The Business Challenge
The business situation the organisation faced before the project began.
Manual CV screening and candidate evaluation create significant bottlenecks in HR workflows. HR teams spend excessive hours manually collecting resumes, parsing unstructured PDF data, scoring candidate qualifications against job criteria, and reformatting resumes into standardized corporate templates. Additionally, lack of automated deployment and observability in model workflows leads to slow iteration cycles and potential downtime.
Why AI Was the Right Solution
Why AI was the right approach — and what alternatives were considered.
High variability in candidate resume layouts made traditional rule-based parsers unreliable and prone to failure. LLM agents provided the necessary semantic understanding to reason over diverse, unstructured CV text, accurately map candidate experience against strict job rubrics, and dynamically synthesize structured profile data into standardized corporate Word documents.
How the Solution Was Delivered
Discovery, design, development, testing and rollout — the journey, not the tooling.
- 1Discovery & requirements gathering with HR team to define scoring parameters and document formats
- 2Data pipeline design for form submission, PDF parsing, and Google Sheets integration
- 3Prototype development using n8n workflows and LLM agent nodes for automated entity extraction
- 4Local RAG implementation using Qdrant vector database and Ollama embeddings for HR knowledge retrieval
- 5Kubernetes pipeline setup featuring Blue/Green gated deployment and Prometheus/Grafana monitoring
- 6Shadow-mode testing, output evaluation, and final production rollout with automated CV generator
Key Technical & Architecture Decisions
Architecture, model selection, workflow and trade-offs.
Implemented a Retrieval-Augmented Generation (RAG) approach using Qdrant and local embeddings over model fine-tuning to ensure strict data privacy and grounded responses. Selected a Kubernetes-based microservice architecture with Blue/Green deployment gates and automated metric tracking via Prometheus and Grafana to balance high availability with zero-downtime model updates.
Challenges & How They Were Solved
Obstacles hit along the way and how they were overcome.
Unstructured PDF parsing initially yielded inconsistent data across varied candidate CV layouts. This was resolved by implementing structured output JSON parsers within the LLM agent pipeline and establishing continuous evaluation quality gates to ensure high extraction accuracy before writing data to target systems.
User Adoption & Change Management
How users responded — training, change management and feedback loops.
Introduced the system through a two-week shadow mode alongside manual HR processes. Recruiters validated AI-generated candidate scores against their manual assessments, providing feedback that refined the prompt rubrics and built strong organizational trust prior to full automated adoption.
Business Outcomes & Impact
The measurable outcomes from the underlying AI Deployment — with the story behind them.
Automated CV processing and candidate scoring by 80%, reducing candidate evaluation turnaround time from days to minutes while standardizing 100% of generated candidate profiles.
Business Outcome Categories
The story behind the numbers
By eliminating repetitive resume parsing and manual formatting, the HR team shifted their focus from operational overhead to strategic talent acquisition. The standardized CV output ensured complete profile consistency for external presentation, while the Local RAG integration empowered recruiters to query internal job requirements instantly without manual spreadsheet searches.
Lessons Learned
What surprised the team, what worked well, and what would be done differently.
- Surprised by how quickly strict JSON output parsers resolved complex PDF extraction edge cases.
- Would implement continuous evaluation testing (eval sets) earlier in the dev lifecycle.
- Local RAG with Qdrant vector store worked remarkably well for accurate knowledge retrieval without latency.
- Automated Blue/Green deployment gates ensured zero downtime during prompt updates.
- Standardized document generation saved significantly more operational hours than initial estimates predicted.
Future Opportunities
Where this solution could go next.
Expand the multi-agent workflow to automate initial candidate screening interviews via conversational AI, and integrate closed-loop feedback from hiring decisions to continuously refine candidate scoring criteria over time.
Professional Reflections
Guidance for another builder tackling a similar problem.
Investing early in robust evaluation sets and local vector retrieval yields far higher long-term reliability than prompt tuning alone. Furthermore, pairing AI agents with production-grade observability (Grafana/Prometheus) is essential for maintaining pipeline stability in enterprise environments.