#!/usr/bin/env python3 """ Generate a synthetic, LINKED "insurance data estate". Everything ties back to the same customers / policies / claims, so you must join + dedup + normalize across heterogeneous sources. ALL personal data is FAKE (Faker). No real individuals. Safe to host publicly. """ import os, json, random, csv, io, textwrap from datetime import date, timedelta from faker import Faker from fpdf import FPDF from PIL import Image, ImageDraw, ImageFont, ImageFilter SEED = 2026 random.seed(SEED) fake = Faker() Faker.seed(SEED) # ---- branding / id scheme (different from the earlier version) ---- COMPANY = "Vantage Mutual" DOMAIN = "vantage-mutual.example" CUST_PREFIX = "CU" POLICY_PREFIX = "PLC-" CLAIM_PREFIX = "CLA-" OUT = os.path.dirname(os.path.abspath(__file__)) def p(*a): return os.path.join(OUT, *a) for d in ["structured", "documents/pdf", "documents/scans", "emails", "transcripts"]: os.makedirs(p(*d.split("/")), exist_ok=True) # ---- scale (representative SAMPLE; README says "design for millions") ---- N_CUST = 2400 N_POLICY = 3100 N_CLAIM = 3900 N_DOCS = 450 # subset of claims that also get a scanned form + PDF N_EMAIL = 700 # email threads N_CALL = 350 # call transcripts LINES = ["auto", "home", "health", "life", "commercial"] CLAIM_STATUS = ["open", "under_review", "approved", "denied", "paid", "closed"] def msgid(): return f"<{fake.uuid4()}@{DOMAIN}>" def phone(): return fake.numerify(random.choice(["(###) ###-####","###-###-####","+1##########","###.###.####"])) def fake_ssn(): return fake.numerify("###-##-####") # synthetic, not a real SSN def fake_policy_no(): return POLICY_PREFIX + fake.numerify("########") def fake_claim_no(): return CLAIM_PREFIX + fake.numerify("#######") # ============================================================ # 1) CUSTOMERS (CRM source) -- with injected PII + dup/noise # ============================================================ customers = [] for i in range(N_CUST): cid = f"{CUST_PREFIX}{500000+i}" name = fake.name() customers.append({ "customer_id": cid, "full_name": name, "email": fake.email(), "phone": phone(), "dob": fake.date_of_birth(minimum_age=18, maximum_age=88).isoformat(), "ssn": fake_ssn(), # <-- PII (fake) "street": fake.street_address(), "city": fake.city(), "state": fake.state_abbr(), "zip": fake.postcode(), "created_at": fake.date_between("-10y", "today").isoformat(), }) real_cust_ids = [c["customer_id"] for c in customers] # the originals, before duplicates # inject ~8% near-duplicate / dirty rows (re-entry, casing, whitespace, format drift) dupes = [] for c in random.sample(customers, int(N_CUST*0.08)): d = dict(c) d["customer_id"] = f"{CUST_PREFIX}{990000+len(dupes)}" # different surrogate key, same person mut = random.choice(["case","space","typo","phonefmt","emaildot"]) if mut == "case": d["full_name"] = d["full_name"].upper() if mut == "space": d["full_name"] = " " + d["full_name"] + " " if mut == "typo": d["full_name"] = d["full_name"].replace("a","@",1) if mut == "phonefmt": d["phone"] = phone() if mut == "emaildot": d["email"] = d["email"].replace("@",".dup@",1) if random.random() < 0.3: d["ssn"] = "" # missing field dupes.append(d) customers += dupes random.shuffle(customers) # ============================================================ # 2) POLICIES + CLAIMS (ERP source) -- different schema/naming # ============================================================ policies = [] for i in range(N_POLICY): cust = random.choice(real_cust_ids) start = fake.date_between("-10y", "-30d") policies.append({ "PolicyNo": fake_policy_no(), "CustomerID": cust, # links to CRM.customer_id "LineOfBusiness": random.choice(LINES), "Premium": round(random.uniform(180, 9500), 2), "EffectiveDate": start.strftime("%m/%d/%Y"), # NOTE: different date format than CRM "ExpiryDate": (start + timedelta(days=365)).strftime("%m/%d/%Y"), "Status": random.choice(["active","lapsed","cancelled"]), }) claims = [] for i in range(N_CLAIM): pol = random.choice(policies) fdate = fake.date_between("-5y", "today") claims.append({ "claim_no": fake_claim_no(), "policy_no": pol["PolicyNo"], # links to ERP.PolicyNo "customer_id": pol["CustomerID"], # denormalized link to CRM "line": pol["LineOfBusiness"], "filed_date": fdate.isoformat(), "amount_claimed": round(random.uniform(250, 75000), 2), "status": random.choice(CLAIM_STATUS), "description": fake.sentence(nb_words=12), }) def write_csv(path, rows): if not rows: return with open(path, "w", newline="") as f: w = csv.DictWriter(f, fieldnames=list(rows[0].keys())); w.writeheader(); w.writerows(rows) write_csv(p("structured","crm_customers.csv"), customers) write_csv(p("structured","erp_policies.csv"), policies) write_csv(p("structured","claims.csv"), claims) # ---- Postgres dump (two "source systems", inconsistent schemas) ---- def sql_val(v): if v == "" or v is None: return "NULL" return "'" + str(v).replace("'", "''") + "'" with open(p("structured","postgres_dump.sql"), "w") as f: f.write("-- Synthetic insurance estate (Postgres) -- all PII is FAKE\n") f.write("-- All PII is FAKE (Faker). Two source systems with inconsistent schemas on purpose.\n\n") f.write("DROP TABLE IF EXISTS crm_customers, erp_policies, claims CASCADE;\n\n") f.write("""CREATE TABLE crm_customers (customer_id TEXT, full_name TEXT, email TEXT, phone TEXT, dob TEXT, ssn TEXT, street TEXT, city TEXT, state TEXT, zip TEXT, created_at TEXT);\n""") f.write("""CREATE TABLE erp_policies ("PolicyNo" TEXT, "CustomerID" TEXT, "LineOfBusiness" TEXT, "Premium" NUMERIC, "EffectiveDate" TEXT, "ExpiryDate" TEXT, "Status" TEXT);\n""") f.write("""CREATE TABLE claims (claim_no TEXT, policy_no TEXT, customer_id TEXT, line TEXT, filed_date TEXT, amount_claimed NUMERIC, status TEXT, description TEXT);\n\n""") for c in customers: f.write("INSERT INTO crm_customers VALUES (" + ",".join(sql_val(c[k]) for k in c) + ");\n") for pol in policies: f.write("INSERT INTO erp_policies VALUES (" + ",".join(sql_val(pol[k]) for k in pol) + ");\n") for cl in claims: f.write("INSERT INTO claims VALUES (" + ",".join(sql_val(cl[k]) for k in cl) + ");\n") # ============================================================ # 3) SCANNED CLAIM FORMS (images) + PDFs -- linked to claims # ============================================================ def load_font(size): for fp in ["/System/Library/Fonts/Supplemental/Arial.ttf", "/Library/Fonts/Arial.ttf", "/System/Library/Fonts/Helvetica.ttc"]: if os.path.exists(fp): try: return ImageFont.truetype(fp, size) except Exception: pass return ImageFont.load_default() def cust_by_id(cid): for c in customers: if c["customer_id"] == cid: return c return None doc_claims = random.sample(claims, N_DOCS) doc_index = [] for cl in doc_claims: cust = cust_by_id(cl["customer_id"]) or random.choice(customers) fields = [ ("Claim Number", cl["claim_no"]), ("Policy Number", cl["policy_no"]), ("Claimant Name", cust["full_name"].strip()), ("SSN", cust["ssn"] or "___-__-____"), ("Date of Loss", cl["filed_date"]), ("Line of Business", cl["line"].title()), ("Amount Claimed", f"${cl['amount_claimed']:,.2f}"), ("Phone", cust["phone"]), ("Address", f"{cust['street']}, {cust['city']}, {cust['state']} {cust['zip']}"), ("Description", cl["description"]), ] # ---- PDF ---- pdf = FPDF(); pdf.add_page(); pdf.set_font("Helvetica", "B", 15) pdf.multi_cell(0, 12, f"{COMPANY.upper()} - CLAIM FORM") pdf.ln(2); pdf.set_font("Helvetica", "", 11) for k, v in fields: pdf.set_x(pdf.l_margin) pdf.multi_cell(0, 8, f"{k}: {v}") pdf.output(p("documents","pdf", f"{cl['claim_no']}.pdf")) # ---- "scanned" image (rendered then degraded) ---- W, H = 1000, 1300 img = Image.new("RGB", (W, H), "white"); dr = ImageDraw.Draw(img) title_f, label_f, val_f = load_font(34), load_font(22), load_font(22) dr.rectangle([30,30,W-30,H-30], outline="black", width=2) dr.text((60,55), f"{COMPANY.upper()} - CLAIM FORM", font=title_f, fill="black") dr.line([(60,110),(W-60,110)], fill="black", width=2) y = 150 for k, v in fields: dr.text((70, y), f"{k}:", font=label_f, fill="black") for j, line in enumerate(textwrap.wrap(str(v), 52)): dr.text((330, y + j*26), line, font=val_f, fill=(20,20,40)) y += max(40, 26*len(textwrap.wrap(str(v),52)) + 14) dr.text((70, H-110), "Signature: ", font=label_f, fill="black") dr.line([(220,H-90),(560,H-90)], fill="black", width=2) # scan degradation: grayscale, rotate, noise, blur, brightness img = img.convert("L").rotate(random.uniform(-1.6,1.6), expand=False, fillcolor=255) px = img.load() for _ in range(int(W*H*0.012)): x_, y_ = random.randint(0,W-1), random.randint(0,H-1) px[x_,y_] = random.randint(0,90) if random.random()<0.5 else random.randint(200,255) img = img.filter(ImageFilter.GaussianBlur(0.5)) img.save(p("documents","scans", f"{cl['claim_no']}.png"), optimize=True) doc_index.append({"claim_no": cl["claim_no"], "policy_no": cl["policy_no"], "customer_id": cl["customer_id"], "pdf": f"documents/pdf/{cl['claim_no']}.pdf", "scan": f"documents/scans/{cl['claim_no']}.png"}) # ============================================================ # 4) SUPPORT EMAIL THREADS -- reference real policy/claim ids # ============================================================ TOPICS = ["claim status update", "premium payment question", "policy renewal", "document request", "address change", "complaint about delay", "coverage inquiry"] email_rows = [] for i in range(N_EMAIL): cl = random.choice(claims); cust = cust_by_id(cl["customer_id"]) or random.choice(customers) topic = random.choice(TOPICS); subj = f"{topic.title()} - {cl['claim_no']}" turns = random.randint(1,4); thread = [] body_intro = f"Hi, I'm writing about my {cl['line']} claim {cl['claim_no']} on policy {cl['policy_no']}. " cust_sig = f"\n\nThanks,\n{cust['full_name'].strip()}\n{cust['email']} | {cust['phone']}" # PII in signature for t in range(turns): frm = (cust["email"] if t % 2 == 0 else f"support@{DOMAIN}") body = (body_intro + fake.paragraph(nb_sentences=3) + cust_sig) if t == 0 else \ (("Thank you for reaching out regarding " + topic + ". " + fake.paragraph(nb_sentences=3) + f"\n\n{COMPANY} Support") if frm.startswith("support") else (fake.paragraph(nb_sentences=2) + cust_sig)) thread.append({"from": frm, "to": (f"support@{DOMAIN}" if frm==cust["email"] else cust["email"]), "date": fake.date_time_between("-2y","now").isoformat(), "subject": ("Re: "+subj if t>0 else subj), "message_id": msgid(), "body": body}) # write .eml-ish text with quoted history (messy) lines = [] for t in reversed(thread): lines += [f"From: {t['from']}", f"To: {t['to']}", f"Date: {t['date']}", f"Subject: {t['subject']}", f"Message-ID: {t['message_id']}", "", t["body"], "", "-"*60, "> quoted earlier message" if t is not thread[0] else "", ""] with open(p("emails", f"thread_{i:04d}.txt"), "w") as f: f.write("\n".join(lines)) email_rows.append({"thread_id": f"thread_{i:04d}", "claim_no": cl["claim_no"], "customer_id": cl["customer_id"], "subject": subj, "turns": turns, "file": f"emails/thread_{i:04d}.txt"}) # ============================================================ # 5) CALL-CENTER TRANSCRIPTS -- reference same claims # ============================================================ DISFL = ["um,", "uh,", "you know,", "like,", "I mean,"] call_rows = [] for i in range(N_CALL): cl = random.choice(claims); cust = cust_by_id(cl["customer_id"]) or random.choice(customers) convo = [] convo.append(("AGENT", f"Thank you for calling {COMPANY}, this is {fake.first_name()}. May I have your name?")) convo.append(("CUSTOMER", f"{random.choice(DISFL)} yeah this is {cust['full_name'].strip()}.")) convo.append(("AGENT", "Thank you. Can you verify your phone number and date of birth?")) convo.append(("CUSTOMER", f"Sure, it's {cust['phone']}, born {cust['dob']}.")) # PII in transcript convo.append(("CUSTOMER", f"I'm calling about claim {cl['claim_no']} on my {cl['line']} policy {cl['policy_no']}.")) for _ in range(random.randint(3,7)): spk = random.choice(["AGENT","CUSTOMER"]) line = (random.choice(DISFL)+" " if spk=="CUSTOMER" and random.random()<0.5 else "") + fake.sentence(nb_words=12) convo.append((spk, line)) convo.append(("AGENT", f"Your claim is currently '{cl['status']}'. Is there anything else?")) convo.append(("CUSTOMER", "No that's all, thanks.")) with open(p("transcripts", f"call_{i:04d}.txt"), "w") as f: f.write(f"# Call {i:04d} | claim {cl['claim_no']} | customer {cl['customer_id']}\n\n") f.write("\n".join(f"{s}: {t}" for s,t in convo)) call_rows.append({"call_id": f"call_{i:04d}", "claim_no": cl["claim_no"], "customer_id": cl["customer_id"], "turns": len(convo), "file": f"transcripts/call_{i:04d}.txt"}) # ============================================================ # MANIFEST # ============================================================ manifest = { "dataset": "data-estate", "purpose": "Synthetic, LINKED insurance data estate for data-pipeline / RAG / fine-tuning exercises.", "all_pii_is_fake": True, "linkage_keys": ["customer_id <-> CustomerID", "policy_no <-> PolicyNo", "claim_no"], "deliberate_quality_issues": ["near-duplicate customer rows", "inconsistent date formats across sources", "inconsistent column naming (CRM vs ERP)", "missing fields", "PII embedded in free text (emails, transcripts, forms)"], "counts": {"customers_rows": len(customers), "policies": len(policies), "claims": len(claims), "scanned_forms": len(doc_index), "email_threads": len(email_rows), "call_transcripts": len(call_rows)}, "structured": ["structured/crm_customers.csv","structured/erp_policies.csv","structured/claims.csv","structured/postgres_dump.sql"], "documents_index": doc_index[:5] + ["...(see documents/)"], "emails_index_file": "emails_index.json", "calls_index_file": "calls_index.json", } json.dump(manifest, open(p("MANIFEST.json"),"w"), indent=2) json.dump(email_rows, open(p("emails_index.json"),"w"), indent=2) json.dump(call_rows, open(p("calls_index.json"),"w"), indent=2) json.dump(doc_index, open(p("documents_index.json"),"w"), indent=2) print("DONE:", json.dumps(manifest["counts"], indent=2))