Koshik Debanath

Software Engineer

Comilla, Chattogram, Bangladesh3 yrs 2 mos experience

AI EnabledAI ML Practitioner

Key Highlights

Increased tender success rate by 70% using LLaMA-2 and GPT-4.
Improved financial data retrieval speed by 40% with a RAG chatbot.
Authored 5 peer-reviewed publications on AI and NLP.

Stackforce AI infers this person is a Fintech-focused Software Engineer with expertise in AI and data-driven solutions.

Contact

Skills

Core Skills

Software EngineeringAi DevelopmentData ScienceGenerative AiNlpMachine LearningWeb DevelopmentInfrastructure

Other Skills

AI AgentsSolidityPython (Programming Language)GPT-3FlaskPineconeFastAPIStreamlitDockerXGBoostSVCLogistic RegressionPyTorchtransformersCLIP

About

I’m a Software Engineer & AI Researcher passionate about turning cutting-edge research into production-ready systems. With expertise in Generative AI, NLP, and multi-modal ML, I’ve built solutions that range from fine-tuned LLMs to full-stack AI platforms deployed at scale. Currently, I develop AI-powered products and decentralized applications—including a YouTube Live AI bot and a DAO governance platform with React, FastAPI, and Ethereum smart contracts. Previously, as a Data Scientist at Manaknightdigital Inc., I: 🚀 Increased tender success rate by 70% using Llama-2-7B + GPT-4. 📈 Improved financial data retrieval speed by 40% and accuracy by 25% via a RAG chatbot with Pinecone and Cohere. Alongside my industry work, I’ve authored 5 peer-reviewed publications on topics including low-resource NLP, misinformation detection, and AI-generated text classification, and actively contribute to open-source projects like OpenLLMetry and Pinecone Canopy. Specialties: Generative AI, LLM Fine-Tuning, RAG, Prompt Engineering, NLP, Full-Stack Development, MLOps, Web3, Smart Contracts.

Experience

3 yrs 2 mos

Total Experience

2 yrs 1 mo

Average Tenure

1 yr 1 mo

Current Experience

Universal machine inc.

Software Engineer - I

Apr 2025 – Present · 1 yr 1 mo · Sunnyvale, California, United States · Remote

I’m happy to share that I’ve joined Universal Machine Inc. as a Software Engineer - I (Remote/Hybrid)!
Universal Machine is a venture studio based in Silicon Valley, working with early-stage startups on meaningful, real-world challenges. I’m truly excited to be part of such a dynamic team and can’t wait to dive into this new journey.
I want to express my heartfelt gratitude to my family, friends, mentors, and everyone who has supported and believed in me along the way. Your constant encouragement has made this possible, and I’m deeply grateful.
Here’s to new beginnings, continued learning, and building something that matters. 🙌

AI AgentsSoliditySoftware EngineeringAI Development

Manaknight digital

Data Scientist

Mar 2023 – Apr 2025 · 2 yrs 1 mo · Toronto, Ontario, Canada · Remote

As a Data Scientist at Manaknightdigital, I led the development of AI-driven solutions across financial services, e-commerce, sports analytics, and generative media. I specialized in designing intelligent systems—from traditional ML/DL models to autonomous AI agents—that deliver real-world impact. Key achievements include:
Conversational AI & RAG Pipelines: Built domain-specific chatbots using GPT-4, Pinecone, and Flask, enabling accurate, real-time financial and sports data retrieval via Retrieval-Augmented Generation (RAG).
Autonomous AI Agents: Developed task-oriented agents for document parsing, data matching, and tender response automation by combining LLMs (LLaMA-2, GPT-4) with vector search, memory modules, and API orchestration.
Fraud Detection: Achieved 90% accuracy using XGBoost, SVC, and Logistic Regression on financial datasets; deployed the model for real-time fraud prevention.
Generative AI Systems: Built a Stable Diffusion–based image generation tool with user-defined presets, leveraging PyTorch, transformers, and Docker for scalable deployment.
AI-Powered Appraisal Platform: Implemented CLIP+FAISS for similarity-based collectible appraisal; developed FastAPI endpoints and Streamlit UI for seamless user interaction.
Data Matching for Tenders: Fine-tuned LLaMA-2 models and applied cosine similarity to segment organizational documents, increasing successful tender matches by 70%.
Scalable Infrastructure: Containerized applications with Docker; utilized asynchronous data pipelines (aiohttp, asyncio) for large-scale image and metadata ingestion.