04Agents

Jarvis

A personal "second brain" agent — the goal is something that knows my context, talks back by voice, and can act across my own systems. Right now it's Phase 2: cleaning personal data and wiring it into the Claude API, built by directing an AI system rather than hand-writing everything myself.

RoleArchitecture, direction

Started2026

ModelClaude API

Phase2 — data + API wiring

Timeline

Click a stop to open the chapter

2026

The idea

Decided Jarvis should be more than a chatbot: something that knows my context, has access to my own data and systems, talks back by voice, and can take actions — not just answer questions.

—About 90% of the code is expected to be AI-written

—The focus is on understanding and directing the system, not memorizing syntax

2026

Synthetic data

Had Claude generate a synthetic life-log dataset to practice on, so I could break the cleaning pipeline without touching anything real first.

—Lower stakes for learning pandas fundamentals

—A safe place to make mistakes before Jarvis sees real personal data

2026

Data cleaning

Cleaned the synthetic dataset in pandas, working through actual errors in data_cleaning.py rather than following a script blindly. Picked up Power Query and pivot tables in Excel along the way as part of the same prep work.

—Debugging pandas errors one at a time, not restarting from scratch

—Excel Power Query and pivot tables picked up as a side effect

Now

Claude API wiring

Current phase: finish cleaning the data, then work through Anthropic Academy courses, then wire the cleaned data into real Claude API calls.

—Roadmap: data cleaning → Anthropic Academy → wire up Jarvis → deploy

—No voice interface or real actions yet — that comes after the core loop works

Later

Deployment

Deliberately not moving to AWS or any cloud deployment preemptively — only once deployment becomes a concrete, real problem rather than something to build ahead of need.

—Avoiding premature infrastructure

—Voice interface and system actions are later-stage goals, not current scope

Milestones

Highlight:

What broke

Didn't know where to start with an AI agent

Decided to learn by building instead of by tutorial — most of the code is AI-written, and my job is understanding and directing the system rather than memorizing syntax.

Real personal data felt risky to experiment on

Had Claude generate a synthetic life-log dataset first, so the cleaning pipeline could be broken and fixed without touching anything real.

pandas errors kept stalling the cleaning script

Worked through data_cleaning.py's errors one at a time instead of restarting — that debugging loop is most of what "Phase 2" has actually been so far.

From the repo

jarvis/data_cleaning.py

# Phase 2 — cleaning a synthetic life-log dataset
# before any real personal data touches the pipeline
import pandas as pd

df = pd.read_csv("synthetic_life_log.csv")
df = df.dropna(subset=["timestamp", "event"])
df["timestamp"] = pd.to_datetime(df["timestamp"])
df = df.sort_values("timestamp").reset_index(drop=True)

# next: feed the cleaned frame into Claude API calls

ReasoningClaude API (Phase 2 integration)

DataPandas, synthetic life-log dataset

LearningAnthropic Academy (next)

DeployAWS — only when it's a real problem

Next project

Bench

Open →