GONZALO
ES Contact
Case

Free AI Audits: automated consulting pipeline

End-to-end system that conducts free business audits via voice questionnaires, transcribes, analyzes with AI, and generates professional PDF reports — turning leads into qualified clients.

active Jul 2026 public
Cover for Free AI Audits: automated consulting pipeline

Step 01

Problem

Consulting audits cost €1,500–€3,000 and require weeks of manual work. Silvigon needed a way to offer free audits at scale, automatically detect process pain points, and generate professional reports that build trust and convert leads into clients.

Step 02

Context and constraints

Built for Silvigon, an automation and ERP company serving the prefabricated concrete, aggregates, and construction sector. The system needed to handle the entire audit lifecycle: from inviting companies, recording voice responses, AI-powered analysis, to generating a McKinsey-grade PDF report.

Role: Full system architect and developer: voice capture, AI pipeline, clustering, PDF generation, and Railway deployment.

Step 03

Key decisions

  • Voice-first questionnaire: each person records answers to 6 questions in ~10 minutes, no typing required. Audio uploaded directly to Cloudflare R2 via presigned URLs.
  • Groq Whisper for transcription (216x real-time, $0.04/hr) with OpenAI Whisper as fallback.
  • GPT-4o mini for exhaustive problem extraction (better 50 false positives than 1 missed problem).
  • Semantic clustering: sentence-transformers (Spanish embeddings) + UMAP + HDBSCAN to consolidate hundreds of problems into ~72 actionable opportunities.
  • LLM judge reviews each cluster to merge/split/reject before final output.
  • 4-step solution pipeline per opportunity: web research (Gemini + Google Search), technical proposal, commercial copy with ROI, and deterministic assembly.
  • Typst 0.15 for PDF generation — 55x faster than WeasyPrint, Knuth-Plass quality, <500ms compile time.
  • Sanitization pipeline removes employee names, uses brackets and neutral plural for anonymity.
  • Refundable pricing model: €299 report fee, fully refunded if client contracts implementation.

Step 04

Outcomes

  • End-to-end automated audit: from voice recording to professional PDF in minutes.
  • Cost per audit: ~$2–$5 in AI API calls, ~$0.0001 in PDF generation.
  • Professional report with 8 sections: executive summary, KPI dashboard, prioritized opportunities, solutions timeline, economic impact, methodology, next steps, guarantees.
  • Zero duplicates in real test: 269 individual problems → 72 consolidated opportunities.
  • Production-ready: deployed on Railway, Docker container with ffmpeg + Typst + embeddings model.

Metrics

  • Pipeline cost: ~$2.60 in AI for 72 opportunities (GPT-4o mini + Gemini + embeddings).
  • Transcription: 216x real-time speed via Groq.
  • PDF compilation: <500ms with Typst.
  • Clustering quality: 0 duplicates in 269 → 72 real test.
  • Anonymity: automatic sanitization of all personal identifiers.
  • ~72 opportunities identified per average audit, each with phase plan, ROI estimate, and evidence.

Step 05

Learnings and next improvement

How it works

The system is an end-to-end pipeline that goes from a company accepting a free audit invitation to receiving a professional PDF report — with zero human intervention.

Pipeline flow

Company accepts → [Invite employees] → [Voice recording (6 questions, ~10min)]
    → [Groq Whisper transcription] → [GPT-4o mini problem extraction]
    → [Semantic clustering (UMAP + HDBSCAN)] → [LLM judge reviews clusters]
    → [Per-opportunity: Web research → Technical solution → Commercial copy → Assembly]
    → [Prioritization matrix (Impact vs Complexity)] → [PDF generation (Typst)]

Key phases

Capture: Admin creates a company, selects departments, generates access tokens. Employees record answers via a web questionnaire with microphone test and frequency bars. Audio uploaded directly to Cloudflare R2 (no backend bottleneck).

Transcription: Groq Whisper Large v3 Turbo processes audio at 216x real-time speed. Files >25MB are compressed with ffmpeg first. OpenAI Whisper serves as fallback.

Analysis: GPT-4o mini extracts problems exhaustively from each transcription — titles, descriptions, verbatim quotes, tools mentioned, frequency, time lost. A second enrichment pass detects speaker roles, relationships, and quantification hints.

Clustering: Sentence-transformers (Spanish, 768d) → UMAP (15d) → HDBSCAN → LLM judge. This transforms hundreds of raw problems into ~72 consolidated, actionable opportunities without duplicates.

Solutions: For each opportunity, a 4-step pipeline runs: (1) Gemini + Google Search for state-of-the-art research, (2) GPT-4o mini for structured technical solutions in 3-5 phases, (3) Commercial copy with ROI estimates and complexity scores, (4) Deterministic assembly with evidence sanitization.

Report: Typst 0.15 generates an 8-section professional PDF with KPI cards, priority badges, phase roadmaps, and economic impact analysis. Compiles in <500ms.

Status: Alpha

The system is fully functional and deployed, but still in alpha phase. We are actively working on:

  • Employee anonymity: ensuring no employee can be identified from the report, to encourage honest answers without fear of reprisal.
  • Tool adoption: making sure the questionnaire and recording flow are intuitive and frictionless for all users.
  • Report quality: iterating on the PDF output to match the quality standards of premium consulting firms.

📄 Download example audit report (PDF, 1.7MB) Sample report generated from a real pipeline test. Company data has been anonymized.

Questionnaire dimensions (6 questions)

🔍 Processes — How do you detect a process is failing? 🧠 Decisions — How do you make decisions with incomplete data? ⚡ Problems — What repetitive problems do you face weekly? 🔀 Exceptions — How do you handle exceptions that don’t fit the system? 🔑 Dependencies — What happens when a key person is absent? 🌟 Vision — What would you automate if you could?

Learnings

  • Voice-first questionnaires dramatically lower the barrier for employee participation — people say more than they type.
  • Exhaustive extraction (more false positives) is the right strategy: LLM-based clustering handles consolidation reliably.
  • A deterministic assembly step after AI generation ensures consistent output formatting and prevents hallucination propagation.
  • Typst is dramatically superior to HTML-to-PDF engines for programmatic report generation.

Step 06

Project visual for Free AI Audits: automated consulting pipeline