Curriculum Vitae
Senior Software Engineer · Dhaka, Bangladesh
info@shamspias.com · shamspias.com · github.com/shamspias · linkedin.com/in/shamspias0
Profile
Software engineer with over five years of professional experience designing scalable backend systems, production machine-learning infrastructure, and developer tooling. Core expertise in Go and Python, with hands-on experience deploying large language models at scale, building multi-tenant platforms, and shipping real-time computer-vision pipelines. Active research interests in computational biology — anti-inflammatory peptide prediction and structure-aware virtual screening for drug discovery. Comfortable across the whole path from research prototype to deployed service, including IoT and embedded edge inference. Open-source contributor and Dify plugin author.
Professional Experience
Senior Software Engineer — Mevrik, Dhaka · Aug 2026 – Present
Mevrik is an AI-powered customer-experience platform serving telecom operators and 700+ SMB customers.
- Work across the AI surface of the product: conversational agents, retrieval, and the model serving infrastructure beneath them.
- Design and document feature specifications for engineering teams and leadership.
Founder — AlgolyzerLab · 2024 – Present
An applied-AI studio for work that does not fit a standard product roadmap.
- Athlete Intelligence — multi-tenant athlete performance management system for elite sport (Go +
chi+ GORM + PostgreSQL, Next.js 16 / React 19 frontend), with an IOC-aligned clinical domain, strict tenant isolation, injury and illness registers, screening, workload, return-to-play tracking, and a CUDA RTMPose (Halpe-26) video pipeline that converts ordinary clips into clinical movement metrics via 2-D geometry and multi-view triangulation. - ChemberIQ — full-stack medical case-management platform for doctors and clinics: Go/Gin REST API, Vue 3 SPA, WebSocket doctor↔patient chat, an offline-first Flutter doctor app that auto-discovers its backend and delta-syncs the medical dataset, and a purpose-built in-memory search engine (prefix → word → contains → token → fuzzy) that replaced Meilisearch.
- Agri Remedy — drone-based crop-disease monitoring: FastAPI backend with a DINOv2 linear-probe classifier, canopy/lesion localization, ISO-6709 GPS extraction from video metadata, bilingual (English / বাংলা) treatment guidance, and a React 19 + TypeScript SPA.
- Cricket Bowling Analyzer — biomechanical analysis of bowling actions from video: 18 parameters at 30+ FPS via MediaPipe pose estimation, automatic ICC 15° elbow-extension compliance checks, and PDF clinical reports.
- Eventaic — event-responsive advertising platform (FastAPI, Vue, PostgreSQL) that turns a live event plus a product into a complete, evaluated ad campaign.
Senior Software Engineer — Evoclick, Moscow (Remote) · Jun 2024 – Jul 2026
- Deployed Qwen 3.5 27B (INT4 and FP8 quantized) on vLLM for production workloads serving 1,000+ concurrent users; tuned throughput through continuous batching, KV-cache sizing, and quantization strategy selection.
- Built an OpenAI-compatible LLM Gateway (FastAPI, httpx) as a routing proxy with rerank support, error enrichment, and benchmark endpoints across multi-tenant workloads.
- Deployed and operated embedding models, vision-embedding models, and document regioning models on production GPU infrastructure.
- Implemented page-index RAG, combining layout-aware region extraction with vector retrieval for document understanding.
- Built an edge-vision pipeline with YOLO on NPU hardware for real-time anomaly detection, with offline-safe buffering for intermittently connected sites.
- Developed multi-agent services with LangGraph and Dify including human-in-the-loop review, retries, and provider fallback; shipped tool integrations over the Model Context Protocol (MCP).
- Added observability hooks — structured logs, trace IDs, and evaluation harnesses — to track accuracy, drift, and SLA breaches across deployed devices.
Software Engineer — Evoclick, Moscow (Remote) · May 2023 – Jun 2024
- Designed legal-tech RAG systems (document intake → retrieval → answer drafting); improved retrieval quality over keyword baselines through better chunking, embedding selection, citation prompting, and reranking.
- Fine-tuned LLMs with LoRA/QLoRA (Unsloth) for domain-specific tasks; led the internal R&D track on agents and retrieval.
- Architected a multi-tenant AI orchestration platform (FastAPI, Vue 3, PostgreSQL/pgvector, Docker) with a DAG-based workflow engine and SSE streaming — comparable in scope to n8n or Dify.
- Built and maintained Dify plugins and workflow integrations for client deployments (GigaChat provider, FLUX Fill Pro, Together AI Image, Telegram Bot).
Software Engineer — OMNAIBLE, Amsterdam (Remote) · Oct 2022 – Apr 2023
- Added ML-powered recommendations to an e-commerce analytics platform, exposed as FastAPI microservices with caching and background jobs.
- Improved API responsiveness through query refactoring and pagination; wrote integration tests that stabilized releases.
Software / DevOps Engineer — CodeSmith Tech Ltd., Dhaka · Mar 2021 – Aug 2022
- Automated CI/CD pipelines with Docker; hardened REST and GraphQL APIs with monitoring and structured logging for faster incident resolution.
- Maintained deployments and unit tests (PyTest), increasing release confidence.
- Worked on IoT-integrated backend systems for sensor data ingestion and device communication.
Education
B.Sc. in Computer Science & Engineering — Daffodil International University, Dhaka · Jan 2016 – Nov 2020
- Coursework: Data Structures, Algorithms, Machine Learning, Data Mining, Computer Networks, Database Systems, Software Engineering, Embedded Systems, Robotics.
- Final-year research on machine-learning applications in computational biology.
Technical Skills
| Area | Tools |
|---|---|
| Languages | Go (Gin, chi, concurrency, actor model), Python (FastAPI, Django), TypeScript/JavaScript, C/C++ |
| Agents & RAG | LangGraph, LangChain, LlamaIndex, Dify, MCP, function calling, DAG workflow engines, evaluation harnesses |
| LLM Ops | vLLM (INT4/FP8 quantization), continuous batching, KV-cache tuning, embedding and vision-embedding models, reranking, benchmarking |
| ML / DL | PyTorch, TensorFlow, LoRA/QLoRA fine-tuning, distillation, GGUF quantization, linear probing, stacking ensembles |
| Computer Vision | YOLO, SAM, DINOv2, MediaPipe, RTMPose, OpenCV, NPU/embedded edge inference |
| Comp. Biology | RDKit, ChemProp, ESM-2 protein embeddings, molecular docking, scaffold-aware splits, ChEMBL/BindingDB/PubChem/COCONUT pipelines |
| Data | PostgreSQL (pgvector), MySQL, MongoDB, Qdrant, Pinecone, FAISS, Redis, ETL |
| Infrastructure | Docker, Nginx, Coolify, CI/CD, GPU deployment, Linux administration, S3/R2/MinIO |
| Frontend | Vue 3, Next.js, React, Flutter (working proficiency for full-stack delivery) |
| Practices | PyTest, observability and tracing, latency/cost budgets, Celery/async jobs, technical specification writing |
Research
AIPpred-Stack: Prediction of Anti-Inflammatory Peptides using Ensemble Learning with an Experimentally Validated Dataset
Abdullah Al Mamun, Shamsuddin Ahmed, Francis M. Bui — IEEE CCECE 2026 · University of Saskatchewan and Daffodil International University
- Two-stage stacking ensemble combining ~2,282 hand-crafted sequence descriptors (AAC, DPC, DDE, CKSAAP, CTD, PAAC, QSO, autocorrelation, physicochemical) with 1,280-dimensional ESM-2 protein language model embeddings.
- The core contribution is a dataset correction: prior work drew negatives from random UniProt proteins, which turns the task into “short peptide vs. long protein”. We use experimentally validated negatives from the same IEDB T-cell assays as the positives, and report the harder, more honest numbers that result.
AI-Driven Discovery of Natural-Product Inhibitors of Viral Proteases (DeepNatProtease, ongoing)
- Integrates per-virus bioactivity data from ChEMBL, BindingDB, PubChem BioAssay, ZINC, and COVID Moonshot with full provenance; standardizes molecules and harmonizes units to nM.
- Enforces scaffold-aware (Bemis–Murcko) splits with zero scaffold overlap, then trains virus-specific model families (ChemProp GNN ensembles, RF/XGBoost/LightGBM/DNN) selected on pre-registered metrics.
- Screens the COCONUT natural-products library (>400k compounds), filters for PAINS/toxicophores and scaffold diversity, and docks survivors against ligand-informed protease structures — prioritizing candidates by consensus of ML probability, docking score, and drug-likeness.
- Targets: HIV-1 protease, HCV NS3/4A, SARS-CoV-2 Mpro, Dengue and Zika NS2B-NS3.
Other interests
- NLP & lexical semantics — context-aware lexical substitution with lightweight models; evaluation methodology for synonym generation (LexSubLM-Lite, benchmarked on SWORDS, ProLex, TSAR-2022).
- AI systems — multi-agent orchestration architectures, efficient LLM inference at scale, and layout-aware retrieval for document understanding.
Selected Open-Source Projects
- Reins — a lightweight agent harness you bolt onto an existing app. Point it at your functions or ORM models and it derives intent-named capabilities, keeps reads and writes strictly separated in code, gates writes behind policy and approval, and runs everything through your own validation instead of raw SQL.
- Veldra — a self-hostable, local-first agent-harness platform where an agent is data, not code: a versioned
AgentSpecrow that the runtime interprets. Natural-language agent construction, MCP connectors, editable RAG knowledge bases, a visual workflow builder, agent teams, and feedback-driven self-improvement. - Fennec — zero-dependency Go library for SSIM/MS-SSIM-guided perceptual image compression: worker-pool batching, auto format selection, Lanczos-3 resize, target-file-size engine. 60–90% size reduction at SSIM ≥ 0.94.
- Clawkido — Go actor-model multi-agent swarm engine: goroutine-per-agent routing, swarm handoff, extensible skills, provider fallback (OpenAI/Groq/Ollama), Telegram and Discord integration. Single binary, ~50 MB RAM.
- VoidMon — high-performance Go TUI system monitor with per-core CPU, cross-platform GPU (NVIDIA/AMD/Intel/Apple Silicon), disk I/O, network, and process tracking.
- DeepNatProtease — the virtual-screening pipeline described above: curation, scaffold-aware splitting, model training, COCONUT screening, docking, and prioritization.
- LexSubLM-Lite — laptop-friendly lexical substitution toolkit: prompted generation via 4-bit causal LLMs, POS and morphological filtering (spaCy + pymorphy3), log-prob and cosine ranking, research-grade metrics (P@1, Recall@k, GAP, ProF1), and a YAML model registry for zero-code model swaps.
- RAG-Scout — test harness that benchmarks sparse, dense, hybrid, late-interaction, and reranked retrieval stacks on any Q/A dataset and reports which one to build on.
- VibeVoice Studio — FastAPI + web UI for Microsoft VibeVoice TTS: voice training from uploaded or recorded audio, multi-speaker synthesis, cloning, and a voice library.
- PlugBot — Next.js dashboard and FastAPI backend for managing many Dify apps and Telegram bots: streaming replies, health checks, encrypted secrets, one-command Docker deploy.
- Dify plugins — GigaChat provider, PDF → image converter, Nmap network scanner, Together AI image, FLUX Fill Pro, and Telegram integration.
Certifications
- HackerRank — Problem Solving (Basic, Intermediate), Python
- Google — Data Analytics, Machine Learning, and AI fundamentals
Languages
- Bangla — native
- English — fluent (professional working proficiency)