Abdul Basit Allahwala

RAG pipeline that automates regulatory-document retrieval for pharma compliance teams — moved from Llama 3 to Mistral 7B, swapped naive retrieval for SBERT embeddings, and tuned ChromaDB for production-scale corpora.

• ▸Built chunking + semantic retrieval with SBERT (~400MB) for low-latency embedding
• ▸Migrated LLM from Llama 3 → Mistral 7B for ~2× faster generation at lower cost
• ▸ChromaDB-backed vector store; FAISS migration in progress for further perf gains
• ▸Designed evaluation harness for retrieval precision on regulatory corpora

Automated platform for meeting participation, transcript processing, semantic search, and business insight generation — built around Recall.ai, OpenAI, Pinecone, FastAPI, and RAG workflows.

• ▸Built automated meeting participation and transcript processing workflows
• ▸Developed webhook pipelines for transcript ingestion, chunking, and vector storage
• ▸Implemented semantic search and RAG for contextual Q&A and SOW document generation
• ▸Integrated OpenAI for summaries, action items, and business intelligence extraction

Computer vision inspection pipeline that fine-tunes YOLO models, filters video frames, detects anomalies, and routes relevant visual evidence to multimodal vision models.

• ▸Fine-tuned YOLO models on custom datasets for automated inspection workflows
• ▸Developed video analysis pipelines with frame filtering and anomaly detection
• ▸Integrated Vision AI models for intelligent defect identification

Structured document extraction system that benchmarks and combines multiple OCR and vision models for scanned documents, forms, and business workflows.

• ▸Evaluated OpenAI Vision, Claude Vision, EasyOCR, PaddleOCR, and Tesseract
• ▸Built structured information extraction pipelines for scanned documents
• ▸Benchmarked OCR performance and optimized extraction accuracy

End-to-end platform that ingests videos and educational content, transcribes them with Whisper, and uses LLMs to generate structured quiz questions.

• ▸Used Whisper for transcription and LLMs for intelligent question generation
• ▸Automated content ingestion through the YouTube API
• ▸Generated assessment-ready quizzes from long-form educational content

Visual + collaborative recommender for fashion items — combines image-feature similarity with user-interaction signals to produce contextual outfit suggestions.

• ▸Built embedding-based item similarity over fashion-image features
• ▸Blended visual similarity with collaborative-filtering signals
• ▸Iterated on retrieval ranking and diversity heuristics

Fine-tuned YOLOv8 detection model for vehicle plates across varied lighting and angles, paired with OCR for end-to-end automated plate reading.

• ▸Fine-tuned YOLOv8 on a curated license-plate dataset
• ▸Integrated Tesseract OCR for plate-text extraction
• ▸Robust to varied lighting, angles, and partial occlusion

Production-grade scrapers using Playwright and Puppeteer for sites that aggressively defeat naive HTTP scraping — handled auth flows, dynamic content, and anti-bot heuristics.

• ▸Built parallel Playwright workers for high-throughput crawling
• ▸Puppeteer pipelines for sites requiring deeper DOM interaction
• ▸Stable session/cookie handling and retry/backoff strategies