Author: Krish Makadiya

Clean Textbook Ingestion for a RAG-Based CBSE Math Learning Platform
Industry: Education / EdTech

At a glance: MyntiQ Tech built a PDF-to-Markdown ingestion pipeline with PyMuPDF for an AI learning platform for CBSE Class 9 Math. It removes repeated watermarks, links figures to their captions, and records both printed and PDF page numbers. That lets retrieval-augmented generation (RAG) answers cite the textbook page a student actually sees.

Overview

This case study covers the first engineering milestone of an AI learning platform MyntiQ Tech is designing for CBSE Class 9 Math. That milestone was turning textbook PDFs into clean, structured text a retrieval system can trust. Answer quality in retrieval-augmented generation (RAG) depends heavily on the source text, so ingestion was built first.

Business Problem
- Watermarks repeated on every page of the textbook PDFs.
- Figures were separated from their captions.
- Printed page numbers did not match PDF page numbers, so answers could not cite the page a student actually sees.
How MyntiQ Tech Helped
- Watermark removal: image hashing finds images that repeat across pages and drops them.
- Figure extraction: each figure is linked to its caption.
- Dual page tracking: both the printed page number and the PDF page number (via pdftotext) are recorded.
- RAG-ready output: Markdown with YAML front matter, ready for chunking and embedding, built with PyMuPDF.
- Platform design: Nomic embeddings, Qdrant and a BGE reranker. Queries are routed between a self-hosted Llama 3.1 8B for low-stakes questions and Claude via AWS Bedrock for teaching-heavy ones.
Business Outcomes Delivered
- Clean, structured source text for the retrieval layer.
- Answers that can cite the textbook page a student actually sees.
- Hybrid model routing that keeps low-stakes queries on self-hosted infrastructure.
Frequently Asked Questions

Why does PDF ingestion matter for RAG?
A RAG system can only be as accurate as the text it retrieves. Watermarks, figures separated from their captions, and mismatched page numbers all reduce answer quality and make citations unreliable.

How do you remove watermarks from textbook PDFs?
MyntiQ Tech’s pipeline hashes the images on each page and drops images that repeat across pages, which is how watermarks show up.

How can an AI tutor cite the right textbook page?
The pipeline records both the printed page number and the PDF page number, so answers can cite the page number printed in the student’s book.

Which models does the CBSE learning platform use?
The design routes low-stakes queries to a self-hosted Llama 3.1 8B and teaching-heavy queries to Claude via AWS Bedrock. Retrieval uses Nomic embeddings, Qdrant and a BGE reranker.

Building an AI product on your own documents? Talk to MyntiQ Tech or browse more case studies.
10/21/2026
Recall-First Clinical NLP Pipeline for Prioritizing Clinician Review
Industry: Healthcare

At a glance: MyntiQ Tech prototyped a clinical NLP pipeline that reads clinical notes and flags the ones that may point to a cancer finding, so clinicians review them first. It starts from a TF-IDF and logistic regression baseline, fine-tunes Bio_ClinicalBERT and PubMedBERT, and tunes decision thresholds for recall. The model does not diagnose; clinicians make every clinical decision.

Overview

This case study covers a prototype natural language processing pipeline built by MyntiQ Tech. It reads clinical notes and flags the ones that may point to a cancer finding, so clinicians can review those notes first. The pipeline is tuned for recall because, in this setting, a missed case costs more than a false alarm.

Business Problem
- Clinical notes are long and full of shorthand, which makes relevant findings easy to miss.
- Positive cases are rare, so a model judged on overall accuracy can look good while still missing the cases that matter.
- Any AI in this workflow has to support clinicians, not replace their judgment.
How MyntiQ Tech Helped
- Explainable baseline: TF-IDF with logistic regression, so every later model has an honest benchmark.
- Domain-tuned transformers: Bio_ClinicalBERT and PubMedBERT, fine-tuned for clinical language.
- Abbreviation expansion: clinical shorthand is expanded before modelling.
- Class imbalance handling: techniques to account for how rare positive cases are.
- Recall-tuned thresholds: more notes are flagged for review in exchange for fewer missed cases.
Business Outcomes Delivered
- A prioritized review queue that puts the notes most likely to matter in front of clinicians first.
- A clear guardrail: the model does not diagnose, and the clinical decision stays with people.
Frequently Asked Questions

Why optimize clinical NLP for recall instead of accuracy?
In clinical text, a missed case usually costs more than a false alarm. Tuning for recall flags more notes for review in exchange for fewer missed cases. Accuracy can look high even when rare positive cases are missed.

Which models work well for clinical notes?
In this pipeline, MyntiQ Tech fine-tuned Bio_ClinicalBERT and PubMedBERT, transformer models pre-trained on medical language. It compared them against a TF-IDF and logistic regression baseline.

Does the model diagnose cancer?
No. It only decides which notes a clinician should read first. Every clinical decision stays with clinicians.

How are clinical abbreviations handled?
Clinical shorthand is expanded before the text reaches the model, because notes are full of abbreviations that general language models misread.

Exploring AI for clinical documentation or healthcare workflows? Talk to MyntiQ Tech or browse more case studies.
10/21/2026
MAEP: A Self-Hosted AI Engineering Platform for Faster, Observable Client Delivery
Industry: Software Engineering / AI (Internal Platform)

At a glance: MAEP (the myntiQ AI Engineering Platform) is a self-hosted platform built by MyntiQ Tech. It runs on a k3s cluster on the company’s own Dell PowerEdge R730 server. It combines Keycloak single sign-on, Gitea, Jenkins CI/CD, Langfuse LLM tracing, Claude models via AWS Bedrock, and ready-made project templates, so every client project starts with the same engineering conventions.

Overview

This case study covers MAEP, the myntiQ AI Engineering Platform. MAEP is the internal platform every MyntiQ Tech client project starts on. It runs on a k3s cluster on a Dell PowerEdge R730 server that MyntiQ Tech owns and operates. It brings source control, CI/CD, single sign-on, AI model access and observability onto infrastructure the team controls.

Business Problem
- Building AI systems for clients needs version control, CI/CD, single sign-on and visibility into every model call.
- Spreading these across a dozen SaaS tools gets messy, and client code and prompts end up in too many places.
- Starting each new project from an empty folder made it hard to apply the same engineering conventions every time.
How MyntiQ Tech Helped
- Single sign-on: Keycloak across every engineering tool.
- Source control and CI/CD: Gitea and Jenkins.
- LLM observability: Langfuse traces prompts, latency and cost on every model call.
- Model access: Claude models via AWS Bedrock, routed by task.
- Reproducible infrastructure: Terraform and Ansible.
- Project templates: template repos for Node/TypeScript, Python/FastAPI, React/Next.js, React Native and Angular. Each ships with linting, tests, a Jenkinsfile and an AGENTS.md file that tells AI coding assistants the team’s conventions.
Business Outcomes Delivered
- Client projects start from proven conventions instead of an empty folder.
- Code and prompts stay on infrastructure MyntiQ Tech controls.
- Every AI call can be observed from the first day of a project.
- Next on the roadmap: one-command developer onboarding.
Frequently Asked Questions

What is MAEP?
MAEP is the myntiQ AI Engineering Platform: MyntiQ Tech’s self-hosted platform for building AI and software systems. It brings single sign-on, source control, CI/CD, LLM tracing and model access together on infrastructure the company controls.

Why self-host an AI engineering platform?
Self-hosting keeps client code and prompts on infrastructure the team controls instead of spreading them across many SaaS tools. It also makes every AI call observable from day one.

How does MyntiQ Tech monitor LLM calls?
MyntiQ Tech traces prompts, latency and cost on model calls with Langfuse, the same way it monitors APIs.

What is AGENTS.md?
AGENTS.md is a file in each MyntiQ Tech template repo. It sets out the team’s coding conventions for AI coding assistants as well as for engineers.

Building an AI product and want an engineering partner with this foundation? Talk to MyntiQ Tech or browse more case studies.
10/21/2026
Phased ERPNext Rollout for an Industrial Automation Manufacturer
Industry: Manufacturing (Industrial Automation)

At a glance: MyntiQ Tech, a software and AI engineering company based in Ahmedabad, India and Danbury, Connecticut, USA, planned a phased ERPNext implementation for an industrial automation manufacturer in Ahmedabad. The plan moves 50–70 users off a desktop ERP in three phases over 26 weeks, starting with 9 core modules and GST compliance.

Overview

This case study covers the ERPNext rollout MyntiQ Tech planned for an Ahmedabad industrial automation company working in SCADA, panel manufacturing and project-based engineering. Around 50–70 users across sales, stores, production and projects are moving off a desktop ERP. Instead of a single big-bang go-live, the rollout is planned in three phases over 26 weeks.

Business Problem
- Sales, stores, production and projects all ran on an ageing desktop ERP.
- Switching every module on at once would put every department through change on the same day, which is a common way for ERP projects to stall.
- Leads came from many sources, including IndiaMart, the website, telecalling, email and referrals, and each needed to be tracked properly.
- Stores, warehousing and project support needed tools that matched how the team actually works.
How MyntiQ Tech Helped
- Phase 1 (11 weeks): 9 core ERPNext modules, email integration and GST compliance.
- Phase 2: 8 additional modules, plus WhatsApp and SMS.
- Phase 3: advanced modules and AI features, built on clean operational data.
- CRM lead sources: set up for IndiaMart, the website, telecalling, email and referrals.
- Barcode scanning: for stores and warehousing.
- Helpdesk: linked to customers and projects.
- Microsoft email integration: so the team keeps working in the tools it already uses.
Business Outcomes Delivered
- A rollout sequence that puts the daily essentials first, depth second, and AI once the data is ready.
- Less disruption, because each department adopts the core system before more modules arrive.
Frequently Asked Questions

How long does an ERPNext implementation take for a manufacturing company?
For this 50–70 user industrial automation manufacturer, MyntiQ Tech planned 26 weeks in three phases. The first phase, covering 9 core modules, email integration and GST compliance, takes 11 weeks.

Why use a phased ERP rollout instead of a big-bang go-live?
A phased rollout lets each department adopt the core system before more modules arrive. That lowers disruption and the risk of the project stalling. AI features come last, once the operational data is clean.

Can ERPNext handle GST compliance and IndiaMart leads?
Yes. In this rollout, GST compliance is part of Phase 1, and the CRM is set up to capture leads from IndiaMart, the website, telecalling, email and referrals.

Who implements ERPNext in Ahmedabad?
MyntiQ Tech implements, customizes and migrates to ERPNext from its office in Prahladnagar, Ahmedabad, including migrations from Tally, SAP and Odoo.

Planning an ERP move for your manufacturing business? Talk to MyntiQ Tech or browse more case studies.
10/21/2026