Every step has a script, every rule has a record.
A seven-step pipeline. Click a step: what is done, what is measured, which tool is used. The same pipeline works when the data source changes; only the content of the first two steps differs.
Scope and queries
keyword set, out-of-scope list, same-name/confusable-entity rules
Query sets are written for the brand, institution and person; same-name people and confusable entities are excluded from the start. The out-of-scope list matters as much as the query.
- What is measured
- Share of irrelevant records in the first pull
- Tools
- Keyword plan, negative list, test pull
- If skipped
- Every later step carries the noise; report volume inflates and the tone distribution comes out wrong.
Collection
raw data from 7 channels (CSV/API)
A period's pull from X, Facebook, Instagram, YouTube, news, Google reviews and forums; raw files are stored untouched. When the data source changes, only this step changes.
- What is measured
- Record counts per channel and gap check
- Tools
- Monitoring platform export, CSV/API
- If skipped
- A missing channel goes unnoticed; period comparison breaks.
Cleaning and relevance
brand-phrase filter, noise signatures, deduplication, AI spam gate
Scraper residue and structural junk are removed with a signature list; a record whose body does not contain the brand phrase cannot pass the gate; duplicates are merged. The AI spam classifier runs here and justifies every decision with a reason code.
- What is measured
- Precision and recall, separately
- Tools
- Python pipeline, LLM prompt v2, regression set
- If skipped
- Irrelevant content leaks into the report; the client says "this is not about us".
Taxonomy and sentiment
L1–L2 topic labels, tone; AI + human audit
Remaining records are labelled against a two-level topic taxonomy and by tone. AI labels, a human audits a sample; contradictions are added to the regression set.
- What is measured
- Number of label↔reason contradictions
- Tools
- Taxonomy file, classification script, audit sample
- If skipped
- Topic distribution drifts from period to period; trend commentary becomes unreliable.
Analysis
pivots, breakdowns, period comparison, sample content selection
Channel, topic and tone breakdowns; comparison with the previous period; sample content is selected by rule, not by whim.
- What is measured
- Consistency across periods
- Tools
- Pivot generator, card selection rules
- If skipped
- A different analyst picks different things each period; the report stays person-dependent.
Report
charts, cards and executive summary auto-placed into the template
Charts, cards and the executive summary are placed into the deck template by script. Font size, bullet count, colour — all read from decision records; the client sees no difference between periods.
- What is measured
- Preparation time (3–4 h → 40 min)
- Tools
- Deck engine, python-pptx, decision records
- If skipped
- Manual layout: hours, and inevitable inconsistency.
QA and delivery
audit scripts, decision-record check, scheduled sending
Gap, tone and metric audits run automatically; once the report gets its ready flag, scheduled delivery sends it, leaving evidence and a log.
- What is measured
- Number of post-delivery corrections
- Tools
- Audit scripts, Playwright delivery, log
- If skipped
- Wrong recipient, forgotten delivery, sends without evidence.
Four areas where we can work together
- 01
Monitoring and reporting setup
Query sets, taxonomy, report template and periodic workflow. Built from scratch or standardising an existing flow.
Deliverable: a working pipeline + runbook
- 02
AI classification setup and audit
Spam/relevance and sentiment prompts, test sets, precision/recall measurement, vendor evaluation. Tells you in numbers how accurate your current AI output is, then fixes it.
Deliverable: prompt + regression set + measurement report
- 03
Report-generation automation
Python automation of the raw-data-to-deck pipeline; QA audits; scheduled delivery.
Deliverable: pipeline + QA scripts
- 04
Corporate knowledge base design
Moving team knowledge into a decision-recorded, searchable structure that AI agents can use.
Deliverable: vault + router + linter
How it works
Short discovery call → scope and proposal → sprints with weekly deliverables. Remote; in-person possible in Istanbul.