Case Studies

SmartBid: AI-Powered Upwork Job Intelligence

A human-in-the-loop platform that scores Upwork jobs across quality, competition, employer, and fit signals — built as an applied AI/ML portfolio project.

OVERVIEW

Helping Freelancers Decide Before They Write

SmartBid is an AI/ML learning project built as a real product: a marketplace intelligence platform that helps freelancers decide which Upwork jobs are actually worth applying to. The problem is simple — freelancers spend too much time guessing — and the answer is measurement, not automation.

The platform analyzes each job opportunity across five signals: job quality, competition, employer quality, employer engagement, and personal fit. It then supports the freelancer with an LLM-assisted proposal draft they review, edit, and submit themselves.

The goal is not to automate the freelancer out of the process. It is to help them apply better judgment before they spend time writing a proposal.

The Challenge

Freelancers on Upwork face a noisy, competitive marketplace with almost no decision support:

  • Hundreds of new job posts a day, with no reliable way to judge which are worth pursuing

  • No visibility into competition levels or an employer's hiring track record

  • Proposal writing is slow, so every bad application is expensive

  • Generic AI auto-apply tools violate platform rules and produce spam, not signal

The product challenge was to build ranking and scoring that a freelancer can trust — explainable, compliant with Upwork's Terms of Service, and always leaving the human in control.

Links

APPROACH

The Approach

I designed and built SmartBid end to end as a compliance-first, human-in-the-loop product:

  • A scoring framework that rates every job on quality, competition, employer quality, employer engagement, and personal fit (Job Fit Score)

  • Analytics pipelines in BigQuery and dbt that turn raw job and employer data into the marts behind each score

  • An LLM-assisted drafting workflow on Vertex AI that generates personalized proposal drafts from the freelancer's profile and onboarding answers

  • Prompt contracts with versioning and evaluation, so generation behavior is deterministic and inspectable rather than a black box

  • Explainable recommendations: every score surfaces the context behind it, so the user understands the why, not just the number

  • Strict non-goals: no auto-submission, no scraping, no stored Upwork credentials — official API only, and the freelancer submits every proposal manually

Core architecture:

  • Next.js web application for job discovery, scoring views, and proposal drafting

  • PostgreSQL for application data; BigQuery as the analytical warehouse

  • dbt marts powering the scoring framework and job insights

  • Vertex AI for the proposal-generation pipeline, with versioned prompt contracts

  • Upwork's official GraphQL API as the sole, read-only data source

  • Separate dev and prod environments on Google Cloud Platform with validated deployments

  • Tiered Free / Starter / Pro subscription model with entitlements and paywalls

Technical Stack

  • Vertex AI

  • BigQuery

  • dbt

  • Next.js

  • PostgreSQL

  • Python

  • Upwork API

  • Google Cloud Platform

RESULTS

Key Outcomes

  • A working end-to-end marketplace intelligence platform, from data ingestion to scored job feed to AI-assisted drafts

  • Five explainable scoring signals covering job quality, competition, employer quality, employer engagement, and personal fit

  • A BigQuery/dbt analytics layer with documented metrics specs and dbt marts

  • A prompt versioning and evaluation workflow for the LLM drafting pipeline

  • A compliance-first design: official API only, read-only, human control at every step

Why This Matters

SmartBid is primarily a learning vehicle for how real marketplace intelligence products work — ranking and scoring systems, analytics pipelines, LLM workflows, and human-in-the-loop product design. Building it as a genuine product, with real constraints, made the lessons concrete:

  • Ranking is a product problem as much as a modeling problem: scores must be explainable to be trusted

  • Compliance constraints (official APIs, no automation) shape the architecture, not just the legal page

  • Prompt versioning and evaluation turn LLM features from demos into dependable product surfaces

One principle kept recurring: AI products are more interesting when they help people apply better judgment, not bypass judgment entirely.

Looking Forward

Planned next steps:

  • Personalized job notifications and saved-search alerts

  • Job tracking and outcome insights (interviews and contracts won)

  • Profile optimization recommendations

  • Richer personalization from like/dislike/save signals in the job feed

  • An expanded evaluation harness for prompts and scoring quality

Interested in building something similar?

We help organizations design scalable AI and analytics infrastructure for:

  • forecasting

  • growth analytics

  • operational automation

  • machine learning platforms

  • agentic AI workflows

  • cloud-native data systems

Let’s build systems that turn data into operational leverage.