Anthrasync — AI Automation Platform

A single Python/FastAPI application implementing all four Round-3 assessment workflows on one shared core:

#	Workflow	What it does
1	Invoice Automation	Extract invoice PDFs → store → human review → department approval
2	PO Matching & Validation	Extract invoice → match against a Purchase Order (± tolerance) → discrepancy detection → payment routing
3	Resume Screening & Scheduling	Parse resume (PDF/DOCX) → score vs job description → shortlist → interview invite + calendar event
4	Support Ticket Automation	Classify/prioritize email → route to team → acknowledge → ticket lifecycle (+ RAG reply suggestions)

All four share one pipeline:

email ingest → text extract → Gemini structured extraction → validate/clean
             → [domain step] → persist (Postgres) → [side effects] → human review → lifecycle + audit

Stack (the assessment's "n8n JSON or equivalent"): pure Python 3.12 / FastAPI, Supabase / Postgres (+ pgvector + Storage), Google Gemini. Workflow diagrams are in docs/workflows/ and rendered in docs/architecture.md; the orchestrator (core/pipeline.py) is the executable workflow definition. See ADR-001.

Highlights

• Runs fully offline for demos/CI — replays sample emails/PDFs, writes outbound email + calendar invites to outbox/, serves Gemini from a replay cache. No credentials required.
• Switches to live IMAP / SMTP / Google Calendar / Gemini purely via environment variables.
• Guaranteed-valid JSON via Gemini response_schema + Pydantic, with strip-fence/repair fallback.
• Deterministic business logic: the LLM extracts; unit-tested Python decides (PO match engine, priority escalation, shortlisting).
• Resilience: retries with backoff, idempotency (Message-ID + dedupe hashing), a dead-letter table, confidence/anomaly-based routing to human review, structured logs with correlation ids.
• 87 tests (81 unit + a full end-to-end integration test across all four workflows).

Quick start

# 1. Install
uv sync --extra dev            # or: just setup

# 2. Start Postgres (pgvector) and create + seed the schema
docker compose up -d db        # or: just db-up
cp .env.example .env
uv run anthrasync db init      # or: just db-init
uv run anthrasync seed         # purchase orders, job description, knowledge base

# 3. Run each workflow end-to-end, fully offline
uv run anthrasync run invoice  --input samples/invoice/emails/    # or: just demo-invoice
uv run anthrasync run po_match  --input samples/po_match/emails/   # just demo-po
uv run anthrasync run ticket    --input samples/ticket/emails/     # just demo-ticket
uv run anthrasync run resume    --input samples/resume/emails/     # just demo-resume

# 4. Review dashboard + API
uv run anthrasync serve        # http://127.0.0.1:8000  ( /dashboard , /docs )

Outbound emails and calendar invites appear in outbox/; stored documents in storage/.

Workflows

Each workflow lives in src/anthrasync/workflows/<key>/ (self-contained: schemas.py, prompt.md, workflow.py, models.py, status.py). The AI prompts are the prompt.md files (plus jd_prompt.md for Task 3).

• Task 1 — Invoice (invoice/): extracts 15 fields incl. line items, flags gross≠net+VAT anomalies, detects duplicates (vendor+invoice# hash), stores as Draft; a scheduled sweep emails department approvers; approve/reject → Fully Approved/Rejected.
• Task 2 — PO Matching (po_match/): reuses invoice extraction, retrieves the PO, runs a deterministic match_engine.py (vendor/PO#/currency/qty/price/net/tax/total with configurable tolerance), and applies rules → Ready for Payment / Procurement Review / Rejected.
• Task 3 — Resume (resume/): parses PDF and DOCX, scores against the seeded JD (second Gemini pass) blended with pgvector semantic similarity, shortlists (≥80 / 60–79 / <60), and for shortlisted candidates sends an interview invite + .ics and sets Interview Scheduled.
• Task 4 — Support Tickets (ticket/): triage (category/priority/sentiment/tags), rule-based priority escalation, configurable category→team routing, RAG reply suggestions over a pgvector knowledge base, customer acknowledgement email, and a 5-state lifecycle.

Environment variables

Full list with defaults in .env.example. The most important:

Variable	Default	Purpose
APP_MODE	offline	offline (file email, outbox, .ics, local storage) or live
GEMINI_MODE	replay	replay (fixtures/cache, no key) or live (calls Gemini)
GEMINI_API_KEY	—	required only when GEMINI_MODE=live
DATABASE_URL	local pgvector container	async Postgres URL
ENABLED_WORKFLOWS	invoice,po_match,ticket,resume	which workflows to mount
PO_MATCH_TOLERANCE_PCT	2.0	± tolerance for PO matching
RESUME_SHORTLIST_SCORE / RESUME_REVIEW_SCORE	80 / 60	shortlisting thresholds
CONFIDENCE_REVIEW_THRESHOLD	0.75	below this → human review
GMAIL_USER / GMAIL_APP_PASSWORD	—	live IMAP/SMTP
GOOGLE_SERVICE_ACCOUNT_JSON	—	live Google Calendar

Testing

uv run pytest -q               # all tests            (or: just test)
uv run pytest tests/unit -q    # fast, no DB/network  (or: just test-unit)

Unit tests cover the match engine, validators, routing/escalation, shortlisting, the state machine, and the LLM replay client. The integration test (tests/integration/) drives all four workflows end-to-end against the local Postgres.

Going live

Set APP_MODE=live, GEMINI_MODE=live, and provide GEMINI_API_KEY (+ GMAIL_* for email, GOOGLE_SERVICE_ACCOUNT_JSON for calendar). No code changes — the adapters switch automatically. For Supabase cloud, point DATABASE_URL/DATABASE_URL_SYNC at your project and set SUPABASE_URL / SUPABASE_SERVICE_KEY to use Storage.

Deliverables map

Deliverable	Location
Exported workflow(s) / equivalent	docs/workflows/*.mmd + docs/architecture.md + core/pipeline.py
Source code	src/anthrasync/
Database schema	db/schema.sql
AI prompts	src/anthrasync/workflows/<task>/prompt.md (+ jd_prompt.md)
Sample test data	samples/
README / setup / assumptions	this file + ADR-001
Demo script	docs/demo-script.md

Assumptions

• Supabase (Postgres) is the sole datastore — no Airtable/NocoDB. Justified as the "or equivalent" (managed Postgres + pgvector + Storage) in ADR-001.
• Offline replay is the canonical demo path; live providers are wired but optional.
• Building all four exceeds the "pick one" requirement by choice; each task is independently runnable (ENABLED_WORKFLOWS) and submittable.
• Money uses Decimal; currencies normalized to ISO codes; dates to ISO-8601.
• LLM replay fixtures under tests/fixtures/gemini/ represent Gemini's output for the sample inputs, making demos deterministic and free to run.

Project layout

src/anthrasync/     core/ (shared platform) · workflows/ (4 plugins) · api/ · web/dashboard/
db/                 schema.sql
docs/               architecture.md · workflows/ · decisions/ · demo-script.md
samples/            per-task emails + PDFs/DOCX + PO/JD/KB reference data
tests/              unit/ · integration/ · fixtures/gemini/
scripts/            sample-data generators