The problem
Hiring a local service provider usually goes one of two ways. Directory sites such as Yelp, Google and Thumbtack leave the seeker cold-calling strangers with no clean way to compare quotes. Lead-gen platforms sell the same lead to five or more providers and push them to undercut each other. Operators have no structured audit trail for disputes or for setting fair prices.
SkillStreamsConnect has four roles:
- Service seekers describe the work they need, compare quoted prices, confirm a provider and leave a structured review when the work is done.
- Service providers are verified professionals who browse routed leads, claim the ones that fit, quote a price range and coordinate with the seeker in chat.
- Provider applicants submit a profile, certifications and service areas, then wait for admin approval.
- Admins run KPI dashboards, moderate users and reviews, process refunds, tune complexity tiers and send targeted notifications.
What I built
I designed and built SkillStreamsConnect as a job-request marketplace. A seeker describes the job in a guided wizard. The routing engine sends it to the top two matched providers in the area. A provider spends credits, priced by the job's complexity tier, to claim it and quotes a price range. The seeker confirms, the work happens off-platform, and both sides leave reviews. If a provider claims and never makes contact, or the seeker never confirms, the credits are refunded automatically.
The goal of that flow is to reward providers who respond and quote fairly, to prevent bait-and-switch pricing, and to keep the platform out of the service transaction itself.
For seekers
- A five-step request wizard with category-specific questions and ZIP-code service-area validation.
- A dashboard of active, completed and closed jobs, with monthly spending trends and top-rated providers.
- Price confirmation: review quotes from claiming providers, then confirm, reject or negotiate in chat.
- Real-time chat scoped to claimed jobs.
- Reviews on completion: an overall rating plus quality, communication, professionalism and value.
- Refund requests with reason codes, routed to the admin queue.
- A notification center with tabs, bulk actions and per-channel preferences.
For providers
- An earnings dashboard with monthly trends, success and completion rates, and category breakdowns.
- A job marketplace filtered by category, location and complexity tier, with live search.
- Credit-based claims: 1 to 5+ credits per claim, set by an automatically computed complexity tier.
- A min/max price-range quote sent to the seeker for confirmation.
- Tabs for available, claimed, completed and refunded jobs, and spam reporting for low-quality leads.
- Credit packs bought through Stripe Checkout.
- A profile with photo, bio, service areas, certifications and a verified badge.
For admins
- A live KPI dashboard covering users, jobs, revenue, funnels and geographic distribution, with heatmaps and CSV export.
- Moderation: ban, suspend, warn, adjust credits and approve applicants, each written to an immutable audit log.
- Job oversight with manual provider assignment, status overrides and refund approvals.
- A form-question builder with eight input types and scored options.
- Complexity-tier and ZIP-code service-area management.
- A three-step wizard for targeted in-app and email broadcasts.
- Review moderation with flag, hide and restore, and an audit-log viewer with IP, reason and before-and-after metadata.
The public side has a marketing site, help and legal pages, and registration with a password-strength meter, email verification and auth state synced across tabs.
Architecture
Routing engine
A custom service decides which providers see each new job. Matching runs in three tiers: the provider's exact service area, then their base ZIP, then a Haversine radius. Candidates are scored on rating, completed bookings, response rate, distance and recency. When nobody matches, the job falls back to MANUAL_FULFILLMENT_REQUIRED for an admin to assign.
Dynamic forms and complexity scoring
Request forms are assembled at runtime from an admin-managed JobFormQuestion collection. Eight input types with weighted options produce a complexity score normalized to 0–100, which maps to a ComplexityTier and from there to a credit cost. Admins add categories and questions without a code deploy.
Credit ledger
Stripe Checkout sessions fund provider credits. Every credit operation writes to an immutable Transaction ledger that snapshots the balance before and after.
Real-time layer
Socket.IO runs with a JWT-authenticated handshake, per-user rooms for direct delivery and role-based rooms for broadcasts. It also handles online presence and unread counts synced across tabs, with a 25-second keepalive and automatic fallback from WebSocket to long-polling.
Scheduled jobs
A cron runs every five minutes and handles three time-based cases:
- Routing expiry. Jobs left in
SENT_TO_PROVIDERSmove to manual fulfillment. - Auto-completion.
CONTACTEDjobs past their window are closed, with credits refunded if the seeker never confirmed. - Stale-claim refund. Credits are returned when a provider claimed a job but never made contact.
Every action the cron takes is written to the admin audit log.
Hard problems
Webhook replays
Stripe retries a webhook whenever the handler returns anything other than a 2xx. A naive grant on checkout.session.completed would add credits again on every retry. Before granting, the handler looks up an existing Transaction by stripeChargeId.
Signature verification and JSON parsing
Stripe signatures are computed over the raw body, but express.json() consumes the stream. Both webhook endpoints mount express.raw({ type: 'application/json' }) before the global JSON middleware. Signatures verify, and every other route still receives parsed JSON.
Ratings that drifted
Provider ratings were first cached on the User document and fell out of sync after review moderation. The dashboard now recomputes them with Review.aggregate() and writes the result back. The UI shows an em dash when reviewCount === 0, so "no reviews yet" does not read as a bad rating.
A modal that flashed the wrong title
The admin category edit modal briefly showed "Create Category" while fading out, because formMode flipped before the close animation finished. I gave formMode its own state, decoupled from the modal's open state, so it changes only after the animation ends.
Replacing the first pipeline
The first version was built on a Leads → Bookings → Services pipeline that did not match how seekers and providers behaved. In a staged refactor in 2026 I unified it into a single Job model, removing about 5 models, 12 routes and 15 client directories. I kept the transaction enums for backward compatibility and changed only the UI labels.
Data model
| Model | What it holds |
|---|---|
| User | Seekers, providers, admins and applicants: profile, role, credit balance, provider metrics, moderation state |
| Job | The request lifecycle from submission through routing, claim, contact, completion and review |
| Category | A service category with its question bindings and routing configuration |
| JobFormQuestion | The admin-managed form schema: eight input types with weighted options for complexity scoring |
| ComplexityTier | Maps a normalized complexity score to a claim cost, globally or per category |
| ServiceArea | Enabled ZIP codes with coverage notes and provider counts |
| Transaction | Immutable credit ledger with the Stripe session ID and before-and-after balances |
| Review | Overall rating plus quality, communication, professionalism and value |
| Notification | 30+ event types with per-channel delivery preferences |
| Conversation | A chat thread between seeker and provider, scoped to a claimed job |
| Message | A typed chat message with read-receipt state |
| Refund | Seeker refund requests with reason codes |
| AuditLog | Every admin action with admin ID, IP, reason and before-and-after metadata |
| Application | The provider-applicant workflow from submission to approval |
Security
- JWTs signed with HS256 and verified server-side, with ban and suspension status re-checked on every authenticated request.
- bcrypt hashing at 10 rounds, with password complexity rules at signup.
express-mongo-sanitizestrips MongoDB operators from request bodies.- Admin search inputs are regex-escaped before use, to prevent ReDoS, and admin list endpoints accept only whitelisted sort keys.
- Rate limits of 15 requests per 15 minutes per IP on auth and 300 requests per minute globally, plus a configurable daily inquiry cap per seeker.
- Helmet defaults, including HSTS, X-Frame-Options and X-Content-Type-Options.
- A 30-second global request timeout against slowloris-style attacks.
- Multer MIME checks with Sharp re-encoding on image uploads, and
isomorphic-dompurifyon all user content rendered to the DOM.
In production
It runs on an Ubuntu 24.04 VPS, with Nginx handling SSL and WebSocket upgrades and PM2 in cluster mode. An idempotent seed.js provisions categories, complexity tiers, ZIP codes and admin users, and is safe to re-run. Morgan and PM2 logs cover observability.
What it demonstrates
- Owning a product end to end: discovery, design, implementation, deployment and operations.
- Payment handling where Stripe retries and replays cannot double-credit an account.
- Real-time sockets and scheduled jobs running side by side in the same Node process family.
- Replacing a core pipeline with a simpler model when the first design stopped fitting.
- Plain infrastructure choices: PM2 on a VPS, MongoDB for dynamic schemas, Nodemailer instead of a paid email API.