AI / Backend Engineer building production systems across computer vision, biometrics, real-time detection, agentic assistants and document intelligence — from custom YOLO and FaceNet models to self-hosted LLM inference on multi-GPU hardware.
Custom YOLO detection & training, Mediapipe landmarks, FaceNet embeddings, pose estimation, Roboflow, OpenCV.
PaddleOCR, Google Vision, DocTR, Tesseract — benchmarked and selected per document quality.
Self-hosted Gemma via llama.cpp, LangGraph routing, prompt-engineered extraction, OpenAI, Redis memory.
FastAPI async services, Docker, multi-GPU serving (5× RTX 3090), TFLite, PostgreSQL / MongoDB.
Migrated a production document-extraction pipeline off the OpenAI API onto self-hosted inference — removing per-document API cost and keeping medical documents inside in-house infrastructure. Gemma 3 27B served in full BF16 across 5× RTX 3090 via llama.cpp layer-split, behind FastAPI.
Real-time detection of live faces vs. photo, video or mask spoofs for mobile authentication.
Mediapipe facial landmarks + temporal ROI movement analysis across frame windows, served over FastAPI with a sliding-window classifier.
Remote fingerprint verification from a mobile app, meeting NADRA image-quality standards — no branch visit needed.
Multi-stage custom YOLO pipeline (hand → left/right → finger region) with preprocessing to the required pore format; TFLite for low-latency inference.
Automated identity-data capture from mobile ID-card photos to speed up bank customer onboarding.
YOLO card detection + perspective correction, then an OCR engine benchmark (Tesseract / DocTR / Google Vision) with structured parsing of the winner.
Real-time event detection & alerting for CCTV / doorbell feeds — fire, weapons, parcels, pets and falls.
Multiple custom YOLO detectors + pose estimation, async FastAPI + WebSocket streaming, and priority-based alert logic (weapon > fire > parcel).
Face-based login for mobile apps using biometric embeddings instead of passwords or PINs.
Mediapipe alignment → FaceNet embeddings, similarity-threshold matching with multi-image enrollment for more robust representation.
Voice-first assistant that answers questions and controls smart devices through one conversational interface.
LangGraph intent routing (knowledge / live search / device command), device-registration DB validation before acting, and Redis short-term memory.
Batch classification & extraction across large medical / legal PDF packets into structured Excel summaries.
Hybrid classification — regex-first with an LLM fallback for ambiguous docs — plus per-type parsers for doctor, dates and metadata, with validation.
Classifies billing documents, extracts patient fields, and creates dashboard records via API — fully automated.
Regex-first + local Gemma fallback classification, multi-engine OCR chosen by quality, prompt-engineered extraction and a validation layer before any record is written.
Open to AI / Backend engineering roles focused on document intelligence, computer vision and LLM systems.