Strategic engineering-hours analysis for Omie's product team. Hybrid AI workflow with explicit role separation: Claude as discovery/strategy agent, local AI executor as data processor. Central question: does more investment in discovery reduce delivery effort?
Status: delivered
A data-driven analysis of Jira/Clockwork hour records, built to support strategic decisions about engineering investment, budget allocation (including R&D incentive programs like Brazil's Lei do Bem), and process quality conversations with business stakeholders.
The output: KPI reports, productivity dashboards, executive summaries — all in Markdown for clean review and reuse.
The repository demonstrates a hybrid AI workflow with explicit role separation between two AI agents:
[Discovery / Strategy] [Delivery / Execution]
───────────────────────── ─────────────────────────
Claude (chat) Local AI executor
• Absorbs core/ and docs/ ────► • Installs dependencies
• Defines analysis methodology ────► • Runs Python scripts
• Generates hypotheses ◄──── • Identifies data gaps
• Frames the business question • Exports summaries to output/
This separation is the point: discovery is human-shaped reasoning amplified by Claude; delivery is deterministic data processing executed by a scripted agent. Each role has a different prompt, a different context window, and a different failure mode.
Most AI workflows collapse both roles into one prompt. That works for trivial tasks and breaks at scale: the model is asked to simultaneously decide what to analyze and execute the analysis, with no checkpoint between strategy and execution.
The hybrid pattern solves three things:
- Auditability — discovery decisions are logged in markdown (
core/,docs/); execution is logged in code (scripts/). - Reproducibility — the executor runs the same scripts regardless of model, regardless of session.
- Stakeholder communication —
core/reads like a strategy brief,output/reads like a delivery report. Each artifact speaks to a different audience.
The same separation appears, in different form, in the Daktus CDSS pipeline: probabilistic reasoning in the LLM, deterministic execution in scripts and validation layers. Probabilistic where it adds value, deterministic where it doesn't.
clockdata/
├── .vibeflow/ # Code conventions and architecture (AI conventions)
├── core/ # Strategic documentation, scope, action plan
├── data/ # Raw data and Excel exports
├── docs/ # Domain context, discovery notes, data dictionary
├── output/ # Analysis results, reports, metrics
├── prompts/ # Structured prompts for the executor agent
├── scripts/ # Python execution (extraction and analysis)
├── MEMORY.md # Quick context index for AI agents
└── README.md
# Prerequisites: Python 3.10+, raw Excel files in data/
# 1. Install dependencies
pip install -r scripts/requirements.txt
# 2. Profile raw data (quality and gaps)
python scripts/extract_and_explore.py
# 3. Run strategic analysis and consolidate views
python scripts/analyze_hours.pyAll outputs (Markdown reports, processed CSV/Excel) land in output/.
| Document | Description |
|---|---|
| Project Overview | Business context, time horizon, objectives, stakeholders |
| Strategic Foundations | Engineering epistemology, quality (Shift-Left, Lean), base metrics |
| Executive Summary | Diagnostic summary with key effort and hours KPIs |
| Methodology Framework | How to cross variables, group tags, and answer the central question |
| Team & Roles | Adherence assumptions and role-level detail (PM, PD, Dev, QA) |
| Action Plan | Prioritized actions split into tactical and strategic views |
| Data Dictionary | Expected schema, relationships, table semantics from Jira |
Python · Pandas · Excel/CSV processing · Structured prompts (markdown contracts for executor agent) · Vibeflow conventions for AI-assisted development
Built by Daniel Martins — product analytics work for Omie's engineering team, coordinated with PM Camila and stakeholders Bia and Edu.
Background: engineering (EFOMM — systems modeling, automation) + medicine (UFJF, final-year). I build AI systems for high-stakes domains where the cost of error is real. This project applies the same separation-of-concerns architecture to a product analytics problem outside healthcare.