Landing page of the app
General manager main view
Region specific view
Autonomous community specific view
Overview of the alerts from an specific comercial delegate
Main comercial delegate page
Overview of an specific closed alert with it's changelog
Chatbot about an alert to get more insights and easier comprehension of the metrics and decisions behind the alert
Action menu for an alert, allowing to register phone calls / emails / visits / other, the result and if the alert should be kept open

Client Alert Manager

Inspiration

INIBSA challenged InterHack-2026 teams to transform dental-clinic purchase history into actionable commercial signals. The core business problem is simple to state but difficult to solve well: sales teams need to know when a clinic should buy again, when demand is drifting to competitors, and when a stable customer is starting to show abandonment risk.

We built Client Alert Manager to turn those signals into a practical workflow for sales delegates and regional managers.

What It Does

Client Alert Manager is a standalone sales intelligence application for dental-clinic demand signals.

Predicts and surfaces purchase-need alerts.
Detects early churn or deterioration signals.
Prioritizes alerts by urgency, customer value, and risk.
Shows interpretable explanations for each alert.
Gives sales delegates a focused alert workflow.
Gives regional managers a map-based performance view with manager, agent, and client drill-down.
Uses an AI assistant to answer contextual questions about individual alerts.
Tracks commercial outcomes so alerts can become a learning system over time.

How We Built It

The solution has three main layers.

1. Analytics and Alert Generation

The IA/ layer processes customer and product purchase behavior and produces alert candidates. It combines purchase timing, customer potential, risk scoring, purchase propensity, product-family context, and explainability fields.

The generated alert dataset is loaded into the backend so the demo behaves like a real operational system instead of a static mockup.

2. Backend and APIs

The backend is built with FastAPI, async SQLAlchemy, Alembic, and SQLite. It exposes APIs for:

sales alerts,
regional dashboard KPIs,
AI chat insight,
authentication flows,
notification-oriented flows,
and audio/voice-oriented integrations.

The architecture keeps the application standalone while leaving a clear path for future CRM or marketing automation integration.

3. Frontend Dashboards

The frontend is built with React, Vite, TypeScript, Tailwind CSS, shadcn-style primitives, TanStack Query, and Framer Motion.

It includes:

a role-based login experience,
a sales delegate alert dashboard,
a regional manager dashboard,
an interactive Spain map using d3/topojson,
localized Catalan and Spanish UI,
and a contextual AI insight panel.

AI and Integrations

We integrated Gemini through the backend to keep credentials server-side and give users natural-language insight into alert context.

We also prepared backend integration points for:

Groq-based LLM workflows,
ElevenLabs voice experiences,
AssemblyAI transcription/audio flows,
Dockerized deployment,
and future CRM or marketing automation activation.

Challenges We Ran Into

The hardest part was translating a broad business challenge into a workflow that feels operational, not just analytical.

We had to balance:

commodity products with recurring demand,
technical products with irregular purchase behavior,
alert prioritization,
explainability,
sales follow-up tracking,
and regional performance visibility.

Another challenge was making the regional map useful for business users. We moved from a rough stylized shape to a more accurate d3/topojson-based Spain map so the dashboard could support real geographic exploration.

Accomplishments That We Are Proud Of

Built an end-to-end standalone application during a hackathon.
Connected ML-generated alerts to a real backend API.
Created two role-specific dashboards: delegate operations and regional management.
Added a contextual AI assistant that explains alert data.
Designed the workflow around traceability: alert, action, outcome, and history.
Built a regional map and drill-down hierarchy for managers, agents, and clients.

What We Learned

Good predictive systems need more than a score. They need:

a clear action,
an explanation,
prioritization,
ownership,
traceability,
and feedback loops.

The project reinforced that model output becomes valuable only when it fits the commercial process and helps users decide what to do next.

What's Next

Connect the daily pipeline to production-grade data ingestion.
Expand model validation with business outcome metrics.
Add CRM export and marketing automation activation.
Replace demo role selection with full backend authorization.
Add stronger monitoring around alert conversion, false positives, and recovery rates.
Extend the architecture beyond Spain for future international rollout.

Built With

React
Vite
TypeScript
Tailwind CSS
shadcn/ui-style components
TanStack Query
Framer Motion
d3-geo
topojson
FastAPI
SQLAlchemy
Alembic
SQLite
Docker Compose
Gemini
Groq
ElevenLabs
AssemblyAI
Python

Contributors

Alvaro - AI Engineer
Yearsuck - Infrastructure and Backend Engineer
Ki-re - Front and Back Engineer

Built With

alembic
assemblyai
docker
elevenlabs
fastapi
gemini
godaddy
numpy
pandas
python
pytorch
react
sqlite
tailwind
typescript
vite
vultr

Submitted to

InterHackBCN
- Winner 1st price - Inibsa Challenge
- Winner [MLH] Best Use of ElevenLabs

Created by

I was one of the core engineers on this project, working across the full stack from the first commit.

On the infrastructure side, I set up the initial repository structure, Docker Compose configuration, and the development environment so the team could iterate quickly from day one. I also handled DevOps concerns like container orchestration, service wiring, volume mounts, and the Alembic migration pipeline.

On the backend, I designed and implemented the FastAPI application structure; routers, async SQLAlchemy models, database schema, and
the Alembic migration chain. I configured SQLite as the default database and built the data loading pipeline that seeds the live database from the ML-generated alerts CSV at startup, so the demo behaves like a real operational system rather than a static mockup.

I also built the Gemini from scratch. This includes the server-side LLM module (llm/), the prompt
engineering system in YAML (diagnostic analyst persona, bilingual CA/ES, BREVITY rule, churn type context, competitive vs. budget displacement reasoning), the Gemini 2.5 Flash async wrapper, and the /ai/chat API endpoint consumed by the frontend AI insight panel.

On the frontend, I laid the initial groundwork for the React/Vite/TypeScript/Tailwind setup and contributed across both the sales delegate dashboard and the regional manager dashboard, including the interactive Spain map, the alert workflow, the changelog for the delegates and other features from our app.

Erik Batiste
DevOps & Infrastructure
Set up the full production deployment pipeline on Vultr VPS (London). Configured Docker Compose with multi-service orchestration (FastAPI backend, Vite/React frontend, Nginx reverse proxy, Certbot for SSL). Managed environment secrets, health checks, and automated migrations on startup. Registered and configured the domain keepinibsa.work with HTTPS via Let's Encrypt.

Frontend
Built and iterated on the React/TypeScript frontend: responsive region map grid, per-agent notification system with optimistic UI, WhatsApp-style audio recording interface with waveform visualizer, bilingual UI (Catalan/Spanish) with full locale coverage across all components and LLM context labels.

ElevenLabs Voice Cloning & TTS Integrated ElevenLabs eleven_multilingual_v2 model for text-to-speech responses in the AI chat panel. Cloned two custom voices — one in Spanish, one in Catalan — and wired dynamic voice selection based on active UI language. Debugged and fixed language_code API incompatibility and environment variable propagation across Docker Compose.

Gemini LLM Integration
Enforced UI language adherence in the Gemini system prompt so the AI always responds in the delegate's selected language regardless of the input language. Localized LLM context block field labels per language.

Notification System
Designed and implemented per-agent scoped notifications with APScheduler daily job, Alembic migration for agent_id column, and a React global AgentContext to wire selected agent from Dashboard to the notification bell.

Ernest Rull Turigas
Machine Learning & AI

Designed and trained three complementary neural network models for B2B commercial intelligence on dental product sales data. The churn/recompra model is a feed-forward network (512→256→128, SiLU activations, BatchNorm, Dropout) with label-smoothed binary cross-entropy that predicts the probability a client will not repurchase a given product. The potencial model is a GRU-based sequential classifier that ingests the full purchase history per client-product pair and outputs a 5-class ordered potential label (muy_negativo → muy_positivo) mapped to a 0–100 score. The días model is a multi-task GRU with two heads: a regression head (Softplus-bounded log-days) predicting days until next purchase, and an auxiliary classification head predicting the repurchase month; both trained jointly on padded, variable-length sequences.

Built the full data ingestion and feature engineering pipeline (`pipeline_ingestion_datos.py`) to transform raw Excel sales sheets (Ventas / Productos / Clientes) into model-ready feature matrices: deduplication, product-level repurchase frequency stats, client-product temporal metrics, category-level spend aggregations, purchase count and return signals, and z-score normalisation of all temporal features. Applied log-transforms to handle heavily skewed monetary and count distributions, one-hot encoded all categoricals, and aligned feature sets between training and inference.

Implemented the inference pipeline (`pipeline_inferencia.py`) with batched DataLoader support for GPU/CPU, padded-sequence collation for the sequential models, and output alignment back to original row indices. Implemented the retraining pipeline (`pipeline_retrain.py`) as a modular multi-step subprocess orchestrator with checkpoint saving, ReduceLROnPlateau scheduling, gradient clipping, and mixed-precision (AMP) training.

Built the alert generation pipeline (`generate_alerts.py`) that converts raw model scores into actionable sales alerts: a configurable temporal window filters the latest qualifying purchase per client-product, client value is segmented by historical spend percentile (Alto/Medio/Bajo) with tiered risk thresholds, and alert type is determined by the number of breaching products (Total / Combinat / Producto X). Alerts are geographically routed from province → Comunidad Autónoma → commercial zone → agent ID, risk levels are assigned by percentile rank, and scores are rescaled to realistic display ranges.

Implemented permutation feature importance explainability (`Explicabilidad/model_feature_importance.py`) for all three models: shuffled each feature column over a 5,000-row sample, measured loss degradation, and aggregated one-hot components back to their original variables. Produced a JSON report (`feature_importance_report.json`) with ranked feature importances and baseline metrics for each model (e.g. recompra: BCE=0.12, accuracy=96.7%; potencial: top driver is product category; días: top driver is client-product repurchase ratio).

Evaluated all models in a Jupyter notebook (`tests/confusion_matrix_metricas_modelos.ipynb`): reconstructed held-out test sequences matching training logic, generated normalised and absolute confusion matrices, computed accuracy, precision, recall, and macro/weighted F1 for classifiers, and MAE/RMSE for the regression head. Wrote a `deanonymize.py` utility for deterministic synthetic clinic name generation to produce GDPR-compliant demo data for the dashboard.

Alvaroost8

Updates

Erik Batiste started this project — May 10, 2026 04:17 AM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.