Architecture v0.2

Charles Dana · Monce AI · April 2026

Request flow

                  HTTPS
                    │
          ┌─────────▼─────────┐
          │   Route53 A       │   alma.aws.monce.ai → 35.180.91.48
          └─────────┬─────────┘
                    │
          ┌─────────▼─────────┐
          │   nginx (443)     │   Let's Encrypt TLS, reverse proxy
          └─────────┬─────────┘
                    │  127.0.0.1:8000
          ┌─────────▼─────────┐
          │  uvicorn/FastAPI  │   systemd: alma.service
          │  api/app.py       │
          └─────────┬─────────┘
                    │
          ┌─────────▼─────────┐       ┌─────────────────────┐
          │  Pipeline (7 st)  │──────▶│  AWS Bedrock        │
          │  stages 0..6+4.5  │       │  eu-west-3          │
          └─────────┬─────────┘       │  Haiku 4.5          │
                    │                 │  Sonnet 4.6         │
                    │                 └─────────────────────┘
                    │
          ┌─────────▼─────────┐
          │  Snake v5.4.5     │   pure Python, in-process
          │  SAT classifier   │   3 models in data/models/
          └───────────────────┘

Pipeline stages

#	Stage	Cost	Typical time	Output
0	Client ID	regex + 1 Haiku call	< 1 s	customer_id, CIF, language
1	Doc Analyzer	regex + 1 Haiku call	< 1 s	page triage, layout
2	Unified Extractor	1 Sonnet call per 3-page chunk	3–6 s	structured JSON
3	Rules Engine	deterministic	< 10 ms	cleaned values
4	Snake Matcher	in-process SAT	< 20 ms/line	matched PN/customer
4.5	Pricing (NEW)	deterministic lookup	< 50 μs/line	tariff-filled unit_price
5	Validation	1 Haiku call	1–2 s	flags & overall confidence
6	Router	deterministic	< 1 ms	auto-approve / review / full review

Data layout

data/
 ├─ articles.json         847 parts (67 real + 180 AXALYS + 600 filler)
 ├─ clients.json          216 customers (16 real + 200 filler)
 ├─ alma_entities.json    1 entity — ALMA COMPONENTES
 ├─ axalys_tarif.json     180 AXALYS refs across 21 product sheets
 ├─ pricing.json          313 priced entries (155 GPAO + 158 AXALYS)
 └─ models/
    ├─ article_matcher.json
    ├─ client_matcher.json
    └─ entity_matcher.json

tarifs/
 └─ TARIF GENERAL 240207.xlsx  (password: soleil)

Deployment topology

Layer	Spec
EC2	t3.medium, Ubuntu 24.04, 20 GB gp3, eu-west-3
IAM	instance profile `safran-po-ec2-profile` → Bedrock InvokeModel
Auth	`AWS_BEARER_TOKEN_BEDROCK` env in systemd unit
Nginx	443 TLS (Let's Encrypt) → 127.0.0.1:8000
Service	`alma.service`, auto-restart, 3 worker coroutines
DNS	Route53 zone `aws.monce.ai`, A-record TTL 300
Cert	certbot renewal cron, valid until 2026-07-21

Extensibility

New customer layout. Add a regex to stage0_client_extraction.py:_CUSTOMER_PATTERNS plus an entry to REAL_CLIENTS in scripts/generate_catalog.py. Retrain Snake → deploy.
New tariff entries. Update tarifs/TARIF GENERAL….xlsx, re-run the extractor (msoffcrypto decrypts with password soleil), deploy data/pricing.json.
New language. Extend _detect_language() and the Stage 2 prompt column dictionary.
New model. Point MODELS in config.py to any Bedrock-hosted Anthropic model ID.