RentEdge
In DevelopmentPricing & revenue management SaaS for rent-a-car companies
Role: Solo full-stack developer
Solo · Since Jul 2026 · Private repo
A multi-tenant SaaS platform that automates pricing, competitive intelligence, and revenue management for rent-a-car companies, built as a modular monolith around a deterministic, auditable rules engine.
Problem
Rent-a-car companies price vehicles manually or with static spreadsheets, missing real-time competitive shifts and leaving margin on the table. Existing enterprise tools are expensive, closed, and slow to adapt.
Solution
A modular-monolith pricing engine with a deterministic rules DSL, hard safety clamps (price floors/ceilings that can never be crossed), automatic competitor rate intelligence, and alerting, all multi-tenant from day one with real JWT/RBAC isolation.
Architecture
- Turborepo monorepo: NestJS API, Next.js web, two Python services (rate-shopper and ml-forecasting) and a backup job, each isolated by workload profile
- Deterministic rules engine: condition/action evaluation, conflict resolution by specificity, hard price floors and ceilings that can never be crossed. It also segments price by length of stay and channel and writes the "why this price changed" text from stored rule facts, with no LLM involved
- Rate intelligence through light adapters that call OTA internal APIs directly (and respect robots.txt) instead of driving a browser. ml-forecasting is a stateless FastAPI service (OLS occupancy forecast, log-log price elasticity, MAD anomaly detection); its output is advisory and the rules engine still decides every price
- PostgreSQL + TimescaleDB for price and occupancy series; Redis + BullMQ for repricing events, with bounded job history and retries with backoff so the audit-log queue cannot drop an entry on a transient DB error
- Money as exact decimal end to end, with row-level locks around every price mutation so two concurrent repricing events cannot overwrite each other
Correctness and security
- Multi-tenant JWT/RBAC with tenant isolation enforced in every use case and verified end to end (403 on cross-tenant access); sessions carry live revocation, so deactivating a user kills their existing token
- SSRF guard with real DNS resolution (not string matching) that re-validates and pins each outbound tenant webhook to its checked IP at delivery time, closing the gap where a hostname is re-pointed at a cloud metadata endpoint after passing validation
- One global middleware writes the activity/audit log for every mutating request, instead of per-endpoint instrumentation
How I checked it
- gitleaks, Semgrep, OSV-Scanner and Trivy found an AES-GCM auth-tag-length gap in 2FA secret decryption (Node accepted 4-byte tags) and an exception-filter ordering bug that turned every domain validation error into a 500 in production
- A backtest of the ML occupancy forecast against the deterministic pickup model, on synthetic history because there are no real clients yet. The deterministic model is about twice as accurate on clean data and OLS wins on noisy data, so the ML stays a suggestion and does not set prices
- Nightly Postgres backups to Cloudflare R2, with the restore itself exercised in CI, not just the backup
- Two audit rounds closed the last rules-engine gaps (seasonal demand index and competitor price deltas became first-class rule facts) and fixed all 17 findings of a UX/accessibility audit, including WCAG contrast on every status badge, ARIA tabs with roving tabindex and visible keyboard focus
- A k6 load test against a local API with a synthetic tenant (81 categories, 405k bookings) showed that one dashboard request ran about 500 queries. The indexes were fine, so the cost was the round trips and the 54k booking rows read into Node. After batching the queries, aggregating occupancy in SQL and letting simultaneous requests share one computation, p95 with 20 users went from 32.8 s to 3.4 s and the response stayed the same. One machine ran everything and the data is synthetic
Not done yet
No production users yet and no public demo: the platform is built toward a pilot with a rent-a-car company that has not started. The ML service only gives advisory suggestions.