Building an AI Agent to Detect and Handle Anomalies in Time-Series Data

As a knowledge scientist engaged on time-series forecasting, I’ve run into anomalies and outliers greater than I can rely. Throughout demand forecasting, finance, visitors, and gross sales information, I hold operating into spikes and dips which can be laborious to interpret.

Anomaly dealing with is normally a grey space, not often black or white, however indicators of deeper points. Some anomalies are actual indicators like holidays, climate occasions, promotions, or viral moments; others are simply information glitches, however each look the identical at first look. The quicker we detect anomalies in information, the quicker motion could be taken to forestall poor efficiency and injury.

We’re coping with essential time-series information, and detecting anomalies is essential. When you take away a real occasion, a beneficial sign information level is eliminated, and in case you hold a false alarm sign, the coaching information accommodates noise.

Most ML-based detectors flag spikes primarily based on Z-scores, IQR thresholds, or different static strategies with none context. With current developments in AI, we have now a greater choice to design an anomaly-handling agent that causes about every case. An agent that detects uncommon conduct, checks context, and decides whether or not to repair the info, hold it as an actual sign, or flag it for overview.

On this article, we construct such an agent step-by-step that mixes easy statistical detection with an AI agent that acts as a primary line of protection for time-series information, decreasing guide intervention whereas preserving the indicators that matter most. We’ll detect and deal with anomalies in COVID-19 information by autonomous decision-making primarily based on the severity of the anomaly, utilizing:

1. Stay epidemiological information from the illness.sh API.
2. Statistical anomaly detection.
3. Severity classification.
4. A GroqCloud-powered AI agent that takes autonomous choices whether or not to:
   • Repair the anomaly
   • Preserve the anomaly
   • Flag anomaly for human overview

That is agentic resolution intelligence, not merely anomaly detection.

Why is conventional anomaly detection alone not sufficient?

There are conventional ML strategies like isolation forests designed for anomaly detection, however they lack end-to-end resolution orchestration. They’re unable to behave on them rapidly sufficient in manufacturing environments. We’re implementing an AI agent to fill this hole by turning uncooked anomaly scores into autonomous, end-to-end choices dynamically on reside information.

Conventional Anomaly Detection

The normal anomaly detection follows the pipeline strategy as drawn beneath:

Limitations of Conventional Anomaly Detection

• Works on static guidelines and manually units thresholds.
• It’s single-dimensional and handles easy information.
• No contextual reasoning.
• Human-driven resolution making.
• Guide-driven motion.

Anomaly Detection and Dealing with with an AI Agent 

The AI Agent anomaly detection follows the pipeline strategy as drawn beneath:

Why does this work higher in follow?

• Works on real-time information.
• It’s multidimensional and may deal with complicated information.
• Works on contextual reasoning.
• Adaptive & self-learning resolution making.
• Take autonomous motion.

Selecting a sensible dataset for our instance

We’re utilizing real-world COVID-19 information to detect anomalies, as it’s noisy, reveals spikes, and the outcomes assist in the development of public well being.

What do we wish the AI Agent to resolve?

The aim is to repeatedly monitor COVID-19 information, discover anomalies, outline their severity, and take autonomous choices and resolve motion to be taken:

• Flag anomaly for human overview
• Repair the anomaly
• Preserve the anomaly

Information Supply

For the info, we’re utilizing free, reside disease.sh information by way of API. This API gives information on each day confirmed instances, deaths and recoveries. For the AI Agent implementation, we’re specializing in each day case counts, which are perfect for anomaly detection.

Information license: This tutorial makes use of COVID-19 historic case counts retrieved by way of the illness.sh API. The underlying dataset (JHU CSSE COVID-19 Information Repository) is licensed beneath CC BY 4.0, which allows industrial use with attribution. (Accessed on January 22, 2026)

How do the items match collectively?

Excessive-Stage system structure of the anomaly detection on COVID-19 information utilizing an AI Agent is as follows:

Constructing the AI Agent Step-by-Step 

Let’s go step-by-step to grasp learn how to load information utilizing disease.sh, detect anomalies, classify them, and implement an AI agent that causes and takes applicable motion as per the severity of the anomalies.

Step 1: Set up Required Libraries

Step one is to put in required libraries like phidata, groq, python-dotenv, tabulate, and streamlit.

pip set up phidata
pip set up groq
pip set up python-dotenv #library to load .env file
pip set up tabulate
pip set up streamlit

Step 2: Atmosphere File Set-up

Open your IDE and create a venture folder, and beneath that folder, create an environmental file “.env” to retailer GROQ_API_KEY.

GROQ_API_KEY="your_groq_api_key_here"

Step 3: Information Ingestion

Earlier than constructing any agent, we want a knowledge supply that’s noisy sufficient to floor actual anomalies, however structured sufficient to cause about. COVID-19 each day case counts are an excellent match as they comprise reporting delays, sudden spikes, and regime adjustments. For simplicity, we intentionally prohibit ourselves to a single univariate time sequence.

Load information from the disease.sh utilizing request URL and extract the date and each day case rely primarily based on the chosen nation and the variety of days for which you need to extract information. The info is transformed right into a structured dataframe by parsing json, formatting date and sorting chronologically.

# ---------------------------------------
# DATA INGESTION (illness.sh)
# ---------------------------------------

def load_live_covid_data(nation: str , days:int):
    url = f"https://illness.sh/v3/covid-19/historic/{nation}?lastdays={days}"
    response = requests.get(url)
    information = response.json()["timeline"]["cases"]

    df = (
        pd.DataFrame(listing(information.gadgets()), columns=["Date", "Cases"])
        .assign(Date=lambda d: pd.to_datetime(d["Date"], format="%m/%d/%y"))
        .sort_values("Date")
        .reset_index(drop=True)
    )
    return df

Step 4: Anomalies Detection

We’ll now detect irregular conduct in COVID-19 time-series information by detecting sudden spikes and speedy progress developments. Case counts are typically secure, and huge deviations or sharp will increase point out significant anomalies. We’ll now detect anomalies utilizing statistical strategies and binary labeling for deterministic and reproducible anomaly detection. Two parameters are calculated to detect anomalies.

1. Spike Detection
   • A sudden spike in information is detected utilizing the Z-score; if any information level falls outdoors the Z-score vary, it have to be an anomaly.
2. Development Fee Detection
   • The day-over-day progress price is calculated; if it exceeds 40%, it’s flagged.

# ---------------------------------------
# ANOMALY DETECTION
# ---------------------------------------
def detect_anomalies(df):
   values = df["Cases"].values
   imply, std = values.imply(), values.std()

   spike_idx = [
       i for i, v in enumerate(values)
       if abs(v - mean) > 3 * std
   ]

   progress = np.diff(values) / np.most(values[:-1], 1)
   growth_idx = [i + 1 for i, g in enumerate(growth) if g > 0.4]

   anomalies = set(spike_idx + growth_idx)
   df["Anomaly"] = ["YES" if i in anomalies else "NO" for i in range(len(df))]

   return df

If there may be an anomaly based on both spike or progress or with each parameters, the “Anomaly” is about to “YES”; in any other case set to “NO”.

Step 5: Severity Classification

All anomalies are usually not equal; we are going to classify them as ‘CRITICAL’, ‘WARNING’, or ‘MINOR’ to information AI Agent choices. Fastened rolling home windows and rule-based thresholds are used to categorise severity. Severity is assessed solely when an anomaly exists; in any other case, Severity, Agent Determination, and Motion parameters within the dataframe are set to ‘clean’.

# ---------------------------------------
# CONFIG
# ---------------------------------------
ROLLING_WINDOW = 7
MIN_ABS_INCREASE = 500

# ---------------------------------------
# SEVERITY CLASSIFICATION
# ---------------------------------------
def compute_severity(df):
    df = df.sort_values("Date").reset_index(drop=True)
    df["Severity"] = ""
    df["Agent Decision"] = ""
    df["Action"] = ""
    for i in vary(len(df)):
        if df.loc[i, "Anomaly"] == "YES":
            if i < ROLLING_WINDOW:
                df.loc[i, "Severity"] = ""

            curr = df.loc[i, "Cases"]
            baseline = df.loc[i - ROLLING_WINDOW:i- 1, "Cases"].imply()

            abs_inc = curr - baseline
            progress = abs_inc / max(baseline, 1)

            if abs_inc < MIN_ABS_INCREASE:
                df.loc[i, "Severity"] = ""
            if progress >= 1.0:
                df.loc[i, "Severity"] = "CRITICAL"
            elif progress >= 0.4:
                df.loc[i, "Severity"] = "WARNING"
            else:
                df.loc[i, "Severity"] = "MINOR"
    return df

Within the above code, to categorise the anomaly severity, every anomaly is in contrast with 7-day historic information (ROLLING_WINDOW = 7), and absolute and relative progress are calculated.

1. Absolute Development

A MIN_ABS_INCREASE = 500 is outlined as a config parameter the place adjustments beneath this worth are thought of very small, a negligible change. If absolutely the progress is lower than MIN_ABS_INCREASE, then ignore it and hold the severity clean. Absolute progress detects significant real-world influence, doesn’t react to noise or minor fluctuations, and prevents false alarms when progress proportion is excessive.

2. Relative Development:

Relative progress helps in detecting explosive developments. If progress is bigger than or equal to 100% improve over baseline, it means a sudden outbreak, and it’s assigned as ‘CRITICAL’; if progress is bigger than 40%, it means sustained acceleration and wishes monitoring, and it’s assigned as ‘WARNING’; in any other case assigned as ‘MINOR’. 

After severity classification, it’s prepared for the AI Agent to make an autonomous resolution and motion.

Step 6: Construct Immediate for AI Agent

Beneath is the immediate that defines how the AI agent causes and makes choices primarily based on structured context and predefined severity when an anomaly is detected.  The agent is restricted to 3 express actions and should return a single, deterministic response for secure automation.

def build_agent_prompt(obs):
    return f"""
You're an AI monitoring agent for COVID-19 information.

Noticed anomaly:
Date: {obs['date']}
Instances: {obs['cases']}
Severity: {obs['severity']}

Determination guidelines:
- FIX_ANOMALY: noise, reporting fluctuation
- KEEP_ANOMALY: actual outbreak sign
- FLAG_FOR_REVIEW: extreme or ambiguous anomaly

Reply with ONLY one in all:
FIX_ANOMALY
KEEP_ANOMALY
FLAG_FOR_REVIEW
"""

Three information factors, i.e., date, variety of instances reported, and severity, are offered to the immediate explicitly, which helps the AI Agent to decide autonomously.

Step 7: Create your Agent with GroqCloud

We at the moment are creating an autonomous AI agent utilizing GroqCloud that makes clever contextual choices on detected anomalies and their severities and takes applicable actions. Three predefined actions for the AI Agent implement validated outputs solely.

# ---------------------------------------
# BUILDING AI AGENT
# ---------------------------------------
agent = Agent(
    title="CovidAnomalyAgent",
    mannequin=Groq(id="openai/gpt-oss-120b"),
    directions="""
You're an AI agent monitoring reside COVID-19 time-series information.
Detect anomalies, resolve based on the anomaly:
"FIX_ANOMALY", "KEEP_ANOMALY", "FLAG_FOR_REVIEW"."""
)
for i in vary(len(df)):
    if df.loc[i, "Anomaly"] == "YES":
        obs = build_observation(df, i)
        immediate = build_agent_prompt(obs)
        response = agent.run(immediate)

        resolution = response.messages[-1].content material.strip()
        resolution = resolution if resolution in VALID_ACTIONS else "FLAG_FOR_REVIEW"
        df = agent_action(df, i, resolution)

An AI agent named “CovidAnomalyAgent” is created, which makes use of an LLM mannequin hosted by GroqCloud for quick and low-latency reasoning. AI Agent runs a well-defined immediate, observes information, contextual reasoning, makes an autonomous resolution, and takes actions inside secure constraints.

An AI Agent isn’t dealing with anomalies however making clever choices for every detected anomaly. The agent’s resolution precisely displays anomaly severity and required motion.

# ---------------------------------------
# Agent ACTION DECIDER
# ---------------------------------------
def agent_action(df, idx,motion):
    df.loc[idx, "Agent Decision"] = motion

    if motion == "FIX_ANOMALY":
        fix_anomaly(df, idx)

    elif motion == "KEEP_ANOMALY":
        df.loc[idx, "Action"] = "Accepted as an actual outbreak sign"

    elif motion == "FLAG_FOR_REVIEW":
        df.loc[idx, "Action"] = "Flagged for human overview"
    return df

AI Agent ignores regular information factors with no anomaly and considers solely information factors with “ANOMALY= YES”. The AI agent is constrained to return solely three legitimate choices: “FIX_ANOMALY“, “KEEP_ANOMALY“, and “FLAG_FOR_REVIEW“, and accordingly, motion is taken as outlined within the desk beneath:

For minor anomalies, the AI agent mechanically fixes the info, preserves legitimate anomalies as-is, and flags essential instances for human overview.

Step 8: Repair Anomaly

Minor anomalies are attributable to reporting noise and are corrected utilizing native rolling imply smoothing over current historic values.

# ---------------------------------------
# FIX ANOMALY
# ---------------------------------------

def fix_anomaly(df, idx):
    window = df.loc[max(0, idx - 3):idx - 1, "Cases"]
    if len(window) > 0:
        df.loc[idx, "Cases"] = int(window.imply())

    df.loc[idx, "Severity"] = ""
    df.loc[idx, "Action"] = "Auto-corrected by an AI agent"

It takes the speedy 3 days of previous information, calculates its imply, and smooths the anomaly by changing its worth with this common. By the native rolling imply smoothing strategy, short-term spikes and information glitches could be dealt with. 

As soon as an anomaly is fastened, the info level is not thought of dangerous, and severity is deliberately eliminated to keep away from confusion. “Motion” is up to date to “Auto-corrected by an AI agent”.

Full Code

Kindly undergo the entire code for the statistical anomaly detection and AI Agent implementation for anomaly dealing with.

https://github.com/rautmadhura4/anomaly_detection_agent/tree/main

Outcomes

Let’s evaluate the outcomes for the nation, “India,” with various kinds of severity detected and the way the AI Agent handles them.

State of affairs 1: A Native Implementation

The primary try is a local implementation the place we detect minor anomalies and the AI Agent fixes them mechanically. Beneath is the snapshot of the COVID information desk of India with severity.

We’ve additionally applied a Streamlit dashboard to overview the AI Agent’s choices and actions. Within the beneath end result snapshot, you may see that numerous minor anomalies are fastened by the AI Agent.

This works finest when anomalies are localized noise fairly than regime adjustments.

State of affairs 2: A Boundary Situation

Right here, essential anomalies are detected, and the AI Agent raises a flag for overview as proven within the snapshot of the COVID information desk of India with severity.

On the Streamlit dashboard AI Agent’s choices and actions are proven within the end result snapshot. You possibly can see that every one the essential anomalies have been flagged for human overview by the AI Agent.

Severity gating prevents harmful auto-corrections in high-impact anomalies.

State of affairs 3: A Limitation 

For the limitation state of affairs, warning and demanding anomalies are detected as proven within the snapshot of the COVID information desk of India with severity.

On the Streamlit dashboard AI Agent’s choices and actions are proven beneath within the end result snapshot. You possibly can see that the essential anomaly is flagged for human overview by AI Agent, however the WARNING anomaly is mechanically fastened. In lots of actual settings, a WARNING-level anomaly needs to be preserved and monitored fairly than corrected.

This failure highlights why WARNING thresholds needs to be tuned and why human overview stays important.

Use the entire code and take a look at anomaly detection for the COVID-19 dataset, with totally different parameters.

Future Scope and Enhancements

We’ve used a really restricted dataset and applied rule-based anomaly detection, however sooner or later, some enhancements could be executed within the AI Agent implementation:

1. In our implementation, an anomaly is detected, and a call is made primarily based on case rely solely. Sooner or later, information could be extra elaborate with options like hospitalization information, vaccination information, and others.

2. Anomaly detection is completed right here utilizing statistical strategies, which can be ML-driven sooner or later to determine extra complicated patterns.

3. Now, we have now applied a single-agent structure; sooner or later multi-agent structure could be applied to enhance scalability, readability, and resilience.
4. Sooner or later human suggestions loop also needs to take care to make improved choices.

Ultimate Takeaways

Smarter AI brokers allow operational AI that makes choices utilizing contextual reasoning, takes motion to repair anomalies, and escalates to people when wanted. There are some sensible takeaways to remember whereas constructing an AI Agent for anomaly detection:

• To detect anomalies, use statistical strategies and implement AI brokers for contextual decision-making.

• Minor anomalies are secure to be autocorrected as they’re typically reported as noise. Important ought to by no means be autocorrected and flagged for overview by area specialists in order that real-world indicators don’t get suppressed.

• This AI agent should not be utilized in conditions the place anomalies straight set off irreversible actions.

When statistical strategies and an AI agent strategy are mixed correctly, they remodel anomaly detection from simply an alerting system right into a managed, decision-driven system with out compromising security.