Md. Faysal Mahmud — Software & Agentic Systems Engineer
Dodging the bullets of life one pixel at a time
The Library
Ancient scrolls of knowledge (Education)
B.Sc. Computer Science & Engineering
Undergraduate thesis on LiteHand3D — a knowledge-distilled vision transformer for real-time, occlusion-robust 3D hand mesh reconstruction.
Higher Secondary Certificate
Science
The Arsenal
Inventory (Skills)
Runes (Languages)
Scrolls (Frontend)
Tomes (Backend & Data)
Artifacts (AI & ML)
Grimoires (LLM, RAG & Agents)
Devices (Mobile, XR & Hardware)
Wards (Testing & DevOps)
The Arena
Dungeons cleared (Selected Projects)
ZenLearn
Champion — BUET CSE Fest 2026AI-integrated learning platform. Multi-agent content generation chains document retrieval, generation and Edge-TTS/Manim media synthesis into one pipeline, with a retrieval layer that parses and chunks each format on its own terms — AST boundaries for code, section-aware for PDFs, slide units for decks. Generated lab exercises run through a sandbox that syntax-checks and unit-tests them before they reach a student.
SynapseED
Finalist — BRACU AI Hackathon 2025AI education management system with real-time collaboration, content generation and exam taking. LangGraph agent workflows are abstracted across OpenAI, Gemini, Groq and OpenRouter, so switching provider is a config change rather than a rewrite.
codeforces-mcp
An MCP server that gives coding assistants access to Codeforces practice data — six read-only tools with typed schemas, caching and rate limiting over the public API. Tested offline against recorded fixtures, with live tests that catch upstream API drift.
Drishty
Voice-controlled assistive Android app for blind and visually impaired users. Full TTS/STT pipeline for in-app navigation, live AI camera scene description, real-time obstacle detection, and voice-guided turn-by-turn directions through the Maps SDK.
TechoPhile
Social platform with a real-time feed, WebSocket chat, WebRTC peer-to-peer video calls, a live collaborative whiteboard, and chat-with-PDF over a RAG pipeline.
PhotonChat
P2P messaging over the BB84 quantum key distribution protocol, simulated in OMNeT++ with photon polarization states and QBER-based eavesdropping detection. Models adversarial channel noise to check the theoretical security bounds hold.
AniConnect
iOS app for movie enthusiasts to discover films, join discussions and get AI-powered personalized recommendations.
The Campaign
Contracts fulfilled (Experience)
Software Engineer
2026 - Present- Built a hybrid search engine in Rust as personal work, demonstrated it, and had it integrated into the product: BM25 and fuzzy lexical retrieval fused with ONNX embedding search via Reciprocal Rank Fusion, under 40 ms, fully offline.
- Shipped it as a reusable, documented library that the Deno backend calls through a C-ABI FFI boundary, with docs covering indexing, ranking and the integration contract.
Full-Stack Developer
2025 - 2026- Built and maintained client-facing applications and REST APIs with Next.js and Node.js, working through Git branching, pull requests and code review with international clients.
Full-Stack Developer
2024 - 2025- Engineered marketplace features for digital asset trading on a Node.js REST backend with a relational data model.
The Archives
Research & published works
LiteHand3D: Lightweight Uncertainty-Aware Knowledge-Distilled Vision Transformer for Real-Time Occlusion-Robust 3D Hand Mesh Reconstruction
Undergraduate Thesis, KUET (2026)
Distilled a 632M-parameter ViT-H teacher into a 33.8M ViT-S student and shipped it as a real-time streaming service; ONNX int8 quantization reached 24.5 FPS on CPU while retaining accuracy. A per-vertex aleatoric uncertainty head, calibrated against physical contact ground truth, routes compute and stabilizes temporal fusion to reduce jitter.
[ Watch Demo → ]Long-Form Bengali ASR & Speaker Diarization via LoRA-Adapted Whisper and Overlap-Aware Powerset Modeling
KUET_TensorBit @ DL Sprint 4.0 — Top 10%
LoRA fine-tuned Whisper-medium for long-form Bengali ASR with chunked inference, paired with overlap-aware powerset diarization.
[ Hugging Face Model → ]NomNet: A Dual-Stream Convolutional Network for Latent Occlusion Mask Estimation and AR Rendering
In preparation
Dual-stream convolutional architecture for estimating latent occlusion masks to drive correct depth ordering in augmented reality rendering.
The Trophy Hall
Loot & titles (Awards, Contests & Leadership)
Competitive Programming3.141f · max 1223
Guild Ranks (Leadership & Community)
- President—Bit2ByteSoftware Research & Development Community, KUET
- President, Technical Affairs—HACKHardware Acceleration Club, KUET
- General Secretary—K-MiNDSKUET Machine Intelligence & Data Science Community
- Key Organizer & Problem Setter—BitFest 2025AI & API Hackathon, KUET