AI Agent for HR: Resume Screening & Candidate Communication Automation

How an AI Agent Accelerates Resume Screening and Improves Candidate Experience

The HR department received 600 resumes in a week after posting a senior developer vacancy. Recruiters work 12 hours a day, but the queue doesn't shrink, and the best candidates get offers from competitors before you can respond. Sound familiar? We built an AI agent that processes a hundred resumes per hour with 89% accuracy, automatically writes personalized rejections and invitations, and schedules interviews via calendar. All without gender or age discrimination—anti-bias filtering is built in by default.

The agent solves three key problems: manual screening (slow and subjective), mass responses (writing each rejection manually is a recruiter's nightmare), and hiring analytics (who dropped out at which stage, which skills are most frequently missing). Below is the technical implementation and a real case from our practice.

HR Agent Components

from pydantic import BaseModel
from typing import Optional, Literal
from openai import OpenAI
import json

client = OpenAI()

class CandidateScreeningResult(BaseModel):
    candidate_id: str
    overall_score: int           # 0-100
    hard_skills_match: int       # % match of hard skills
    experience_match: int        # % match of experience
    red_flags: list[str]         # Stop-factors
    green_flags: list[str]       # Strengths
    recommendation: Literal["strong_yes", "yes", "maybe", "no"]
    next_step: str
    personalized_rejection_reason: Optional[str]

def screen_resume(
    resume_text: str,
    job_description: str,
    required_skills: list[str],
    nice_to_have: list[str],
) -> CandidateScreeningResult:
    """Screen resume against job requirements"""

    response = client.beta.chat.completions.parse(
        model="gpt-4o",
        messages=[{
            "role": "system",
            "content": """You are an experienced recruiter. Objectively assess the candidate's fit for the vacancy.
DO NOT make assumptions—if experience is not explicitly stated, consider it absent.
Be honest in evaluating stop-factors."""
        }, {
            "role": "user",
            "content": f"""Job description:
{job_description}

Required skills: {required_skills}
Nice-to-have skills: {nice_to_have}

Candidate resume:
{resume_text}"""
        }],
        response_format=CandidateScreeningResult,
        temperature=0,
    )

    return response.choices[0].message.parsed

How We Achieve 90%+ Accuracy?

The magic is not in the model but in the prompt and post-processing. The system prompt above prohibits inferring skills—critical for honest screening. Additionally, we run the result through an anti-bias filter and log every call for an audit trail.

Compare: a human reviews 100 resumes in 4.5 hours, the agent does it in 18 minutes. Concordance rate of 89% means the agent agrees with the recruiter in 9 out of 10 cases. Better than a human? No, but 15 times faster. According to LinkedIn Talent Solutions, the average time-to-hire in IT is 35 days.

Order an audit of your hiring funnel—we will select the agent architecture for your stack.

Automated Responses to Candidates

def generate_candidate_response(
    candidate_name: str,
    decision: str,
    position: str,
    feedback: str = None,
) -> str:
    """Personalized response to candidate"""

    templates = {
        "invite_interview": f"""Dear {candidate_name},

Thank you for your interest in the {position} position. We found your experience interesting and would like to invite you for an interview.

Available slots: [CALENDAR_LINK]

The interview will take about 45 minutes. Format: video call.

Best regards,
Recruitment Team""",

        "rejection": None,  # Generate personalized
    }

    if decision == "rejection" and feedback:
        response = client.chat.completions.create(
            model="gpt-4o-mini",
            messages=[{
                "role": "system",
                "content": "Write a polite rejection to the candidate. Tone: respectful, without clichés like 'you are not a fit'. Specify a concrete reason (without humiliating wording)."
            }, {
                "role": "user",
                "content": f"Candidate: {candidate_name}, Position: {position}, Reason: {feedback}"
            }],
        )
        return response.choices[0].message.content

    return templates.get(decision, "")

Batch Screening Pipeline

import asyncio
from typing import List

async def batch_screen_resumes(
    resumes: List[dict],
    job_description: str,
    required_skills: List[str],
    concurrency: int = 10,
) -> List[dict]:
    """Parallel screening of multiple resumes"""

    semaphore = asyncio.Semaphore(concurrency)

    async def screen_single(resume: dict) -> dict:
        async with semaphore:
            result = await asyncio.to_thread(
                screen_resume,
                resume["text"],
                job_description,
                required_skills,
                [],
            )
            return {
                "candidate_id": resume["id"],
                "name": resume["name"],
                "email": resume["email"],
                "screening": result,
            }

    results = await asyncio.gather(*[screen_single(r) for r in resumes])

    # Sort by score
    return sorted(results, key=lambda x: -x["screening"].overall_score)

Practical Case: Hiring 80 Call Center Operators

Task: Hire 80 call center operators in 3 months. Incoming flow: 600+ resumes per week. One recruiter.

Screening Criteria: customer service experience (required), good written communication (required), CRM knowledge (nice-to-have), willingness to work night shifts (required).

Agent Pipeline:

1. Parse incoming resumes from job boards (hh.ru/Avito API)
2. Screen via LLM (50 resumes in 8 minutes vs 4 hours manually)
3. Top 30% → invitation for phone screening
4. Rejections → personalized response automatically
5. After screening → schedule individual interview (Calendly integration)

Results:

• Time to screen 100 resumes: 4.5h (manual) → 18min (agent)
• Concordance rate (agent vs recruiter): 89% (verified on 200 jointly assessed resumes)
• False rejection rate (qualified rejected): 4.1%
• Time-to-hire: 42 days → 28 days
• Recruiter focus: shifted to interviews and onboarding

The client saved $8,000 per month on a second recruiter's salary, the agent took over 70% of the workload. Implementation costs were recouped in two months through reduced time-to-hire.

Legal limitation: the final hiring decision is made by a human. The agent provides a recommendation; the recruiter confirms.

Why Implement an AI Agent?

Get a consultation on implementation—we'll show how the agent fits into your current workflow.

What's Included in AI Agent Development?

Anti-bias Filtering

ANTI_BIAS_PROMPT_ADDENDUM = """IMPORTANT: When evaluating:
- DO NOT consider name, gender, age (if indicated), nationality
- Evaluate only professional competencies and experience
- Do not make assumptions based on personal data
- Apply the same criteria to all candidates"""

Timeline

• HR screening agent: 2–3 weeks
• Integration with job board API (hh.ru, etc.): 1–2 weeks
• Automated responses + calendar: 1 week
• Calibration with recruiter: 1–2 weeks
• Total: 5–8 weeks

Order an audit of your hiring funnel—we'll select the agent architecture for your stack. We'll evaluate your project in 2 days. Contact us by email or Telegram to get a cost and timeline estimate.

Based on technology: LLM OpenAI GPT-4o, LangChain, ChromaDB.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.