codeproofed

A technical recruitment platform and bot immunity using automated sandboxes and qualitative assessment with the Gemini API

Comment

📖 CodeProofed | The Project Story

This is the official narrative of how CodeProofed was conceived, designed, built, and optimized during the hackathon.

💡 Inspiration: The Broken Gate of Tech Recruiting

The current engineering recruitment ecosystem is experiencing a severe structural crisis. On one side, job applications are flooded by automated mass-application bots that submit hundreds of inflated, copy-pasted, or AI-generated resumes in milliseconds. On the other side, companies attempt to defend themselves by using rigid, static keyword-based Applicant Tracking Systems (ATS).

This creates a tragic double-loss:

  1. Recruiters are buried under oceans of spam applications, making it nearly impossible to find real, qualified developers.
  2. Talented Developers are routinely filtered out by automated bots because their resumes lack exact keyword matches, denying them a fair chance to demonstrate their actual coding abilities.

CodeProofed was born to break this cycle. Our vision was to create a platform that grants technical immunity to companies against spam bots while guaranteeing fair, interactive, and comprehensive evaluations to human programmers.

🛠️ How We Built It: Architecture & Stack

We designed CodeProofed as a highly modular, secure, and cost-effective B2B SaaS platform.

graph TD
    A[Frontend: HostGator Shared Hosting] -->|Authentication & Data| B[Firebase Auth & Firestore]
    A -->|Chat Interaction & Analysis| C[Gemini API]
    A -->|Monaco Editor| D[Cloud Functions / Sandbox API]
    D -->|Isolated Micro-containers| E[E2B.dev / Piston]
  style A fill:#1e1b4b,stroke:#818cf8,stroke-width:2px;
  style B fill:#0c0a09,stroke:#e7e5e4,stroke-width:2px;
  style C fill:#030712,stroke:#a21caf,stroke-width:2px;
  style D fill:#030712,stroke:#c084fc,stroke-width:2px;
  style E fill:#030712,stroke:#f43f5e,stroke-width:2px;

The Stack:

  • UI/UX Layer: Developed in React.js + Tailwind CSS v4, utilizing Outfit & Inter typography for a premium, high-contrast dark mode dashboard experience. Hosted on a shared Apache environment (HostGator) to minimize fixed web server costs.
  • Serverless Backend: Built using Firebase Cloud Functions (Node.js) to orchestrate data handling, security validations, and API handshakes.
  • Data & Session Persistence: Firebase Authentication secure flows combined with Cloud Firestore for real-time storage of candidate state tracking, challenges, and qualitative scores.
  • Code Sandbox: Integrates isolated cloud container runners (E2B.dev and Piston) to securely execute candidate code inputs against automated test cases.
  • AI Evaluation Engine: Integrated the Google Gemini API to drive the dynamic soft-skills chat recruiter and analyze clean code quality metrics.

🧬 Mathematical Modeling (LaTeX Support)

To design a platform that is both technically sound and financially viable, we developed mathematical models for evaluation scoring and API token cost control.

1. The Recruiter Combined Score Model

A candidate's final ranking is determined by a weighted average of their technical performance ($S_{\text{tech}}$) and their soft skills/communication capabilities ($S_{\text{soft}}$):

$$S_{\text{combined}} = w_{\text{tech}} \cdot S_{\text{tech}} + w_{\text{soft}} \cdot S_{\text{soft}}$$

Where $w_{\text{tech}}$ and $w_{\text{soft}}$ are customizable weights defined by the recruiter based on the requirements of the job opening, satisfying:

$$w_{\text{tech}} + w_{\text{soft}} = 1.0 \quad \text{where} \quad w_{\text{tech}}, w_{\text{soft}} \ge 0$$

2. Cognitive vs. Mechanical API Cost Optimization

To prevent spam bots from blowing up our Gemini API budget, we designed a mechanical-first gate. Candidates must first pass isolated, local test cases before their code is sent to Gemini for qualitative analysis (Big O, structure, readability).

Let:

  • $C_{\text{mech}}$ be the cost of running code in the isolated sandbox.
  • $C_{\text{cog}}$ be the token cost of a qualitative Gemini API assessment.
  • $P_{\text{pass}}$ be the probability of an application passing the automated test cases.

The expected evaluation cost per application, $E[C]$, is modeled as:

$$E[C] = C_{\text{mech}} + P_{\text{pass}} \cdot C_{\text{cog}}$$

Since spam bots and invalid scripts fail the mechanical test cases immediately ($P_{\text{pass}} \approx 0.15$ in pilot runs), the expected cost is dramatically optimized:

$$E[C] \approx C_{\text{mech}} + (0.15) \cdot C_{\text{cog}}$$

This strategy yields a 85%+ reduction in AI token costs, preserving profit margins.

3. SaaS Break-Even Equation

Given our annual fixed hosting costs ($C_{\text{fixed}}$) and the variable token cost per active customer subscription ($C_{\text{var}}(N)$), the annual net profit ($\Pi$) as a function of active clients ($N$) charging a monthly subscription ($P_{\text{sub}}$) is:

$$\Pi(N) = 12 \cdot N \cdot P_{\text{sub}} - \left( C_{\text{fixed}} + C_{\text{var}}(N) \right)$$

By defining the variable cost linearly as $C_{\text{var}}(N) = N \cdot V$, where $V$ is the yearly API token overhead per client, the break-even volume of corporate clients ($N_{\text{BE}}$) is solved by:

$$N_{\text{BE}} = \left\lceil \frac{C_{\text{fixed}}}{12 \cdot P_{\text{sub}} - V} \right\rceil$$

Using our active pricing ($P_{\text{sub}} = \$1,500\text{ MXN/month}$) and infrastructure parameters ($C_{\text{fixed}} = \$767\text{ MXN/year}$):

$$N_{\text{BE}} = \left\lceil \frac{767}{12 \cdot 1500 - 300} \right\rceil = 1 \text{ Client}$$

⚡ Challenges We Faced & Overcame

  1. Vulnerability in Sandboxes (Code Injection):
    • Challenge: Running arbitrary code submitted by candidates is highly dangerous. A candidate could attempt a fork bomb, memory exhaustion, or network scanning from inside our runner.
    • Solution: We isolated the execution entirely by using micro-containers hosted in the cloud via E2B. Every run spins up an ephemeral, resource-constrained container with strict timeouts (2 seconds max) and disabled local network access.
  2. React Dynamic Rendering Latency:
    • Challenge: Rendering complex code highlight blocks and live chat inputs created minor input lag on lower-spec mobile devices.
    • Solution: We optimized state propagation in React, lazy-loaded the heavy Monaco Editor component, and implemented debounced text changes to guarantee a fluid 60fps interaction.

🧠 What We Learned

  • Token Economics (Tokenomics): We learned that building with AI requires careful cost-minimization logic. Moving cognitive checks behind mechanical validations is essential for scaling B2B SaaS systems.
  • Asynchronous Chat Loops: Designing prompt architectures that prevent LLMs from hallucinating or falling out of character requires strict systemic instruction structures.
  • Deployment on Hybrid Infrastructures: Deploying a static React frontend to shared hosting (HostGator) while routing backend logic to serverless structures (Firebase) taught us how to configure domain-level routing and cross-origin resource sharing (CORS) rules.

🔮 What's Next for CodeProofed?

  1. Integration with Enterprise ATS: Building custom connectors for Greenhouse, Workday, and Lever so recruiters can access their top 5 shortlists without leaving their daily workflow.
  2. Multi-file Project Sandboxing: Allowing candidates to solve complex, multi-module challenges (such as React components or microservices) in an expanded editor directory layout.
  3. Advanced Plagiarism Detection: Integrating semantic analysis to determine if a candidate copy-pasted answers from external AI interfaces during their technical code run.

Built With

Share this project:

Updates