保留正文与补采附件。源页面当前未重新核验；具体限制见交付清单。

Revisit consent button We value your privacy We use cookies to enhance your browsing experience, serve personalised ads or content, and analyse our traffic. By clicking "Accept All", you consent to our use of cookies. Customise Reject All Accept All Customise Consent Preferences We use cookies to help you navigate efficiently and perform certain functions. You will find detailed information about all cookies under each consent category below. The cookies that are categorised as "Necessary" are stored on your browser as they are essential for enabling the basic functionalities of the site. ... Show more NecessaryAlways Active Necessary cookies are required to enable the basic features of this site, such as providing secure log-in or adjusting your consent preferences. These cookies do not store any personally identifiable data.

Functional Functional cookies help perform certain functionalities like sharing the content of the website on social media platforms, collecting feedback, and other third-party features.

Analytics Analytical cookies are used to understand how visitors interact with the website. These cookies help provide information on metrics such as the number of visitors, bounce rate, traffic source, etc.

Performance Performance cookies are used to understand and analyse the key performance indexes of the website which helps in delivering a better user experience for the visitors.

Advertisement Advertisement cookies are used to provide visitors with customised advertisements based on the pages you visited previously and to analyse the effectiveness of the ad campaigns.

Uncategorised Other uncategorised cookies are those that are being analysed and have not been classified into a category as yet. No cookies to display. Reject All Save My Preferences Accept All Toggle SearchToggle Mobile Navigation Publish AI, ML & data-science insights to a global community of data professionals. Sign inSubmit an Article LatestEditor's PicksNewsletterShipAI

Write For TDS LinkedInXToggle Search Agentic AI

Building My Own Personal AI Assistant: A Chronicle, Part 2

Improving Fernão: Building a personal AI assistant is rarely a single, monolithic effort Ivo Bernardo April 16, 20268 min read You probably remember the first part of my journey building Fernão , my personal AI agent. Now it’s time to continue the story and let’s dive into the second part! In this post, I’ll walk you through the latest improvements to Fernão , refining existing features and adding new capabilities to the agent. Let’s start with what changed. ··· Remember the function that fetched the calendar through ICS (the universal calendar format) and extracted my calendar tasks? That function was a mess and it reflected a poor architectural decision. ICS calendars don’t support native filtering, which means every request required pulling all events from the calendar and filtering them afterward. In practice, Fernão was downloading my entire schedule just to extract a few relevant meetings. Learn this step by step with the interactive AI Agents roadmap. This was like bringing home an entire library just to look up a single sentence in one book. The more I tried to optimize the function, the less I would get anywhere, because I needed a new system solution. I needed to change the way the system was getting the calendar and I would NEVER get around this bottleneck with ICS. So I dug deeper and found that Google provides calendar access through an API that supports native filtering. Fernão now only retrieves the events it actually needs. This improved the schedule generation, that dropped from nearly five minutes to about twenty seconds. With this new pipeline in place, I also refactored the surrounding logic. The entire function is now significantly cleaner and faster. We now have a beautiful way to fetch the calendar events via API: python

def get_events_for_date(target_date=None, use_api=True):
    """
    Fetches events for a specific date from Google Calendar.
    Tries API first (if use_api=True), falls back to ICS if API fails.
    
    Args:
        target_date: datetime.date object for the target day. If None, uses today.
        use_api: If True, try Google Calendar API first. If False, use ICS only.
    
    Returns:
        List of event dictionaries.
    """
    if use_api and GCAL_API_AVAILABLE:
        print("[GCal] Attempting to use Google Calendar API...")
        try:
            events = get_events_for_date_api(target_date)
            if events is not None:
                print(f"[GCal] Successfully fetched {len(events)} events via API")
                return events
            else:
                print("[GCal] API returned None. Falling back to ICS...")
        except Exception as e:
            print(f"[GCal] API failed with error: {e}")
            print("[GCal] Falling back to ICS...")
    
    # Fallback to ICS
    print("[GCal] Using ICS feed method...")
    return get_events_for_date_ics(target_date)

.. and here is our API fetch: python

def get_events_for_date_api(target_date=None):
    """
    Fetches events for a specific date from Google Calendar using the Calendar API.
    
    Args:
        target_date: datetime.date object for the target day. If None, uses today.
    
    Returns:
        List of event dictionaries, or None if API call fails.
    """
    service = get_calendar_service()
    if not service:
        return None
    
    day_start, day_end, target_date, tz_name, local = _get_local_time_range(target_date)
    
    print(f"\n[GCal API] Fetching events for {target_date.strftime('%Y-%m-%d')}")
    print(f"  Timezone: {tz_name}")
    
    # Get calendar IDs from environment, or use primary
    calendar_ids_str = os.getenv('GCAL_CALENDAR_IDS', '')
    if calendar_ids_str:
        calendar_ids = [cid.strip() for cid in calendar_ids_str.split(',')]
    else:
        calendar_ids = ['primary']
    
    all_events = []
    
    # Fetch from each calendar
    for calendar_id in calendar_ids:
        try:
            print(f"[GCal API] Fetching from calendar: {calendar_id}")
            
            # Call the Calendar API with timeoutlobally
            old_timeout = socket.getdefaulttimeout()
            socket.setdefaulttimeout(10)
            
            try:
                events_result = service.events().list(
                    calendarId=calendar_id,
                    timeMin=day_start.isoformat(),
                    timeMax=day_end.isoformat(),
                    singleEvents=True,
                    orderBy='startTime'
                ).execute()
            finally:
                socket.setdefaulttimeout(old_timeout)
            
            events = events_result.get('items', [])
            print(f"[GCal API] Found {len(events)} event(s) in {calendar_id}")
            
            # Parse events
            for event in events:
                # Get start time
                start = event['start'].get('dateTime', event['start'].get('date'))
                end = event['end'].get('dateTime', event['end'].get('date'))
                
                # Parse datetime
                if 'T' in start:  # DateTime
                    start_dt = datetime.fromisoformat(start.replace('Z', '+00:00'))
                    end_dt = datetime.fromisoformat(end.replace('Z', '+00:00'))
                    
                    # Convert to local timezone
                    start_local = start_dt.astimezone(local)
                    end_local = end_dt.astimezone(local)
                    
                    start_str = start_local.strftime("%H:%M")
                    end_str = end_local.strftime("%H:%M")
                else:  # All-day event
                    start_str = "00:00"
                    end_str = "23:59"
                
                all_events.append({
                    "title": event.get('summary', 'Untitled Event'),
                    "start": start_str,
                    "end": end_str,
                    "location": event.get('location', ''),
                    "description": event.get('description', '')
                })
        
        except Exception as e:
            print(f"[GCal API] Error fetching from {calendar_id}: {e}")
            continue
    
    # Sort by start time
    all_events.sort(key=lambda x: x["start"])
    
    print(f"[GCal API] Total events: {len(all_events)}")
    return all_events

Beyond the back-end improvements, I’ve also added new features to the assistant. In the schedule view, I can now mark tasks as completed. The cool thing is that and when I do it, they’re automatically synced with Microsoft To-Do app. Checking tasks in the App - Image by Author Checking tasks in the App - Image by Author Marvelous! It makes me think that, over time, many of us will end up building our own “personal operating systems”, assembling workflows from our preferred tools, integrating what we need, and replacing components whenever better options appear. If features become easy to replicate, loyalty to specific platforms will weaken. This also raises an interesting question: will companies eventually try to lock down their systems by restricting APIs and external integrations? Possibly. But isolating themselves is unlikely to work in the long run. If users can’t connect tools to their own workflows, they’ll simply move elsewhere (… and will this be a behavior only done by power users?) With this new capability in the Schedule Maker, I went on to implement another feature for Fernão , suggested by one of the readers. Introducing: the Task Breaker. The new Fernão role is to break tasks - Image by Author - Generated by Gemini ··· The Task Breaker follows a simple workflow:

Fernão Task Breaker - Image by Author Here’s how the Task Breaker appears in the sidebar: Task Breaker Module - Image by Author And here’s the current (still rough) front end for the Task Breaker : Task Breaker Module - Image by Author Let’s take a real example. One of the larger tasks sitting in my To Do list is “ Project Documentation at DareData.” This isn’t a small operational task, it’s a structural project. The goal is to consolidate and formalize the internal knowledge of the company in Notion, making sure the main questions about the enterprise are clearly answered in our Knowledge Hub. That means reviewing every department, identifying gaps, creating or refining pages, structuring information properly, and assigning clear ownership for each section. In practice, this requires auditing, writing, and governance decisions. It’s not something you “just do” in one sitting and I want to complete this within three weeks. Realistically, I can dedicate between 30 minutes to one hour per day. This is exactly the kind of task that benefits from decomposition, so I’ll use some context information in the Task Breaker and turn this giga-task into an actionable plan: markup

I’m currently building our Knowledge Hub in Notion. Here’s the current structure of Notion and Departments:# DareData Hub## Handbook for Network Members[DareData Network Member Handbook ]## Teams### Sales & Partnerships### Marketing (Brand and Gen-OS)### Finance### Delivery[Tech Specialists][Principals][Account Managers]I’ll need to:- Go Through Every department and create the pages that make sense (search the web for more context on what DareData is if you need ideas of the pages I need to create on every department or use your knowledge on what definitely needs to be documented in a 130 people B2B company).- Make sure that every page has an Owner- Revisit the Suggestions recorded by my team in the suggestions, I can probably pick one or two suggestions every day- Add some departments that are not in there: Admin, Core Members, Finance, ProductI have around 30 minutes to 1 hour to work every day and want to complete this project by 7 March 2026.

After hitting "breakdown", Fernão will start to forge: Fernão Forging the Task - Image by Author Marvelous x2! Task Breakdown - Image by Author We can now review the tasks and plan generated and submit them to my Microsoft To-Do. I also need to assign which list on Microsoft To-Do they will be assigned to: Destination List in Microsoft To-Do - Image by Author

Cool! Let’s see how it looks in Microsoft app: Perfecto. While testing this, I realised I want to adjust two things.

Let's break another task after making these changes: Fernão now forges in a proper medieval vest: Fernão is forging again - Image by Author And now we can also use the nice “ Submit All ” button: Submit new tasks in To-Do - Image by Author I’ll likely remove the emoji. That kind of “ emojification ” is very typical of LLM-generated code in the front-end, and it really annoys me. Anyway, building this new feature for Fernão was genuinely satisfying. If you have ideas for additional features, I’m open to suggestions! Below is the current prompt I’m using for task breakdown. I’ll continue refining it as I experiment and observe how it performs in real use cases. yaml

name: task_breakdown
description: Break down a task into 20-minute actionable subtasks
max_tokens: 4096

variables:
- task_name
- task_context
- current_date
template: |
You are a productivity expert helping break down complex tasks into manageable 20-minute chunks.

**CURRENT DATE:** {current_date}
**TASK TO BREAK DOWN:**
{task_name}
**CONTEXT PROVIDED BY USER:**
{task_context}

**YOUR JOB:**

Break this task into specific, actionable subtasks that can each be completed in approximately 20 minutes.
**RULES:**
1. Each subtask should be concrete and actionable (starts with a verb)
2. Each subtask should take ~20 minutes (can be 15-25 min, but aim for 20)
3. Subtasks should follow a logical order
4. Be specific - avoid vague tasks like “work on X”
5. If the task is already small enough, you can create 1-3 subtasks
6. If it’s large, create > 5 subtasks
7. Consider the context provided - use it to make subtasks relevant and specific
8. **SCHEDULING:** Based on the user’s context (e.g., “every other day”, “weekends only”), suggest a specific Due Date for each task starting from the Current Date.

**OUTPUT FORMAT:**

Return ONLY a markdown list of subtasks in this format:
- {task_name} - [Subtask description] (20 min) [Due: YYYY-MM-DD]

Example:
- {task_name} - Create project repository (20 min) [Due: 2026-02-13]
- {task_name} - Configure CI/CD pipeline (20 min) [Due: 2026-02-15]
Do NOT include any other text or explanations. Just the list.

In parallel, I’m already working on several new modules:

Stay tuned for the next modules and hope this inspires you for your personal projects as well!

Register Now to Continue Reading

Unlock Towards Data Science with a Free Account

All articles on Towards Data Science are available to registered users. Create a free account today to access this article and thousands more. Register for freeSign in to your account

Written By

Ivo Bernardo

See all from Ivo Bernardo Ai AgentAi AssistantAi Product ManagementArtificial IntelligenceProgramming Share This Article

Towards Data Science is a community publication. Submit your insights to reach our global audience and earn through the TDS Author Payment Program. Write for TDS

Related Articles

Tool Calling, Explained: How AI Agents Decide What to Do Next

Agentic AI Understanding how LLMs interact with the world around them, from returning data to taking action Maria Mouschoutzi June 21, 202611 min read

Code Is Cheap. Engineering Judgement Is Now the Scarce Resource

Agentic AI The barriers to building have collapsed. That shifts the bottleneck to ownership, validation, taste, and deciding what should actually exist Clara Chong June 2, 20266 min read

6 Tips to Complete Udacity's Data Scientist Nanodegree in 1 month

Artificial Intelligence Maximize your learning with minimal time and cost Louis Teo February 4, 20216 min read

Target-encoding Categorical Variables

Artificial Intelligence One nice alternative to One-hot encoding your categories Vinicius Trevisan March 17, 20227 min read

How to Build Your Own Agentic AI System Using CrewAI

Agentic AI Step-by-Step Guide to Developing a Multi-Agent System for Social Media Marketing Destin Gong November 9, 202511 min read

How to Organize All of Your Coding Agent Tasks

Agentic AI Optimise how you interact with your coding agents Eivind Kjosbakken July 30, 202611 min read

Vibe Coding with AI: Best Practices for Human-AI Collaboration in Software Development

Agentic AI Accelerate coding with AI while staying in control and building reliable, production-ready software. Partha Sarkar March 19, 202614 min read

Hello! I am PAMI

Programming A new Pattern Mining Python library for Data Science Uday Kiran RAGE November 18, 20215 min read

Python Lists Are Overrated

Programming Consider these potentially superior alternatives Aashish Nair October 4, 20215 min read

Your home for data science and AI. The world's leading publication for data science, data analytics, data engineering, machine learning, and artificial intelligence professionals. Subscribe to Our NewsletterWrite For TDS•About•Advertise•Privacy Policy•Terms of Use © Insight Media Group, LLC 2026•The New Stack•roadmap.sh•Creator Network

Recommended Articles

Close