I build the infrastructure layer that makes LLM and AI-agent systems run in production, reliably, on serverless and decentralized GPUs, at a fraction of managed-API cost. Published ML researcher (7 peer-reviewed papers, 3 in Q1 journals) who ships the hard part: multi-agent orchestration, RAG, and GPU deployment, not just notebooks.
- 🚀 Production LLM / agent deployment on RunPod, Modal, and Vast.ai. Dockerized serverless endpoints with CI/CD, plus self-contained installers that survive real GPU-provider failures (driver mismatches, multi-GPU OOM, topology-aware placement)
- 🧠 Multi-agent systems with latent-space collaboration, where agents exchange hidden states instead of generated text, cutting tokens and inference time (open-sourced as LatentMAS-SLoRA)
- 📚 RAG / document intelligence: embedding retrieval, citation tracking, multi-hop reasoning, and LLM-as-judge evaluation
- 🎛️ Dynamic LoRA serving: hot-swappable, domain-routed PEFT adapters on a single base model, with custom LoRA training built in
- 💸 Inference cost optimization through serverless autoscaling with near-zero idle cost
📨 Tell me your model and your latency or cost target, and I'll reply with a concrete plan.
|
Osmantic / ODS (4,000+ ⭐) is a production, self-hostable AI-deployment platform for local and decentralized GPUs. I've opened 23 PRs, 14 merged upstream across the platform's schema, API, and deployment surface. Merged into main:
Open, in review:
Stack: Bash installer hardening · Docker Compose · GPU tier detection · FastAPI · pytest · ShellCheck · Vast.ai quirks |
|
Featured as community extension #5 in Gen-Verse / LatentMAS, the official repo for the ICML 2026 Spotlight paper Latent Collaboration in Multi-Agent Systems (arXiv:2511.20639). Listed alongside extensions from MIT's LAMM group. A multi-agent reasoning framework that augments LatentMAS with role-specialized, dynamically switchable LoRA adapters and latent-space collaboration, where agents communicate through hidden states rather than generated text. Includes VLM support (Qwen2.5-VL-7B), RAG integration, and RunPod serverless deployment with CI/CD. Pipeline: Planner → Critic (latent) → Refiner (latent) → Judger, with the Critic and Refiner operating purely in hidden-state space at ~200ms each. On honesty about scope: the repo documents exactly what it implements via PEFT (dynamic loading, hot-swapping, weighted merging, LRU eviction) and what it does not have versus true S-LoRA (no custom CUDA paging kernels, no heterogeneous batching). Benchmarks against vLLM S-LoRA are on the roadmap, not claimed as done.
|
|
LLM / Agent Infra
ML / DL
MLOps / Infra
Languages & Data
|
Multi-agent reasoning with dynamic LoRA routing and latent-space collaboration. Featured in Gen-Verse/LatentMAS.
|
Retrieval pipeline over URL/JSON/CSV docs with citation tracking and automated LLM-as-judge evaluation.
|
|
Serverless RunPod worker for FLUX.2 [klein] text-to-image and image-to-image, autoscaled with near-zero idle cost.
|
End-to-end serverless inference ecosystem for diffusion models with Dockerized endpoints and CI/CD.
|
|
Topology-aware multi-GPU pipeline with per-GPU VRAM capping (multi-GPU OOM fix) and self-host UI.
|
Self-hosted Flux LoRA platform tuned for A40 (48 GB) through H200, with 4-bit quantization and batch inference.
|
|
Serverless speech-to-text worker, a scalable transcription endpoint on demand.
|
Low-cost, CPU-friendly self-hosting kit for local AI tooling with secure external sharing.
|
More projects: LoRA training, NLP fine-tuning, and low-VRAM inference
| Project | What it does | Tech |
|---|---|---|
| Lora_Trainer_Imgen_Flux | Custom LoRA training pipeline for Flux image models | PEFT Flux |
| bangladeshi-tts-finetuning | Fine-tuning TTS for Bangla speech | TTS Fine-tuning |
| bangla-punctuation-restoration | Transformer-based punctuation restoration for Bangla | NLP Transformers |
| time-series forecasting | Modern time-series forecasting models | PyTorch Forecasting |
| Wan2GP | Low-VRAM video generation for consumer GPUs | PyTorch Gradio |
| Ltx-Image-to-Video | Image-to-video generation with a Gradio interface | LTX Gradio |
7 peer-reviewed papers · 3 in Q1 journals · Google Scholar · ORCID · ResearchGate
| # | Paper | Venue | Metrics | Year | Link |
|---|---|---|---|---|---|
| 1 | Machine-learning-assisted revelation of the best-performing single heterojunction thermophotovoltaic cell | Sustainable Energy Technologies & Assessments (Elsevier) | JIF 7.4 · Q1 (Scopus) | 2025 | DOI · arXiv |
| 2 | Machine-learning-enabled performance exploration of AuCuSe₄ in a thermophotovoltaic cell | Solar Energy (Elsevier) | JIF ~6.0 · Q1 | 2024 | DOI |
| 3 | Numerical studies on a ternary AgInTe₂ chalcopyrite thin-film solar cell (first author) | Heliyon (Cell Press) | JIF ~3.5 · Q1 (Multidisciplinary) | 2023 | DOI |
| 4 | Numerical prediction of the photovoltaic performance of a CZTS-based thin-film solar cell | Nano Select (Wiley) | JIF 3.5 · Q2 · ESCI | 2023 | DOI |
| 5 | Unleashing the power of open-source transformers in medical imaging (brain MRI: 99.60% classification accuracy, 0.90 Dice) | IJACSA | Peer-reviewed | 2024 | DOI |
| 6 | Spectrum estimation for voiced speech using average weighted linear prediction | [ ] | — | 2024 | * * |
| 7 | Enhancement of bone-conducted speech using deep transfer learning | [ ] | — | 2024 | [ ] |
Research themes: applied ML for materials and PV optimization, medical-imaging transformers, and speech enhancement. Journal metrics reflect the most recent JCR/Scopus release; quartile assignment varies by database and subject category.
- 🎓 B.Sc. Electrical & Electronic Engineering, University of Rajshahi, Bangladesh (Solar Energy Laboratory)
- 🏅 SQL Advanced and REST API Intermediate (HackerRank) · Deep Learning with TensorFlow (IBM) · Prompt Engineering (Vanderbilt)