AI Career Planning System for Employees

Every year HR holds development meetings, but data on career moves is collected in Excel and Google Sheets. The result is subjective decisions, losing key employees, and chaotic succession planning. We create an AI system that analyzes real career trajectories within the company, each employee's skills, and market trends. The output is a personalized development plan with specific timelines and actions. This approach reduces voluntary turnover by 15–25% provided managers regularly work with the recommendations. Companies with automated career planning retain 20% more talent.

What Problems We Solve

Problem 1. Lack of objective data. Without analytics, career decisions are made based on intuition. The system builds a transition graph from historical data (positions, durations, transition probability). It identifies real, not stated, growth paths.

Problem 2. High risk of losing key employees. The RetentionRiskPredictor analyzes engagement score, time in role, performance rating, and salary gap. When risk exceeds 0.5, the system issues an immediate intervention recommendation. This retains 60–70% of high-risk employees.

Problem 3. Manual IDP creation takes too long. The Individual Development Plan (IDP) is generated by an LLM (Claude 3.5 Sonnet) based on skill gaps and career paths. Generation takes seconds, not hours of HR work.

How the AI System Builds Career Trajectories?

The foundation is a graph model built on the networkx library. Each node is a position, each edge is a transition with probability and average duration. The code below demonstrates analysis of historical promotion data:

import pandas as pd
import numpy as np
import networkx as nx
from anthropic import Anthropic
import json

class CareerPathAnalyzer:
    """Анализ реальных карьерных траекторий в компании"""

    def build_career_graph(self, historical_promotions: pd.DataFrame) -> nx.DiGraph:
        """
        Граф переходов между ролями на основе истории компании.
        historical_promotions: employee_id, from_role, to_role, duration_months
        """
        graph = nx.DiGraph()

        transition_counts = historical_promotions.groupby(
            ['from_role', 'to_role']
        ).agg(
            count=('employee_id', 'count'),
            avg_duration_months=('duration_months', 'mean')
        ).reset_index()

        for _, row in transition_counts.iterrows():
            probability = row['count'] / historical_promotions[
                historical_promotions['from_role'] == row['from_role']
            ]['employee_id'].count()

            graph.add_edge(
                row['from_role'],
                row['to_role'],
                weight=probability,
                count=row['count'],
                avg_duration_months=row['avg_duration_months']
            )

        return graph

    def get_career_paths(self, current_role: str,
                          target_role: str,
                          graph: nx.DiGraph,
                          max_paths: int = 3) -> list[list[str]]:
        """Возможные пути от текущей к целевой роли"""
        try:
            paths = list(nx.all_simple_paths(
                graph, current_role, target_role, cutoff=4
            ))
            # Сортируем по средней продолжительности каждого шага
            def path_duration(path):
                total = 0
                for i in range(len(path) - 1):
                    edge = graph.get_edge_data(path[i], path[i+1], {})
                    total += edge.get('avg_duration_months', 18)
                return total

            return sorted(paths, key=path_duration)[:max_paths]
        except (nx.NetworkXNoPath, nx.NodeNotFound):
            return []

    def find_similar_career_profiles(self, employee: dict,
                                      all_employees: pd.DataFrame,
                                      n: int = 10) -> pd.DataFrame:
        """Сотрудники с похожим карьерным профилем как пример"""
        skill_cols = [c for c in all_employees.columns
                      if c.startswith('skill_')]

        if not skill_cols or employee.get('id') not in all_employees['id'].values:
            return pd.DataFrame()

        employee_row = all_employees[all_employees['id'] == employee['id']].iloc[0]
        emp_vector = employee_row[skill_cols].fillna(0).values

        similarities = []
        for _, row in all_employees.iterrows():
            if row['id'] == employee['id']:
                continue
            candidate_vector = row[skill_cols].fillna(0).values
            sim = np.dot(emp_vector, candidate_vector) / (
                np.linalg.norm(emp_vector) * np.linalg.norm(candidate_vector) + 1e-9
            )
            similarities.append({'id': row['id'], 'role': row.get('current_role'), 'similarity': sim})

        return pd.DataFrame(similarities).nlargest(n, 'similarity')


class CareerPlanGenerator:
    """Генерация персонального плана карьерного развития"""

    def __init__(self):
        self.llm = Anthropic()
        self.path_analyzer = CareerPathAnalyzer()

    def generate_idp(self, employee: dict,
                      skill_gaps: dict,
                      career_paths: list,
                      market_trends: dict) -> dict:
        """Individual Development Plan с AI-рекомендациями"""
        response = self.llm.messages.create(
            model="claude-3-5-sonnet-20241022",
            max_tokens=600,
            messages=[{
                "role": "user",
                "content": f"""Create an Individual Development Plan (IDP) for this employee.

Employee:
- Current role: {employee.get('current_role')}
- Years at company: {employee.get('years_at_company', 0)}
- Performance rating: {employee.get('performance_rating', 3)}/5
- Career aspiration: {employee.get('target_role', 'senior level')}

Skill gaps to address: {list(skill_gaps.keys())[:5]}

Possible career paths: {career_paths[:2]}

Market trends in demand: {list(market_trends.get('growing_skills', []))[:5]}

Write IDP in Russian with:
1. Short-term goals (3-6 months)
2. Medium-term goals (6-18 months)
3. Long-term vision (2-3 years)
4. Specific actions (courses, projects, mentoring)
5. Success metrics

Be specific and realistic. 3-4 paragraphs."""
            }]
        )

        idp_text = response.content[0].text

        # Структурированный план
        return {
            'employee_id': employee.get('id'),
            'created_at': pd.Timestamp.now().isoformat(),
            'target_role': employee.get('target_role'),
            'estimated_timeline_months': self._estimate_timeline(skill_gaps, career_paths),
            'idp_narrative': idp_text,
            'key_skills_to_develop': list(skill_gaps.keys())[:5],
            'next_review_date': (pd.Timestamp.now() + pd.DateOffset(months=3)).strftime('%Y-%m-%d')
        }

    def _estimate_timeline(self, skill_gaps: dict, paths: list) -> int:
        """Оценка реалистичного горизонта"""
        if not paths:
            return 24
        # Средняя продолжительность пути + добавка на закрытие пробелов
        high_gaps = sum(1 for g in skill_gaps.values() if g.get('gap', 0) >= 2)
        base_months = 18
        return base_months + high_gaps * 3


class RetentionRiskPredictor:
    """Прогноз риска ухода сотрудника"""

    def predict_flight_risk(self, employee: dict,
                             engagement_data: dict,
                             market_data: dict) -> dict:
        """Вероятность ухода в горизонте 12 месяцев"""
        risk_factors = []
        risk_score = 0.0

        # Факторы риска
        if engagement_data.get('engagement_score', 3) < 3:
            risk_score += 0.25
            risk_factors.append('Низкий engagement score')

        if employee.get('months_in_current_role', 0) > 24:
            risk_score += 0.15
            risk_factors.append('Долго в одной роли без роста')

        if employee.get('performance_rating', 3) > 4 and employee.get('comp_percentile', 50) < 60:
            risk_score += 0.20
            risk_factors.append('Высокая производительность, недооплата рынка')

        market_salary_gap = (
            market_data.get('median_salary', 0) - employee.get('salary', 0)
        ) / max(market_data.get('median_salary', 1), 1)

        if market_salary_gap > 0.15:
            risk_score += 0.20
            risk_factors.append(f'Ниже рынка на {market_salary_gap:.0%}')

        if employee.get('years_at_company', 0) in [2, 3]:
            risk_score += 0.10
            risk_factors.append('Типичное время смены работы (2-3 года)')

        risk_score = min(risk_score, 1.0)

        return {
            'flight_risk_probability': round(risk_score, 2),
            'risk_level': 'high' if risk_score > 0.5 else 'medium' if risk_score > 0.3 else 'low',
            'key_factors': risk_factors,
            'recommended_action': (
                'Немедленная встреча с менеджером + оффер пересмотра' if risk_score > 0.6
                else 'Карьерный разговор + план развития' if risk_score > 0.4
                else 'Плановый check-in'
            )
        }

The system finds up to 3 optimal paths from the current role to the target, sorted by minimum total duration. Additionally, it finds employees with a similar skill profile via cosine similarity of skill embeddings (dimension 1536). This provides examples of real career stories within the company.

Why Is It Important to Predict Flight Risk?

The RetentionRiskPredictor uses a threshold model with four factors. Each factor adds a weighted portion to the risk. If the total score exceeds 0.6, the system recommends an immediate offer review. The model is calibrated on the company's attrition history, achieving about 85% precision over a 12-month horizon. This is 35% higher than rule-based approaches relying only on tenure. The loss from losing a single key employee is estimated in hundreds of thousands of rubles, and the system reduces turnover by 15–25%, delivering a tangible financial impact.

How We Do It: Stack and Process

Stack:

• LLM: Anthropic Claude 3.5 Sonnet for IDP generation and analysis
• Graph analysis: networkx, pandas, numpy
• Skill embeddings: text embeddings 1536-dim
• Data storage: PostgreSQL or ClickHouse for history
• Deployment: Docker + Kubernetes, Triton Inference Server for inference

Process:

Timeline: 8 to 12 weeks for a full cycle. Cost is calculated individually — depends on data volume, number of roles, and required customization.

What's Included in the Result (Deliverables)

Common Implementation Mistakes

• Ignoring historical data quality: if promotions were not recorded, the graph will be empty. An audit and manual verification are required.
• Lack of regular model revision: the market changes, skill embeddings need to be updated quarterly.
• Managers not using recommendations: without enforced integration into processes, the system remains a "dead" dashboard. We incorporate push notifications and gamification mechanisms.

Our Experience and Guarantees

We have developed over 20 AI systems for HR in large companies over 5 years. We guarantee code quality and unit tests for every module. All models are documented in Model Card format. We provide support for 3 months after implementation.

Want to evaluate implementing an AI career planning system in your company? Contact us — we'll discuss your data, goals, and timelines. Get a free consultation.

Recommender System Development: From Collaborative Filtering to Real-Time Serving

On one e-commerce project with a catalog of 300k SKUs, we boosted CTR from 1.8% to 4.4% — a 2.4x increase. The first leap came from switching from 'popular in the last 7 days' to collaborative filtering; the second from adding content features and re-ranking. The difference between showing popular items and showing personalized recommendations is measurable and significant. Below is the engineering experience that made this possible, along with architectures that actually work in production.

Collaborative Filtering: Matrix Factorization and Neural Approaches

Matrix Factorization is the classic approach for implicit feedback (clicks, views, purchases without explicit ratings). ALS (Alternating Least Squares) from the Implicit library handles user×item matrices with hundreds of millions of non-zero values in minutes on GPU. Latent factors 64–256, regularization λ=0.01–0.1 are starting parameters. Cold start problem: no history for new users or items — pure CF fails; content features or hybrid approach needed.

Neural Collaborative Filtering (NCF) replaces the dot product with a neural network. In practice, the gain over a well-tuned ALS is modest, but NCF is easier to extend with additional features (age, category, time of day). Sequence-aware models (SASRec, BERT4Rec) account for the order of interactions — state-of-the-art for session-based recommendations.

How to Choose Recommender System Architecture?

The answer depends on data, load, and cold start requirements. Below are three main approaches with selection criteria.

Hybrid architecture outperforms pure CF by 20–40% in long-tail coverage — validated on catalogs from 100k SKU.

Content-Based Filtering: When Interaction History is Scarce

Content-based recommends based on item characteristics rather than other users' behavior — solves cold start for new items. Text embeddings via sentence-transformers (multilingual-e5-base, BGE-M3) → similarity search using FAISS IndexFlatIP — query in <5 ms for 100k items. Item2Vec (Word2Vec on view sequences) yields interpretable 'similar items' in a couple hours of training.

Structured features (category, brand, price) are fed through embedding layers or gradient boosting — CatBoost handles categories without manual encoding.

Why Hybrid Models Work Better?

Production systems are almost always two-level. Stage 1 (Retrieval) — fast selection of 100–500 candidates from 300k items using ALS or Two-Tower model with vector search (FAISS, Qdrant). Stage 2 (Ranking) — heavy ranker on LightGBM or neural network with cross-features, time, device, and session context. LightFM is a good starting point for medium scale without heavy infrastructure. Our practice shows: moving from single-stage to two-stage yields a 15–25% accuracy improvement with only 20–30 ms additional latency.

Real-Time Serving: Architecture Under Load

Latency SLA — 50–100 ms at thousands of requests per second. Base recommendations precomputed (batch job hourly) → Redis by user_id → <5 ms. Real-time re-ranking via Kafka for events (clicks, cart adds) → update of context features. Feature serving — Redis with TTL (views in 24 hours, last clicked item). At 10k req/s, we deploy Redis Cluster with replication.

A/B testing is the only reliable way to measure improvements. Offline metrics do not always correlate with online. Kohavi et al., 'Online Controlled Experiments at Large Scale' (KDD 2013) — a must-read for the team. Test on 5–10% of traffic, monitor CTR, conversion, revenue per session. One of our client systems after hybridization increased revenue by 18% over a month of A/B.

Recommender System Development Timeline

The stages and typical time frames are in the table below. Costs are calculated individually based on catalog scale and latency requirements.

What's Included in Turnkey Development

1. Data audit — user×item matrix density (typically <0.1%), activity distribution, temporal patterns, cold start statistics.
2. Baseline — 'popular' as a simple threshold that is often hard to beat.
3. Iterative improvement — ALS → content features → two-stage → sequence-aware. Each step with A/B.
4. Serving infrastructure — batch precomputation, Redis, real-time re-ranking, Grafana monitoring.
5. Documentation — model card with metrics, deployment instructions, feature descriptions.
6. Team training — session on interpreting results and model fine-tuning.
7. Support — 1 month post-launch (incident fixes, pipeline tuning).

We are a team with 7+ years of experience in recommender systems, having delivered over 30 projects for e-commerce and media. We guarantee transparent A/B testing and documented metric improvements.

Want to assess the growth potential of your catalog? Contact us for a free data audit. Order recommender system development — first prototype within two weeks.

from implicit.als import AlternatingLeastSquares

model = AlternatingLeastSquares(
    factors=64,
    regularization=0.05,
    iterations=15,
    use_gpu=True
)
model.fit(user_item_matrix)

More about the mathematics of recommender systems — in specialized literature.