Problem

• Scattered candidate data across sourcing websites
• No unified system to track candidate progress
• Manual coordination with external code testing platforms
• Endless back-and-forth emails for scheduling interviews
• Limited visibility into bottlenecks or drop-offs
• Recruiters were spending more time managing operations than evaluating talent.

Solution Exploration

• Hiring additional recruitment coordinators to manage operations
• Building a fully customized ATS with native integrations
• Embedding AI-driven workflow automation into the existing setup

UI Design

How Dezoko Helped

• Unified candidate intake: Applications from all platforms were automatically collected and merged into a single pipeline.
• Smart stage tracking: The AI classified each candidate into pipeline stages like Applied, Code Test, Technical Round, Leadership Round, and Hired.
• Seamless coordination:
• At the Code Test stage, the system sent out test links via integrated vendors.
• At the Technical Round stage, candidates received personalized emails and automated calendar booking links.
• During the Leadership Round, scheduling links were sent based on leadership availability without any manual coordination.
• Once hired, candidate data was pushed to internal systems or CRMs automatically.

Real-Time Analytics and Control

• Number of candidates per stage
• Source-wise conversion and quality
• Time spent at each step of the pipeline
• Interview delays or drop-offs
• Recruiter-wise productivity metrics

Key Outcomes

• All sourcing platforms now feed into one central workflow
• Recruiters spend 40 percent less time on administrative tasks
• Candidate response and scheduling time improved by 60 percent
• Real-time insights replaced manual reports and guesswork
• Candidate experience improved with faster, consistent communication

Tech Stack

• We implemented AI-driven pipeline orchestration through Dezoko using the following technologies:

BigQuery for Centralized Candidate Repository

• We built a data lake using BigQuery as the backbone.
• Every lead, regardless of the source, was cleaned and stored in a unified schema.
• Regular batch jobs processed incoming data from CSVs, email parsers, or APIs into BigQuery.
• Scheduled queries tracked candidate movement between stages.
• BigQuery was a reliable, scalable store and handled analytical queries well.

AI-Powered Stage Intelligence (RAG + Prompt Chaining)

• We implemented Retrieval-Augmented Generation (RAG) for certain automations:
• Use Case: Candidate Reply Drafting
• We used OpenAI with a custom vector store for context on JD, company policy, and FAQs.
• A FastAPI service handled the RAG pipeline:
• Embedding generation via text-embedding-3-small
• Context pulled from BigQuery + Notion docs
• Final prompt fed into GPT to generate personalized email replies.
• Challenge: For niche tech roles, responses sometimes missed the tone. We added a manual approval checkpoint.

Automated Calendar Scheduling Pipelines

• Code test stage triggered an API call to HackerRank or internal test system.
• On technical round movement, the pipeline auto-generated:
• Email with candidate summary
• Calendly link with interviewer availability
• Booking confirmation updated the candidate status in BigQuery and notified recruiters via Slack.
• These workflows reduced scheduling time by 60%.

CRM & Dashboard Integration

• All interactions were logged to BigQuery and surfaced in Looker Studio dashboards.
• Dashboards showed:
• Stage-wise drop-off rates
• Time spent at each stage
• Top sources by hire ratio
• Recruiter productivity

Pipelines That Did Not Work Well (and Why)

• Resume Screening via GPT
• Initial pipeline parsed resumes and used GPT-4 to generate match scores against JD.
• Accuracy was poor for non-standard resumes. Many false positives.
• Fix: We replaced it with rule-based screening + human shortlist + GPT for summary.
• End-to-End Auto-reply without Context
• Tried auto-replying to every inquiry using GPT with generic prompts.
• Lacked personalization. Backfired in some cases.
• Fix: Switched to semi-automated replies with pre-filled drafts and human editing.

Pipelines That Performed Well

• Code test automation
• Interview scheduling with Slack/Calendar integration
• Candidate pipeline reporting (Looker Studio + BigQuery)
• RAG-based reply generation with custom context

Tech Stack Summary

• AI Model: OpenAI
• Embedding Model: text-embedding-3-small
• Retrieval Layer: Pinecone + BigQuery
• Backend: FastAPI, Flask, Node.js workers
• Data Store: Google BigQuery
• Scheduler: Cloud Functions + Pub/Sub
• Dashboard: Looker Studio
• Code Test Integration: HackerRank API
• Calendar: Google Calendar API, Calendly

Outcomes with Technical Impact

• Reduced average hiring time by 6 days
• 40% decrease in manual effort
• Real-time reporting improved leadership visibility
• Human-in-the-loop ensured high-quality responses
• Scalable BigQuery backend enabled performance and cost efficiency