All Projects

Central production CV+analytics API for a British American Tobacco retail-execution program — live Hugging Face deployment (merges the BAT_Master flagship and its earliest prototype lineage).

Unilever Bangladesh retail-execution API — 10+ versioned YOLO models with cross-model reconciliation (Display Audit, QPDS, SOS, MTSOS), live Hugging Face deployment.

Retail item-detection + agent face-verification API for a Malaysian deployment.

Multi-module retail-audit API: Display Audit, POSM, Sachet, and Share-of-Shelf.

Mobile-financial-services (Nagad/bKash/Rocket/Tap) branding detection API.

Versioned (v12) snapshot of the Nagad detection API.

Async object-detection & counting microservice (reusable detection pattern).

Object distance/spacing estimation from detections (pixel-to-metric).

Competitor-product detection experiment (YOLO + Gemini vision).

Team recognition + directory-lookup API.

Browser AR camera demo (WebSocket + OpenCV) — maps to the AR validation tool.

Multi-module AI model store/API unifying BAT, Unilever, face recognition, OCR, and conversational-NLP services behind one Hugging Face deployment.

Identity verification by matching a selfie against an NID photo (face embeddings).

Two-image face verification API (dlib encodings).

Real-time face & landmark detection (OpenCV + MediaPipe Holistic).

Custom PyTorch Bangla intelligent character recognition (CNN + CTC).

Detection-driven OCR field-extraction demo (detect-then-read).

Hybrid OCR: Google Cloud Vision + Gemini 1.5 Flash + NLTK.

Production real-time conversational voice agent (WebRTC → STT → LangGraph → TTS).

Hospital appointment-booking voice & chat agent — LangGraph slot-memory core with a Gemini Live audio bridge, Twilio/email tool-calling, and WebRTC transport.

Multi-agent, self-correcting RAG over MongoDB (Planner→Executor→Critic).

Clean MongoDB retrieval-augmented-generation pipeline (E5 embeddings).

Bangla document RAG Q&A with strict grounding (FAISS + Gemini).

LangGraph learning notebooks & agent-workflow experiments.

Intent chatbot dispatching to operational data functions (UBL variants).

Intent-classification chatbot over spreadsheet data (bag-of-words).

Rule-based & TF-IDF retrieval chatbots (pre-LLM techniques).

Async Telegram service bot with inline menus & encrypted config.

Local, GPU-accelerated voice assistant loop (whisper → local LLM → TTS).

Bangla text-to-speech fine-tuning (Coqui VITS + grapheme frontend, expanding toward XTTSv2).

Speech-to-text dataset builder (Cloud Speech + Gemini normalization).

Audio brand-mention detection for BAT (AssemblyAI + fuzzy regex).

Campaign-phrase compliance detection from field audio (NLTK + fuzzy match).

Two self-hosted Ubuntu GPU servers hosting production AI + training (internal).

GPU instance selection & provisioning automation (Vast.ai cost optimization).

Docker & containerization practice (Compose, cheat sheet).

ML/CV engineering toolbox — dataset prep, YOLO annotation, training utils.

Fastfetch config presets & ASCII art (Arch / Hyprland dotfiles).

Hybrid IP-gated attendance system (internal production) — Fernet-encrypted creds.

Django REST API demo + Python client.

First Django project (StudyBud discussion-rooms clone).

Django practice sandbox (models/views/URLs/templates).

HTML/CSS learning lessons (numbered, per-page stylesheets).

MongoDB setup & usage notes underpinning the DB-backed projects.

Algorithm practice — LeetCode solutions in notebooks.

Personal configuration store (reusable setup bundle).

Image filtering via convolution — blur, sharpen, edge detection.

Global, adaptive & Otsu thresholding for binarization/segmentation.

Erosion, dilation, opening, closing, gradient & hat operations.

Contrast enhancement — global equalization & adaptive CLAHE.

Vanishing-point detection (Canny + Hough + RANSAC).

Homography & perspective warping with interactive click-to-warp.

Chessboard calibration & lens-distortion correction (intrinsics).

Affine transforms — translation, rotation, scale, shear.

Feature matching from classical keypoints up to GPU LightGlue.