Plataforma de Gestión de Demanda con Agentes IA
Digitises the full lifecycle of a technology demand through 5 AI-assisted stages with mandatory human validation.
- Role
- Full-stack development
- Year
- 2026
- Status
- Corporate project · internal
- Client
- Argentina's largest oil company
- 5
- AI-assisted stages
- HITL
- mandatory human validation
- Multi-agent
- LangGraph orchestration
- Ports & adapters
- swappable infrastructure
Overview
The company's technology department received business demands as loose documents and meetings: each initiative went through requirements gathering, solution research, architecture design, estimation and business case preparation, with weeks of back-and-forth between analysts, architects and cybersecurity, and no traceability of who validated what. The platform turns that process into a guided 5-stage workflow assisted by AI agents, each with mandatory human validation.
Context
The problem
Weeks of back-and-forth between analysts, architects and cybersecurity for every initiative, with artifacts scattered across loose documents and no traceability of who validated what.
The solution
A single, auditable workflow with persisted state. The business user talks to an agent that gathers the requirement, and the system automatically produces each stage's artifacts: a requirements document as PDF, a comparison of internal and external solutions, an editable architecture diagram, an estimate and a business case. The AI does the heavy lifting of writing and analysis; people keep the decision.
What it does
Conversational requirements gathering
The business user describes their need in natural language and an agent conducts the interview, producing the formal document as a PDF.
Solution comparison
The system contrasts solutions that already exist inside the company against market alternatives, avoiding building the same thing twice.
Editable architecture
It generates a proposed architecture diagram the architect can edit rather than draw from scratch.
Estimation and business case
It closes the cycle by producing the effort estimate and the business case ready to present to committee.
Human-in-the-loop at every stage
No stage advances without explicit human validation. The AI proposes, the person approves, and everything is recorded.
Technical decisions
- 01
Ports & adapters in the infrastructure layer
LLM model, persistence, telemetry, messaging and blob storage are chosen by environment variable. Adding a provider means a new file plus one line of registration, without touching the application core or the agents.
- 02
Pluggable telemetry
Langfuse hooks in through configuration: callbacks are registered in the model factory, so the agents don't change when observability is switched on or off.
- 03
Multi-agent orchestration with LangGraph
Each stage of the cycle is a node in the graph with its own specialised agent and its own human validation point, rather than a single monolithic prompt.
Stack
Frontend
- Next.js 14
- React 18
- TypeScript
- Tailwind CSS
- TanStack Query
Backend
- NestJS
- Python
- LangGraph
- Anthropic API
Data
- PostgreSQL
Infra
- Azure
- Blob storage
- Langfuse
- Ruff
- pytest
Next project