Evidra

A cognitive talent intelligence platform that discovers hidden engineering capabilities by auditing GitHub repositories and extracting technical evidence beyond the resume.

Comment

Inspiration

The technical recruiting industry is fundamentally broken. Traditional Applicant Tracking Systems (ATS) rely on keyword matching, forcing candidates to stuff their resumes with buzzwords while completely ignoring actual engineering capabilities. We realized that the best engineers often have terrible resumes but incredible GitHub repositories, while candidates with perfectly optimized resumes sometimes lack practical system design experience.

We built Evidra to shift the paradigm from claims-based recruiting to evidence-based intelligence. We wanted to build a platform that doesn't just ask "Did they write 'Python' on their resume?", but instead asks, "Do their Git commits demonstrate an understanding of distributed systems?"

What it does

Evidra is an evidence-driven Talent Intelligence Platform that evaluates software engineers using verifiable technical signals derived from resumes, repository activity, and demonstrated engineering work.

  • Cognitive Parsing: It digests traditional resumes and extracts not just skills, but the context in which those skills were applied.
  • GitHub Auditing: It seamlessly connects to the candidate's GitHub, analyzing commit histories, repository architectures, and open-source contributions to verify if their actual code matches their resume claims.
  • Hidden Strength Discovery: Using a multi-stage AI reasoning pipeline, Evidra surfaces evidence-backed hidden strengths (e.g., identifying that a candidate who only listed "Backend Developer" actually has deep expertise in "IoT Architecture" based on their repository structure).
  • Blind Review: It strips all personally identifiable information (PII) to eliminate unconscious bias, presenting recruiters with pure, unadulterated technical capability metrics.

Every recommendation is paired with evidence sources and explainability data, ensuring recruiters understand not only what the system recommends, but why.

Why It's Different

Most recruiting platforms rank candidates based on what they claim. Evidra ranks candidates based on evidence.

The platform combines semantic understanding, repository intelligence, explainable reasoning, and blind-review workflows to surface overlooked technical talent.

How we built it

We architected Evidra as a decoupled, microservices-oriented platform prioritizing speed, scalability, and AI integration:

  • Backend Intelligence: Built entirely on Python and FastAPI for high-performance asynchronous execution. We integrated SQLModel and PostgreSQL with the pgvector extension to handle high-dimensional vector storage for our semantic search engine.
  • Cognitive Engine: We implemented a dual-LLM architecture. We utilize Featherless (Qwen2.5-7B-Instruct) and Groq (Llama-3) for rapid, structured JSON extraction of candidate narratives, and SentenceTransformers (all-MiniLM-L6-v2) to generate dense vector embeddings for intelligent candidate matching.
  • Frontend Architecture: The user interface is crafted using Vanilla JavaScript, HTML5, and CSS3. We intentionally avoided heavy frontend frameworks, utilizing modern CSS Grid, Flexbox, and Glassmorphism for a premium, native-feeling user experience.
  • Deployment: We deployed the edge-optimized frontend on Vercel and proxy API requests directly to our Render-hosted backend, ensuring secure and seamless cross-origin communication.

Challenges we ran into

One of our toughest engineering challenges was Semantic Hallucination Management. Early iterations of our LLM pipeline would hallucinate skills that candidates never possessed because the model was trying to be overly helpful. We solved this by implementing strict JSON-schema enforcement and a multi-pass reasoning loop that forces the AI to cite specific evidence sources for every skill it extracts.

Additionally, managing Render Port Binding Timeouts during cold-starts was a significant DevOps hurdle. Downloading a 90MB Hugging Face embedding model during the Uvicorn boot sequence consistently crashed our container. We engineered around this by deferring the model load out of the FastAPI lifespan hook and lazy-loading the transformer only upon the first user interaction, reducing boot time from over 60 seconds to under 1 second.

Accomplishments that we're proud of

We are incredibly proud of the Authenticity Index. By successfully bridging an LLM's natural language understanding with deterministic GitHub repository signals, we created a metric that helps identify gaps between claimed expertise and publicly available technical evidence. Seeing the platform successfully surface these insights was our "Eureka!" moment.

We're also proud of the UI. Building a fluid, dynamic, glassmorphism interface entirely from scratch without React allowed us to achieve highly optimized performance through a lightweight frontend architecture.

What we learned

  • Vector Databases: We mastered pgvector, learning how to index and query dense embeddings inside PostgreSQL using cosine similarity instead of relying on expensive managed vector databases.
  • Prompt Engineering & Structured AI Outputs: We learned that carefully engineered prompts, schema validation, and evidence-first reasoning dramatically improve reliability compared to free-form AI responses. Designing prompts around strict JSON schemas reduced hallucinations, improved consistency, and made downstream processing significantly more robust.
  • Explainable AI Design: We discovered that recruiters trust AI recommendations far more when every conclusion can be traced back to observable evidence sources and reasoning steps rather than opaque scores.
  • Asynchronous Python: We pushed our understanding of asyncio and httpx to their limits to ensure GitHub audits didn't block the main event loop.

What's next for Evidra

This is just the foundation. Our immediate roadmap includes:

  • Multi-Platform Auditing: Expanding beyond GitHub to index Stack Overflow activity and Kaggle projects.
  • Enterprise SSO: Integrating SAML/OAuth2 for seamless HR department onboarding.
  • Automated Technical Interviews: Using candidate embeddings to generate personalized technical interview questions tailored to verified skill gaps.

Evidra isn't just a hackathon project; it's a step toward more explainable and evidence-driven technical hiring.

Built With

Share this project:

Updates