SAG-RAG: Speculative Agentic Graph RAG Platform
Built a RAG system combining speculative query planning, hybrid retrieval, reranking and evidence-grounded synthesis with provenance tracking for reliable answers.
Hi, I'm Atikul Islam Munna
AI-driven Software Engineer
learning ✦ exploring ✦ experimenting
- 52 public repositories
- 2.2k contributions in the last year
- 131 pull requests
- 137 stars earned
About Me
I'm an AI Engineer working on computer vision, LLM-powered systems and applied machine learning. I currently work at The Data Island, where I help develop production computer vision pipelines for retail analytics and industrial surveillance.
My research interests include deep learning, multimodal systems, edge AI and robotics, with work accepted and under review at peer-reviewed venues.
North South University
BS in Computer Science and Engineering
January 2022 - December 2025
Major: Artificial Intelligence
Chittagong College
Higher Secondary Certificate (HSC)
2018 - 2020
Group: Science
Experience
- Maintain and operate the U-Lens production Retail Intelligence system, ensuring continuous shelf compliance and reporting for Unilever retail analytics.
- Train, fine-tune and deploy AI models to improve detection accuracy and support new SKUs or product categories as they're introduced.
- Manage inference and data pipelines for product detection, SKU clustering and classification.
- Resolve daily AI-related production issues and tickets in collaboration with the customer support team.
- Collaborate with cross-functional AI team to develop and optimize scalable computer vision pipelines for production environments
- Oversee end-to-end dataset lifecycle, including data collection, cleaning and annotation
- Conduct PoC validation for early-stage computer vision projects
Tools and Technologies
Languages
- Python
- JavaScript
- TypeScript
- Bash
- C
LLM/NLP
- Hugging Face (Transformers, Accelerate)
- sentence-transformers
- spaCy
- LangChain
- LangGraph
ML/DL
- PyTorch
- TensorFlow
- OpenCV
- ONNX Runtime
- TensorRT
- CUDA
Databases
- Qdrant
- Elasticsearch
- Neo4j
- Redis
- SQL
- PostgreSQL
- MongoDB
Backend and APIs
- FastAPI
- Node.js
- Express
- WebSocket
- Kafka
Cloud and DevOps
- Docker
- AWS (Lambda, S3, SQS)
- GitHub Actions
- Cloudflare R2
- OpenTelemetry
- Prometheus
- Vercel
- Railway
- Render
- Supabase
- Upstash
Frontend & Visualization
- React
- Next.js
- Vite
- Streamlit
- Gradio
- Grafana
Mobile
- Flutter
- Kotlin
- Dart
Testing and Quality
- Pytest
- Jest
- Vitest
- MyPy
- Ruff
- ESLint
Projects
Common Ground: AI-Mediated Deliberation Platform
Built and deployed a full-stack AI-assisted deliberation platform that transforms multi-participant positions into grounded steelmans, shared foundations, and clearly classified disagreements. Implemented secure authentication, moderation, privacy-aware exports, email notifications, automated quality benchmarks, and AWS backups and monitoring.
SmartForm: AI-Based Exercise Form & Rep Tracking
Built an on-device Android app for real-time pose and hand tracking to analyze form, count reps and provide posture feedback. Added gesture-based controls and per-user calibration with robust rep quality scoring.
Townhall API
Built a production-ready multi-tenant workspace management REST API with passwordless auth, role-based access control and real-time email invitations, deployed live on Railway.
Additional Projects
SourceCast: Evidence-Grounded Media Research Workspace
Built and designed an AI-powered research workspace that turns long-form audio and video sources into timestamped transcripts, searchable evidence and citation-backed answers.
Ledgerly Invoice Processor
Built a serverless AI invoice processing platform with direct uploads, structured extraction, validation, human review and monitoring.
VaultCLI
Built a security-focused Python CLI for encrypted file vaults with outer and hidden-volume workflows, including authenticated encryption, passphrase-based key derivation, integrity verification and streaming file handling.
DevLens: GitHub Repository Intelligence
Built a full-stack AI-powered repository intelligence platform that analyzes GitHub codebases and delivers architecture dashboards plus retrieval-augmented chat over indexed code.
Decision Archaeology
Built a full-stack decision intelligence platform for capturing time-locked decisions, tracking outcomes and surfacing reasoning patterns through AI-powered analysis.
CivicFix: Community Issue Reporting Platform
Built a full-stack civic-issue reporting platform with map-based reporting, community verification and a role-gated admin workflow that tracks each report from submission to resolution.
Multi-Step Research Assistant Platform
Built a production-style research automation platform with multi-node planning, evidence gathering, analysis and citation-aware report generation with live progress tracking.
WebTorrent P2P Stream App
Built a browser-based P2P watch-party streaming app with WebTorrent/WebRTC that supports room signaling, synchronized playback, chat, reconnect recovery, tracker failover, subtitle upload (.vtt/.srt) and validation reporting.
COWFS: Copy-on-Write Filesystem
Built a userspace copy-on-write filesystem with FUSE that versions every write, supports snapshots/restore/GC/diff and ships with tested CI/CD and automated PyPI releases.
TraceBack: Explainable Backup & Restore Tool
Built a Rust-based explainable backup and restore tool with deduplicated snapshots, a guided terminal UI, integrity checks, restore rehearsal, storage analysis and filesystem remote sync.
unuMCP
Built a full-stack platform that turns any OpenAPI specification into a tested, ready-to-ship TypeScript MCP server, using deterministic code generation with a human approval gate, isolated Docker-sandbox testing, and an LLM-powered self-repair loop.
Dataset Quality Analyzer (DQA)
Built and deployed a secure computer-vision dataset auditing platform for YOLO and COCO, with asynchronous cloud audits, integrity and leakage detection, actionable reports, and automated CI/CD on AWS.
Predictive Maintenance MLOps Platform
Built an end-to-end predictive maintenance platform with validation-gated ensemble selection, production inference APIs and a dashboard for model decisions and live checks.
Realtime Fraud Detection Pipeline
Built an end-to-end fraud detection pipeline with offline training, stream-style scoring, analyst feedback ingestion, online model updates and Prometheus/Grafana observability for latency, anomaly flow and update health.
High-Throughput Distributed File Service
Built a distributed file service with resumable chunk uploads, HTTP range downloads, checksum validation and reliable state transitions with auth and audit logging.
Loom: Log-Observer & Monitor
Built a high-performance real-time log aggregation CLI with multi-source tailing, structured parsing, resilient rotation handling and a live WebSocket dashboard for throughput, severity trends and health metrics.
dgvis: Dependency Graph Visualizer
Built a Python CLI that parses dependencies across 7 ecosystems (npm, pip, Go, Rust, Ruby, Maven, YAML) and builds directed graphs from scratch. Detects cycles, ranks critical packages via PageRank, computes blast-radius impact, and runs offline security (OSV) and license (SPDX) audits, all visualized in an interactive D3.js dashboard.
NativeNote: Offline AI Notepad
Built an offline-first AI note-taking app in Flutter with rich-text editing, local persistence and on-device summarization using an ONNX-based runtime pipeline.
Resolve: Habit Streak Timer
Built an offline Android app in Flutter to track habit streaks with a panic mode, relapse insights, milestone notifications and a native Kotlin home-screen widget.
Complexity Lens
Built a VS Code extension that displays inline cyclomatic complexity via CodeLens with a sortable sidebar report and JSON/CSV export across JavaScript, TypeScript, Python, Java and C# codebases.
Federated Learning with Differential Privacy
Built a research framework for comparing FedAvg, client-level DP-FedAvg and sample-level DP-SGD across synthetic, MNIST and CIFAR-10 datasets. The project includes configurable federated simulations, RDP privacy accounting, IID/non-IID partitioning, experiment sweeps, report generation and a dashboard for privacy-utility analysis.
Anomaly-Aware Probabilistic Forecasting for Multivariate Time Series
Built a research framework for joint probabilistic forecasting and anomaly detection on multivariate time series. The project implements MDN-LSTM and regime-aware joint models, full SMD benchmark evaluation, calibrated thresholding, persistence-based false-alarm reduction, CRPS/Energy Score metrics and reproducible experiment reporting with tested pipelines.
Cross-Chain AI Provenance Verification
Designed and implemented a blockchain protocol for carrying privacy-preserving AI model attestations across chains, combining zero-knowledge verification, permissioned-to-public relay flows and automated end-to-end orchestration.
Research & Publications
Bengali MemeNet (BMN): A Multimodal Framework for Sentiment Classification of Bengali Memes
ACM Transactions on Intelligent Systems and Technology, 2026.
LinkBone Fracture Detection Using Vision Transformers: A Comparative Analysis of the Pooling-based Vision Transformer (PiT) and the Causal Transformer (CaFormer) Models
2nd IEEE Conference on Secure and Trustworthy Cyberinfrastructure for IoT and Microelectronics (SaTC), 2026.
LinkFeature Engineering and Ensemble Classifiers for Robust Cardiovascular Disease Detection
5th International Conference on Sentiment Analysis and Deep Learning (ICSADL), 2026.
LinkWorks in Progress
-
ChainAttest: Cross-Chain Verification of Privacy-Preserving ML Attestations from Permissioned to Public Blockchains
-
HAMP: Hardware-Aware Mixed-Precision Quantization via Sensitivity-Guided Evolutionary Search
-
Edge-Native Autonomous Manipulation: A Hybrid Imitation-Reinforcement Learning Framework on Constrained Hardware
-
Artificial Dexterity: A Unified Survey of Robotic Hand Architectures, Multimodal Perception and Embodied AI