Build a private LLM agent to automate work lists from reports

Teams drowning in monthly reports can spend days manually extracting meaningful work items. When each developer writes updates in different formats, the task becomes slow, inconsistent, and error-prone. Worse, sending reports to cloud AI tools raises security concerns for companies handling sensitive data. The solution? A console-based AI agent that runs entirely on your own servers.

We built a lightweight system that normalizes chaotic report formats, enriches entries with Jira details, filters out vague noise, and produces a clean list of completed tasks automatically. By running on CPU-only hardware, the agent eliminates external data exposure while keeping costs low. Here’s how it works.

Why manual report processing fails teams

Most managers follow the same painful routine: collect dozens of developer reports, scan through hundreds of entries, and manually pick out the few that actually describe completed work. The process is straightforward but riddled with problems.

First is data inconsistency. Developers write updates in wildly different styles—some include detailed Jira references, others use cryptic shortcuts like “fixed issue” or “adjusted header.” When someone unfamiliar with the project reviews these later, crucial context is lost. Did “refactored code” touch feature X or module Y? The original intent becomes unclear.

Second is duplicate work. Tasks spanning multiple days often get logged repeatedly, creating near-identical entries across reports. Managers may also re-add items from previous months without realizing they’re already tracked. Without automated comparison, overlaps persist.

Third is security risk. Early attempts to use cloud AI tools like ChatGPT worked well for cleaning data but required uploading entire monthly reports to external servers. For finance, healthcare, and other regulated industries, this level of exposure is unacceptable.

How the private LLM agent transforms raw reports

Our solution is a console application that runs on internal servers, triggered automatically at month-end via cron job or API call. The agent processes raw reports for each active project, applies structured transformations, and outputs a polished list of work items—all without ever leaving your network.

The system uses Ollama to serve a local instance of Gemma 4 E2B running on CPU-only hardware. For duplicate detection, we leverage nomic-embed-text, a compact embedding model measuring just a few megabytes. Here’s the complete workflow:

1. Normalization: turning chaos into structured data

A single project may generate 80+ reports per month, each formatted differently. The agent’s first step converts this free-form text into a consistent, machine-readable structure. This transformation alone turns unstructured chaos into reliable, processable data.

2. Chunking: respecting CPU token limits

Local LLM inference on CPU comes with constraints. Our Gemma 4 model supports a 4,096-token context window—not enough for full monthly reports from busy teams. The solution? Intelligent chunking.

The system estimates token counts per batch, splitting reports into groups of about 20 entries each. This ensures the LLM never exceeds capacity while processing every chunk thoroughly. In one test, 22 raw reports were divided into 94 distinct work items after splitting multi-task entries like “Did A, did B, did C.”

3. Jira enrichment: adding missing context

Developers often write terse updates assuming Jira ticket IDs provide enough detail. Later, when these cryptic references appear as “AAA-123 – done,” they tell managers nothing about actual accomplishments.

Our agent automatically detects Jira IDs in reports and fetches full descriptions via the Jira API. This replaces vague entries with professional summaries of completed work, giving managers clear visibility into team achievements.

4. Noise filtering: removing empty statements

Not every report entry deserves inclusion. Generic phrases like “working on…” or “following up” carry no meaningful information. We implemented a bad-word filter as part of our intelligent document processing pipeline.

The LLM scans each chunk against our exclusion list and flags vague entries. In testing, this filter removed 69.1% of entries, leaving only 29 meaningful items from an initial pool of 94. What remained were concrete, specific descriptions of completed tasks.

5. Candidate selection: choosing the top work items

After filtering, the system selects the most relevant entries for presentation. The exact number varies by project and is stored in our internal reporting database. To allow for additional filtering in the next step, we typically start with a larger pool—around 80 items.

6. Vector duplicate detection: preventing repeats

Duplicate detection uses vector embeddings to compare new entries against historical data. By calculating semantic similarity, the system identifies near-identical tasks that might represent the same work logged multiple times. This ensures managers see each accomplishment only once, maintaining clean, accurate records.

Security and cost benefits of on-premise AI

Running entirely on internal servers means sensitive project data never leaves your network. Companies in regulated industries can finally use AI for document processing without compliance risks. Performance remains strong even on standard CPU hardware, making the solution accessible to enterprises with limited GPU resources.

The agent’s modular design allows easy updates to models, filters, and workflows as reporting standards evolve. With automation handling the tedious work of report parsing and task extraction, managers regain time for strategic planning instead of manual data cleanup.

What’s next for private LLM report automation

As LLMs continue improving, future versions could incorporate real-time report processing during the month rather than waiting for month-end cycles. Enhanced natural language understanding might detect nuanced accomplishments currently missed by simple filters.

For now, the current system delivers reliable, private automation that turns months of chaotic reports into clean, actionable work lists—without compromising security or breaking the budget.