Home Projects Skills About Contact
Data Scientist & ML Engineer

M. Ahmad
Siddiqui

Strong foundation in machine learning, NLP, and computer vision. Final-year BS Data Science student at KFUEIT.

● Plains of Punjab, Pakistan

View Projects Download Resume GitHub
Selected Work

Projects

NLP · RAG · Edge AI · FYP QuestionPrep: AI-Based Interview Simulation Platform github.com/Copperhorse/QuestionPrep
RAG Ingestion Pipeline
Docling OCR parse MD Chunker mask-unmask ChunkEval filter LFM2.5-1.2B metadata + Q/A SimHash dedup BGE+Chroma store
Answer Scoring Pipeline
User Answer + reference rapidfuzz lexical score BGE bi-encoder semantic score DeBERTa NLI entailment check Final Score + length guard
  • Three-stage answer scoring: rapidfuzz bidirectional partial-ratio (lexical) → BGE-small-en bi-encoder (semantic) → DeBERTa cross-encoder NLI (entailment/contradiction), with a length-ratio guard that penalises cut-off answers and a contradiction hard-cap that overrides high bi-encoder scores.
  • Stress detection runs entirely in the browser — AudioWorklet captures raw PCM into a ring buffer, a Web Worker computes log-mel spectrograms in pure JS, and ONNX Runtime Web runs the TCN with WASM pre-cached by the service worker for offline use.
  • RAG ingestion: Docling OCR → mask-unmask MarkdownChunker → ChunkEvaluator → two-pass LFM2.5-1.2B enrichment (metadata + Q/A generation) → SimHash deduplication → BGE embeddings in ChromaDB, all via FastAPI background tasks with SSE progress streaming.
NLP · LoRA · LLM Fine-tuning Hybrid Profanity Detection & Review Rewriting System github.com/Copperhorse/profanity_distilbert_lora
Inference Pipeline
Input review text DistilBERT+LoRA classify (95.67%) Clean pass through Flagged needs rewrite Adapter Switch load QLoRA adapter QLoRA Rewriter LFM2.5-1.2B Output clean
  • Fine-tuned DistilBERT with LoRA (1.2% trainable parameters) for profanity classification at 95.67% accuracy. The classifier decides clean or flagged before anything else runs.
  • Flagged reviews trigger a dynamic adapter switch on the same LFM2.5-1.2B base model, loading the QLoRA rewriter adapter at inference time, no second model load required.
  • Documented generation collapse following classification fine-tuning as a known adapter interference phenomenon, not a LoRA implementation bug.
RAG · Legal NLP · LLM PocketLaw: Legal Chatbot Using Retrieval-Augmented Generation github.com/Copperhorse/PocketLaw e
RAG Query Flow
User Query legal question Embed dense vector Vector Search top-k chunks LoRA LLM fine-tuned Grounded Answer banking law
  • RAG chatbot over Pakistani banking law corpus, combining dense retrieval with a LoRA fine-tuned LLM for grounded, jurisdiction-specific responses.
  • LoRA fine-tuning improved adherence to legal terminology and reduced hallucination on domain-specific queries.
Computer Vision · Transfer Learning · Accessibility Currency Classifier for the Visually Impaired github.com/Copperhorse/CurrencyClassification
Inference Flow
Camera live frame EfficientNet-B0 transfer learning PKR Note Class 10/20/50/100... Confidence Gate threshold filter Audio TTS feedback
  • Real-time Pakistani currency note classifier using EfficientNet-B0 transfer learning, for visually impaired users to identify denominations via camera.
  • Confidence threshold gate prevents incorrect spoken announcements on ambiguous or partially-occluded frames before passing to TTS.
Machine Learning · Cybersecurity · Ensemble Cyber Attack Detection System (Edge IIoT) github.com/Copperhorse/edge_iiot
Detection Pipeline
IIoT Traffic network flows Feature Eng. flow statistics Ensemble Model RF + XGB Attack Class DDoS / MITM / ... Alert 94.67% acc.
  • Multi-class attack detection on the Edge-IIoT dataset targeting DDoS, MITM, and other vectors using ensemble methods (RF + XGB).
  • Reached 94.67% accuracy after feature engineering on network flow statistics, with handling for significant class imbalance across attack categories.
Toolkit

Skills & Technologies

Core ML / DL
Python PyTorch Hugging Face LoRA / QLoRA ONNX Runtime llama.cpp
NLP & Retrieval
RAG Pipelines ChromaDB BGE Embeddings DeBERTa NLI Docling OCR
Computer Vision & Edge
YOLOv8 / YOLO11 EfficientNet ByteTrack TCN ONNX Runtime Web
Backend & Data
FastAPI Polars & DuckDB Linux uv
Background

About Me

M. Ahmad Siddiqui

Final-year Bachelor of Science in Data Science student at Khawaja Fareed University of Engineering & Information Technology (KFUEIT) CGPA: 3.35/4.00.

When I'm not deep in a model architecture or debugging an inference pipeline, I enjoy exploring mythology, working through logical puzzles, and critically reading AI-generated content on professional platforms.

Get in Touch

Contact

Download Resume