AI-Personalized Fitness Programs: Adaptive Workouts Based on Biometrics

Standard fitness apps offer static plans that don't respond to the user's actual physiological state. The result: progress plateaus, overtraining, and injuries. We build AI systems that adapt load based on biometrics: HRV, heart rate, sleep quality, and training history. Our solutions increase adherence rates from 35–45% to 65–75% — a 1.7x improvement. Injury risk drops by 30–40% (Medicine & Science in Sports & Exercise). Personalized plans boost user LTV by 25–30% and reduce churn by 15–20%. We are a team of AI/ML engineers with 7 years of experience in sports physiology, with over 15 completed projects in workout personalization. Assess the implementation opportunity — contact us for a preliminary audit.

Why AI personalization outperforms static plans?

The key metric is not meeting a norm, but matching load to current recovery. Even the perfect plan becomes useless if the user has low HRV or poor sleep today. We deploy algorithms that compute a readiness score every morning (0–100). This score adjusts: workout type, intensity (via a load modifier), and recovery recommendations. This approach ensures smoother progress and reduces injury probability by 1.5x compared to static programs. Static plans typically achieve 40% adherence, while AI adaptation reaches 70% — 1.75x higher.

How does AI determine the optimal daily load?

The system considers four key factors: HRV, resting heart rate, sleep quality, and previous day's load. RecoveryMonitor computes a readiness score that directly influences training intensity. If the score is below 55 — only light recovery or rest is recommended. If above 75 — a heavy workout at 100% intensity is possible. Periodization is also applied: 3 weeks of increasing load, then a recovery week. Example: user wakes up with HRV 45ms (baseline 55ms), resting heart rate 62 (baseline 58), sleep 72 points, and had a heavy workout yesterday. RecoveryMonitor calculates: HRV dropped 18% → minus 25 points; RHR increased by 4 → minus 15 points; sleep 72 (above 60) → no penalty; high load yesterday → minus 10 points. Final readiness score = 50 — only light activity at 60% intensity.

Required data for personalization

Minimum set: 1–2 weeks of workout history and biometric indicators (HRV, resting heart rate, sleep) from a wearable tracker. If data is insufficient, we generate a profile based on anthropometry and goals, then calibrate it as real measurements come in. All major trackers are supported: Whoop, Oura, Apple Watch, Garmin, Polar.

Technical implementation: Python + Anthropic

Below is the real code for a plan generator we use in production systems. Core stack: Python 3.12, Anthropic Claude 3.5 (via Anthropic SDK), Pandas for analytics, NumPy for math.

import numpy as np
import pandas as pd
from dataclasses import dataclass
from typing import Optional
from anthropic import Anthropic
import json

@dataclass
class AthleteProfile:
    user_id: str
    age: int
    sex: str
    weight_kg: float
    height_cm: float
    fitness_level: str  # beginner, intermediate, advanced
    primary_goal: str   # weight_loss, muscle_gain, endurance, general_fitness
    available_days_per_week: int
    equipment: list    # ['dumbbells', 'barbell', 'pull_up_bar']
    injuries: list     # ['lower_back', 'knee']
    vo2max: Optional[float] = None

class FitnessPlanGenerator:
    """Generate and adapt training plan"""

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

    def calculate_training_zones(self, profile: AthleteProfile) -> dict:
        """Heart rate zones for cardio workouts"""
        # Tanaka formula (more accurate than 220-age)
        max_hr = 208 - 0.7 * profile.age

        return {
            'max_hr': int(max_hr),
            'zone1_recovery': (int(max_hr * 0.50), int(max_hr * 0.60)),
            'zone2_aerobic': (int(max_hr * 0.60), int(max_hr * 0.70)),
            'zone3_tempo': (int(max_hr * 0.70), int(max_hr * 0.80)),
            'zone4_threshold': (int(max_hr * 0.80), int(max_hr * 0.90)),
            'zone5_vo2max': (int(max_hr * 0.90), int(max_hr * 1.00)),
        }

    def generate_weekly_plan(self, profile: AthleteProfile,
                              recent_performance: list[dict]) -> list[dict]:
        """Weekly training plan"""
        training_zones = self.calculate_training_zones(profile)

        # Periodization: 3 weeks of increasing load + 1 recovery week
        # Determine current week of periodization from history
        week_in_cycle = self._get_week_in_cycle(recent_performance)
        load_modifier = [0.85, 1.0, 1.15, 0.70][week_in_cycle % 4]

        response = self.llm.messages.create(
            model="claude-3-5-sonnet-20241022",
            max_tokens=700,
            messages=[{
                "role": "user",
                "content": f"""Create a personalized weekly training plan.

Profile:
- Fitness level: {profile.fitness_level}
- Goal: {profile.primary_goal}
- Available days: {profile.available_days_per_week}
- Equipment: {profile.equipment}
- Injuries to avoid: {profile.injuries}
- Age: {profile.age}, Weight: {profile.weight_kg}kg

Current training intensity: {load_modifier:.0%} of base load
Training zones: Zone 2 aerobic = {training_zones['zone2_aerobic']} bpm

Recent performance (last 5 sessions):
{json.dumps(recent_performance[-5:], ensure_ascii=False)[:400]}

Create {profile.available_days_per_week} training sessions. Return JSON array:
[{{
  "day": "Monday",
  "session_type": "strength|cardio|hiit|recovery",
  "duration_min": 45,
  "exercises": [{{"name": "...", "sets": 3, "reps": "8-10", "rest_sec": 90}}],
  "cardio_zone": "zone2",
  "notes": "..."
}}]"""
            }]
        )

        try:
            return json.loads(response.content[0].text)
        except Exception:
            return []

    def _get_week_in_cycle(self, performance: list[dict]) -> int:
        if not performance:
            return 0
        return len(set(p.get('week_number', 0) for p in performance)) % 4


class RecoveryMonitor:
    """Recovery monitoring from biometrics"""

    def compute_readiness_score(self, biometrics: dict) -> dict:
        """
        Readiness score (0-100).
        Data: HRV, RHR, sleep_score, previous_day_load.
        """
        score = 100.0
        factors = []

        # HRV (Heart Rate Variability) — main indicator
        hrv = biometrics.get('hrv_ms')
        hrv_baseline = biometrics.get('hrv_baseline_ms', 50)
        if hrv and hrv_baseline:
            hrv_ratio = hrv / hrv_baseline
            if hrv_ratio < 0.85:
                score -= 25
                factors.append(f'HRV low ({hrv:.0f}ms vs {hrv_baseline:.0f}ms baseline)')
            elif hrv_ratio > 1.15:
                score += 5  # Good recovery

        # Resting Heart Rate
        rhr = biometrics.get('resting_hr_bpm')
        rhr_baseline = biometrics.get('rhr_baseline_bpm', 60)
        if rhr and rhr_baseline:
            if rhr > rhr_baseline + 5:
                score -= 15
                factors.append(f'RHR elevated ({rhr} vs {rhr_baseline} baseline)')

        # Sleep
        sleep_score = biometrics.get('sleep_score', 80)  # 0-100
        if sleep_score < 60:
            score -= 20
            factors.append(f'Poor sleep (score: {sleep_score})')
        elif sleep_score < 75:
            score -= 10

        # Previous day's load
        previous_load = biometrics.get('yesterday_training_load', 0)  # AU (Arbitrary Units)
        high_load_threshold = biometrics.get('weekly_avg_load', 300) * 0.4
        if previous_load > high_load_threshold:
            score -= 10
            factors.append('High load yesterday')

        score = float(np.clip(score, 0, 100))

        if score > 75:
            recommendation = 'Great day for an intense workout'
            intensity_modifier = 1.0
        elif score > 55:
            recommendation = 'Moderate workout — reduce intensity by 15%'
            intensity_modifier = 0.85
        elif score > 35:
            recommendation = 'Only light recovery or rest'
            intensity_modifier = 0.60
        else:
            recommendation = 'Active rest or day off'
            intensity_modifier = 0.0

        return {
            'readiness_score': round(score),
            'recommendation': recommendation,
            'intensity_modifier': intensity_modifier,
            'limiting_factors': factors
        }


class ProgressTracker:
    """Track progress and adjust plan"""

    def analyze_progress(self, training_logs: pd.DataFrame,
                          profile: AthleteProfile,
                          weeks: int = 8) -> dict:
        """Progress analysis over period"""
        recent = training_logs[
            training_logs['date'] >= pd.Timestamp.now() - pd.Timedelta(weeks=weeks)
        ]

        if recent.empty:
            return {}

        return {
            'sessions_completed': len(recent),
            'sessions_planned': weeks * profile.available_days_per_week,
            'adherence_rate': len(recent) / (weeks * profile.available_days_per_week),

            # Progress on key exercises
            'strength_progress': self._compute_strength_progress(recent),
            'endurance_progress': self._compute_endurance_progress(recent),

            'avg_session_duration_min': recent.get('duration_minutes', pd.Series([45])).mean(),
            'total_volume_kg': recent.get('total_volume_kg', pd.Series([0])).sum(),
        }

    def _compute_strength_progress(self, logs: pd.DataFrame) -> dict:
        """Change in max weights for main exercises"""
        if 'exercise_name' not in logs.columns:
            return {}

        key_exercises = ['squat', 'bench_press', 'deadlift', 'overhead_press']
        progress = {}

        for exercise in key_exercises:
            exercise_logs = logs[logs['exercise_name'] == exercise]
            if len(exercise_logs) < 2:
                continue
            first_max = exercise_logs.nsmallest(3, 'date')['max_weight_kg'].mean()
            last_max = exercise_logs.nlargest(3, 'date')['max_weight_kg'].mean()
            progress[exercise] = round((last_max - first_max) / max(first_max, 1) * 100, 1)

        return progress

    def _compute_endurance_progress(self, logs: pd.DataFrame) -> dict:
        if 'pace_min_per_km' not in logs.columns:
            return {}
        cardio = logs[logs['session_type'] == 'cardio']
        if cardio.empty:
            return {}
        early = cardio.head(3)['pace_min_per_km'].mean()
        recent = cardio.tail(3)['pace_min_per_km'].mean()
        improvement = (early - recent) / early * 100  # Lower pace = improvement
        return {'pace_improvement_pct': round(improvement, 1)}

Comparison: static plan vs AI adaptation

What's included in the work?

Implementation process

1. Analytics — study biometrics, workout types, current stack (1–2 days).
2. Design — determine architecture: which LLMs, how to store exercise embeddings (pgvector).
3. Implementation — write RecoveryMonitor, FitnessPlanGenerator, ProgressTracker modules (2–6 weeks).
4. Testing — A/B test on 50–100 users, evaluate adherence improvement (1–2 weeks).
5. Deploy — deploy microservices in Kubernetes with GPU nodes for inference (Triton Inference Server).

Economic efficiency

The cost of implementing a basic solution is recovered within 6–12 months by increasing user LTV by 25–30%. The average savings from developing a custom AI solution compared to buying a ready-made platform is 40%. Additionally, reducing churn by 15–20% directly increases revenue. Evaluate the economic effect for your product — contact us for a preliminary calculation.

Typical mistakes in AI personalization implementation

• Ignoring recovery data. A plan based only on goals (weight loss/muscle gain) without considering HRV and sleep leads to overtraining. We always include readiness score as a corrective factor.
• One model for all. General-purpose LLMs (basic GPT-4) give template advice. We use custom fine-tuned models based on LLaMA 3, trained on your data.
• No fallback mechanism. LLM failures (latency, toxic responses) should trigger a rule-based engine. Our architecture includes this.

Get a consultation on architecture and pipelines — contact us to discuss metrics and implementation plan for your fitness product.

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.