Mubeen Sahar
Available · Open to Opportunities

Software Engineer & AI Builder.

I build intelligent systems, developer tools, and human-centred products — across AI, ERP architecture, and full-stack engineering. Close to three years of shipping real things to real people.

// TruthLens · 50-Sample Eval Dataset
Macro F1 Average 91.3%
Hallucination Shield94%
Bias Detector91%
Manipulation Radar88%
DeBERTa-v3 BART-large-mnli MV3 Extension Wikipedia NLI
🛡️
AI Safety
3 platforms protected
Real-time NLI
Zero latency penalty
3+
Years Experience
7
Featured Projects
3
AI Platforms Supported
91.3%
Macro F1 Score
Flagship Work

TruthLens — AI Integrity
in Real Time

A three-signal AI safety layer running live inside ChatGPT, Claude and Gemini. Detects hallucinations, manipulation and bias on every response — no backend required.

✦ Shipped Product · 2024–2025

Real-time Hallucination,
Bias & Manipulation
Detection

A production-grade Chrome extension running a two-phase NLI pipeline on every AI response. Phase one delivers instant rule-based analysis. Phase two upgrades asynchronously with DeBERTa-v3 cross-encoder NLI and BART-large zero-shot classification — all without a server, directly inside the browser.

DeBERTa-v3 NLI BART-large Zero-shot Chrome Extension Wikipedia API JavaScript HuggingFace
91.3%
Macro F1 Score
3-in-1
Signals in One Pass
50
Ground-Truth Samples
0ms
Server Overhead
Performance
Hallucination Shield94%
Bias Detector91%
Manipulation Radar88%
Macro F1 Average91.3%

Two-phase, no server

Phase one delivers instant regex results. Phase two upgrades asynchronously with DeBERTa-v3 cross-encoder NLI and BART-large zero-shot via HuggingFace — running entirely in the browser, with Wikipedia as a live fact-check oracle. Evaluated on a 50-sample ground-truth dataset across ChatGPT, Claude and Gemini.

Projects

Selected Projects

A mix of shipped tools, active research, and promising experiments worth showing because each one reveals how I think about product, systems, and impact.

PaperTrail
01 · Justice & Systems
PaperTrail

A crowdsourced map of how broken systems actually work versus how they claim to. Courts, hospitals, government offices — the hidden manual nobody publishes.

CrowdsourcingNLPMaps APIFastAPI
CodeDissect
02 · Developer Tools
CodeDissect

Right-click any file, selection, or git diff — severity-tiered issues appear with gutter badges, hover tooltips, Dissect's Take, session history, and markdown export. Senior engineer review, no signup.

VS Code ExtensionHuggingFaceFastAPITypeScript
Lore
03 · AI Infrastructure
Lore

Organizational memory tool. Logs every Slack message and Jira ticket into a pgvector database — ask "why was the mobile app deprioritized?" and get an answer with exact source, author, and timestamp.

RAGpgvectorFastAPIGroqSlack
Uncanny Valley of Text
04 · AI Research
The Uncanny Valley of Text

Every AI model has a statistical fingerprint invisible to humans. This identifies which model wrote something, which version, and what prompt structure likely produced it.

StylometryTransformersForensics
ERPWhisperer
05 · ERP
ERPWhisperer

Talk to your ERP in plain English. "Show invoices from Q3 where payment was late and client is in Lahore." It translates, runs the query and shows results. No filters, no dropdowns.

FrappeNL2QueryERPNextPython
The Empathy Compiler
06 · Human · Creative
The Empathy Compiler

Rewrites your codebase as a human narrative told from the perspective of the data flowing through it. Makes invisible systems visible in a way no diagram ever could.

Code AnalysisNLPLLMPython
Capabilities

The Stack

Tools I reach for
when building real things.

AI & Machine Learning
PyTorchHuggingFaceTransformersLangChainFine-tuningNLPRAGSHAP
ERP & Business Systems
FrappeERPNextOdooWhite-labellingERP Architecture
Backend & Systems
PythonFastAPINode.jsPostgreSQLRedisDockerWebSockets
Frontend & Tooling
ReactJavaScriptChrome ExtensionsGitHub ActionsLinux
AI & NLP Engineering94%
ERP Architecture88%
Transformer Fine-tuning85%
Backend Engineering82%
Developer Tooling78%
DevOps & Cloud70%
About

The person
behind the work

"I am not looking for my first challenge. I am looking for a bigger one."

I am a software engineer with close to three years of experience working across AI, ERP systems, and full-stack development. I have built systems for detecting hallucinations, bias and manipulation in real-time AI responses, running directly inside ChatGPT, Claude and Gemini.

Alongside that, a significant part of my professional experience has been in architecting and customising ERP platforms — working with Frappe and ERPNext, delivering full implementations to clients, making architecture-level decisions, and leading projects from requirements all the way through to deployment.

I have worked in teams, mentored others, and shipped things that are currently in production. I take my work seriously and I hold myself to a high standard.

2025 · Present
AI Safety Research & TruthLens
DeBERTa-v3 · BART-large · Chrome Extension
2025 – 2026
ERP Architect & Developer
Frappe · ERPNext · White-labelling · Client delivery
2023 – 2026
Full-stack Development
Python · FastAPI · React · Production systems
2022
BSc Software Engineering
University of Sargodha
Contact
Open to the
right
opportunity.
Available · Open to opportunities