PDF REDACTION
AI document-redaction system using Gemini to auto-detect and redact PII from sensitive medical records — with version control for revisions and audit trails.
► UPLINK ESTABLISHED · PORTFOLIO_2026
SRC NLP & LLMs · KNOWLEDGE GRAPHS · OCR / DOCUMENT AI
ROLE AI ENGINEER · STATUS ACTIVE · 2026
► 01 · OPERATOR PROFILE
I'm Muhammad Awais Bin Adil — an AI engineer specializing in NLP, knowledge graphs, and document intelligence. I've shipped PII-redaction systems for medical records and live video, clinical cohort search over unstructured data, and OCR pipelines that read licenses, insurance cards, and radiology reports — moving models out of the notebook and into production.
↳ KEEP SCROLLING — THE BRAIN IS ABOUT TO CHANGE ITS MIND
► 02 · GRADIENT DESCENT, LIVE
LANGUAGES · ENGLISH C1 · URDU C2
► 03 · ATTENTION IS ALL YOU NEED
AI document-redaction system using Gemini to auto-detect and redact PII from sensitive medical records — with version control for revisions and audit trails.
Real-time broadcast moderation that detects and redacts PII, illicit content, and violent threats mid-stream, fusing computer vision, NLP, and streaming pipelines.
Edit audio by editing the transcript — speech-to-text, AI voice cloning, and synthesis keep the speaker's voice seamless when words are added or changed.
Schema design and an automated construction pipeline ingesting insurance-claim data into a real-time Neo4j knowledge graph.
Natural-language search of clinical-trial criteria across unstructured records — RAG on Milvus with recursive chunking, Mxbai reranking, and Llama 3.1 8B, built atop an earlier NER-driven cohort engine.
Extracts names, DOB, and member IDs from driving licenses, insurance cards, and radiology/pathology reports using PaddleOCR, Tesseract, YOLO, GOT, and Llama 3.2 Vision.
► 04 · SHIPPED IN PUBLIC
An autonomous fine-tuning pipeline where every agent has veto power. Describe a task in plain English — agents source data, build an independent hash-locked eval, train a LoRA, and ship to Hugging Face only if the run earns it. A justified refusal ("don't fine-tune — here's a prompt that already wins") counts as success.
A "23andMe for open-weight models" — fingerprints an LLM from its weights alone and
tells you, with a calibrated probability, which base model it descends from, no
matter what the README claims. Byte-range sampling reads a 7B model in ~2 min, and
decompose even recovers a merge's parent mixture weights.
Turn a phone video into a language-queryable 3D scene. It composes three frozen foundation models — Depth Anything 3, SAM 2, and SigLIP — with no per-scene training: build on a free Colab T4, then type "the red backpack" and watch it light up in your browser, all queried on a plain CPU.