All Case Studies
Ab
Abstrabit Technologies
AI Process Automation
Case Study
ISO Compliance · RAG Pipeline

6–12 Hours
Per SOP.
Now 40 min.

An ISO compliance firm's consultants were spending 30% of their time on documentation instead of auditing. We built a RAG pipeline that turns audio, handwritten notes, and emails into auditor-ready documents — with zero hallucination.

RAG Pipeline Multi-Modal Ingestion Document Generation Hallucination Control
18×
Faster SOP production
40 min
Per SOP (was 6–12 hrs)
3–5 days
Full QMS cycle (was 2–4 wks)
30→8%
Consultant time on docs
<2 min
Clause coverage check
Industry
ISO Compliance Consulting
Location
Salt Lake City, UT
Team
100+ employees
Revenue
$5M+ ARR
Engagement
7 weeks
The Problem

Expert Auditors. Spending 30% as Typists.

🎙️
Audio Recordings
Plaud recorder on-site
Hours-long interviews
No timestamps or transcripts
📝
Handwritten Notes
Process diagrams on paper
Margin notes from factory floors
Warehouse-lighting photos
📧
Scattered Text
Email threads with key facts
Partial docs from prior audits
Version drift across consultants
1–2h
Transcribe
Audio → text
1h
Read ISO Clauses
Manual reference
2–4h
Draft Document
Clause by clause
2–3h
Cross-Reference
Verify coverage
1–2h
Formatting
Template + flowcharts
6–12h
Per SOP, every single time
30%
Consultant capacity lost to docs
100%
Manual clause verification
System Architecture

Three Inputs. Five Sources.
Zero Hallucination.

Stage 1 — Multi-Modal Ingestion (all on-premise)
🎙️ Audio → Transcript
Whisper (local) · speaker diarization · timestamped output
no cloud STT
📝 Handwriting → Text
Tesseract OCR · deskew + contrast preprocessing · handles poor lighting
fully local
📧 Text → Facts
Entity extraction: roles, equipment, control points · source attribution preserved
timestamped + cited
↓ All streams produce atomic "fact objects" with source attribution
Stage 2 — RAG Across Five Knowledge Sources
①
ISO Standards
9001 · 14001 · 27001
②
Templates
200+ doc templates
③
Historical SOPs
Anonymized benchmarks
④
Regulatory
FDA 21 CFR · AS9100
⑤
Client Context
Org charts · process maps
pgvector semantic search + metadata filtering → Cross-encoder re-ranker prioritizes by clause
↓ Retrieved context + facts → two-step generation
Stage 3 — Two-Step LLM (Hallucination-Controlled)
STEP 01
🔍 Extract
GPT-4o extracts structured facts from raw inputs. Only what was explicitly stated. Every fact linked to source.
extract-only · no inference
→
STEP 02
📋 Compose
Second pass: facts + ISO clauses + template → formal document. If a fact wasn't extracted, it cannot appear.
traceable · auditor-trusted
↓ Three deliverables generated
Outputs
📄
ISO-Compliant DOCX
Template-enforced · house style
✅
Clause Coverage Matrix
Every clause mapped · gaps flagged
🔀
Process Flowchart
AI-generated · editable in React Flow
Project Proof

Not Described. Shown.

Actual system outputs — clause coverage matrix, AI-generated flowcharts, and fact extraction traces from the live pipeline.

ISO 9001:2015 · QMS-SOP-003 · Receiving Inspection 2 gaps flagged
Clause
§8.4
§8.5
§8.6
Status
8.4.1 General
✓ Covered
8.4.2 Control type
⚠ Partial
8.4.3 Ext. providers
✓ Covered
8.5.1 Controlled
✓ Covered
8.5.6 Changes
✕ Gap
8.6 Release
✓ Covered
Fully addressed
Partial
Gap
N/A
Fact Extraction · QMS-SOP-003 11 facts extracted
✓Incoming materials inspected by QC Lead within 24h · audio 14:32
✓Inspection checklist: Form QC-07, Rev 3 · notes p.2
✓Non-conforming items quarantined, NCR logged · audio 16:44
✓Release: QC Manager sign-off for Tier-1 suppliers · email 09:14
⚠Change control for approved suppliers: NOT STATED — flagged
Receiving Inspection · SOP-003 · §8.4–8.6
Export Edit Mode
Goods Received
↓
QC Lead Notified
within 24 hours
↓
Inspection OK?
↙ Yes
Tag & Release
QC Manager sign-off
No ↘
Quarantine
NCR logged QC-07
↓
↓
Update QMS
inventory confirmed
Supplier NCR
corrective action
⚡ Generated in 5–10 min via AI · drag-and-drop editing · replaces 1–2 hrs manual Visio work
Hallucination Control — Two-Step Architecture

Step 1 extracts only explicit statements. No inference, no gap-filling.

Step 2 composes only from Step 1 output. Unstated facts cannot appear.

Measured Impact

Capacity Freed. Quality Standardised. Audit Risk Eliminated.

18×
Faster SOP production
(12 hrs → 40 min)
22%
Consultant capacity
returned to billable work
<2 min
Clause coverage check
(was 2–3 hours)
SOP Production Time
Before
6–12 hrs
After
~40 min
Full QMS Documentation Cycle
Before
2–4 weeks
After
3–5 days
Consultant Time on Docs
Before
~30%
After
~8%
Full Stack
Whisper (local) GPT-4o (2-step) pgvector React Flow Tesseract OCR python-docx Cross-encoder re-rank Celery + RabbitMQ Django REST PostgreSQL AWS EC2 + S3
Metric
Before
After
SOP production
6–12 hours
<40 minutes
Full QMS cycle
2–4 weeks
3–5 days
Consultant doc time
~30%
~8%
Clause verification
2–3 hrs manual
Auto, <2 min
Flowchart creation
1–2 hrs (Visio)
5–10 min (AI)
Doc consistency
Consultant-dependent
Template-enforced
Key Architectural Decision
Separating extraction from composition made the system auditor-trustable. Consultants verify facts, not prose — reducing review from hours to minutes.
Abstrabit Technologies
AI Process Automation · Incubated at NSRCEL, IIM Bangalore