5. 05Magic-link portal + sandbox billing
WHAT WORKS TODAY
  • Staging on Oracle Cloud: 6 healthy containers and readyz 200
  • 134/134 user journeys validated
  • JEV: 39/39 real evaluations (p50 448 ms)
  • Complete billing loop in sandbox, deliberately switched off
HARD TECHNICAL DECISIONS
  • The model never governs state, permissions or money
  • A hard boundary in code and tests between interpretation and decision
  • A resumable, idempotent pipeline: nothing lost, nothing double-charged
  • Keep commercial billing off until external approvals exist
MEDIA AND DEMO
VIDEO
3:33 narrated walkthrough (English)
DEMO
Private staging pilot
SCREENSHOTS
Shared with coordinated access
WHAT IS STILL PENDING
  • The pilot lives in staging: commercial billing stays off until external approvals exist —an explicit decision, not a forgotten to-do.
  • JEV is a probabilistic evaluator: it audits, orders and lowers confidence; authority remains in code.
  • No credentials or tenant data are published; the demo is shared with coordinated access.

01 / PROBLEM

Conversation is where information goes to die.

Commitments, dates and decisions sit buried inside voice notes and documents. The obvious fix —letting an LLM handle it— is dangerous: models hallucinate and should never touch your money or your permissions.

02 / DESIGN DECISION

A deterministic core with LLMs at the edge.

The LLM interprets and extracts; the application governs states, permissions and billing. JEV audits every interpretation with closed options and confidence, under hard authority limits.

SYSTEM / END TO END

The product as an explicit flow.

Each phase has a distinct responsibility. That separation makes it possible to observe, test and fix the system without relying on intuition.

  1. 01

    Capture

    Voice, documents and photos come in over WhatsApp with signed webhooks.

    CAPTURE
  2. 02

    Interpret

    DeepSeek reads intent; JEV audits it with closed options and confidence.

    AUDIT
  3. 03

    Confirm

    Every field waits for human review; nothing executes without confirmation.

    CONFIRM
  4. 04

    Remember

    Canonical memory with provenance and retrieval isolated by account, workspace and owner.

    MEMORY
  5. 05

    Bill

    Sandbox billing loop: immutable quotes, idempotent webhooks and entitlements.

    BILLING

PROOF / WHAT IS ACTUALLY SHOWN

Evidence, not decoration.

The labels indicate whether the data was observed, comes from the artifact or belongs to the professional record of the case.

STAGING EVIDENCE · SEP 2026134 / 134user journeys in the final validation
ADVERSARIAL GO7 / 7adversarial review with zero critical or high
OBSERVED MEASUREMENT39 / 39real JEV evaluations · p50 448 ms
OBSERVED HEALTH6 / 6healthy containers · readyz 200

SCOPE / HONEST BOUNDARIES

What is demonstrated and what still needs validation.

A technical portfolio must also show where the available evidence ends.

  • The pilot lives in staging: commercial billing stays off until external approvals exist —an explicit decision, not a forgotten to-do.
  • JEV is a probabilistic evaluator: it audits, orders and lowers confidence; authority remains in code.
  • No credentials or tenant data are published; the demo is shared with coordinated access.

CASE 10 / LÍA

The LLM interprets; the application decides.

Watch the case video View the full portfolio