Debjyoti Ray

I'm a Software Dev Engineer Intern at AWS Q for Business Teams; and Research Lead at Sapphire, where I work on machine learning, and applied AI engineering.

My research interests include multimodal AI, diffusion models, reinforcement learning, interpretability of language models, and multi-agent systems. I've worked on scalable OCR pipelines, biomedical knowledge graphs, and ensemble trading agents. I'm pursuing a B.Tech in Electronics and Communication Engineering at IIIT Allahabad and a B.S. in Data Science at IIT Madras. I am currently really interested in image-vision diffusion models, agentic systems and scalable architectures.

Email  /  LinkedIn  /  GitHub  /  CV

Education

Experience

My professional experience spans AI engineering, machine learning research, and quantitative finance.

• Packaged an independent Python‑based text→SQL CLI tool (text_to_sql.py) that parses 4+ parameters (vehicle type, sensor, date range, aggregation) from natural language and emits valid PostgreSQL queries with 98% syntactic accuracy, cutting manual query writing time by 70%.
• Engineered multi‑layered SQL validation (basic & schema‑aware) to block dangerous operations and mis‑typed tables/columns, reducing invalid‑query failures by 85%.
• Developed an intelligent correction system—rule‑based fixes for common typos, LLM‑based fallback for complex errors, plus a self‑learning loop—that auto‑corrects 92% of syntax/name mistakes and preempts 60% of recurring error patterns.

• Resolved 40+ support tickets via Notion, Linear, and Slack; on user onboarding by deploying seamless OCR and parameter extraction pipelines from submitted legal documents, improving user intake efficiency by 23%.
• Engineered a multi-agent document extraction pipeline, increasing OCR accuracy by 45% and reducing data-entry errors by 29%, using: Mistral-ocr-latest for markdown conversion; Pixtral-12B; Mistral-2B SLM as a post-processing agent.

• Built a biomedical knowledge graph pipeline processing 500+ PubMed abstracts with MedPalm-2, achieving 92% accuracy in protein-pathway mapping cross-referenced with STRING-db, and Reactome.
• Scaled automated graph construction by 22.71%, creating 3000+ entities and relationships, used multi-agent architecture for, leveraging small language models (SLMs) to reduce inference time and latency.

• Developed a novel ensemble trading agent using multiple SOTA reinforcement learning algorithms combined with LLMs on stock market data, achieving a 20% increase in annualized returns and a 15% reduction in volatility.
• Directed an 8-member team under a Board of Advisors in parallel for algorithm development and interface deployment; achieving a 30% growth in net positive independent portfolio over 3 months.
• Top 4 teams out of 150,000 entrepreneurs in the STPI [Startup India Finals] – Neuron, Mohali OCW 7.0
• Raised a research grant with one of the highest valuations from NewGen IEDC, IIIT-Allahabad (Incubator) and Xartup

• Engineered a fine‑tuned BioBERT model by curating ~500 Q&A pairs from 30 PubMed papers—boosting structured‑format adherence by 45% and factual accuracy by 30%.
• Developed an Entrez‑powered citation module using NCBI E‑utilities to fetch top n PubMed results and rank via review‑weighting + 0.85ⁿ decay—achieving 88% relevancy precision.
• Built a high‑throughput backend orchestrator handling 50 queries/sec, merging LLM answers with PMID/DOI links in < 3s.

• Developed over 150 predictive alphas to forecast market trends in the USA and China, using comprehensive price-volume and sentiment datasets; delivered alpha submissions with Sharpe ratios under 3.45 and fitness above 1.2.
Certificate

Publications

A comprehensive knowledge graph-based portal for drug discovery research focusing on biomarker exploration and target identification, integrating data from multiple biomedical databases.

This paper explores the application of Deep Q-Learning to create an intelligent agent capable of mastering the classic Snake game, demonstrating the effectiveness of reinforcement learning in game environments.

Currently Working On

• Processed 4 full-text PDFs (≈240 total 256-token chunks) and embedded each chunk into a FAISS index (sub-100 ms lookup); migrating to Weaviate and ChromaDB for richer metadata support and horizontal scaling.
• Retrieved top-5 similar chunks per entity (avg. cosine ≥ 0.83) and fed them to a fine-tuned GPT-4o model to extract 150+ entity–relation triples into a standardized JSON schema.
• Computed Pearson correlations over 1,200+ node pairs, applied Louvain community detection to identify 5 biologically coherent clusters (pathways/diseases), and validated cluster significance (p < 0.05) across 300+ tests.
• Calculated degree, betweenness, and modularity metrics for 200+ nodes; visualized insights in Streamlit; planning Neo4j migration to support >10 K edges in production.
Streamlit Demo / Hosted API / GitHub / Documentation and report [to be out soon]

• Prototyping an LLM-driven multi-agent simulation that manages a shared resource pool with configurable consumption limits and regeneration rates.
• Designing and testing inter-agent negotiation and conflict-resolution protocols to balance individual utility against group sustainability under scarcity.
• Integrating LLM API calls into each agent's decision logic to weigh self-interest versus collective welfare in real time.
• Monitoring resource metrics and agent behaviors across simulation runs to extract insights on ethical trade-offs and long-term system health.

• Researching on a comprehensive survey of 100+ diffusion-based generative models to map the rapidly evolving image generation and editing landscape.
• Designing an application-centric taxonomy that categorizes methods across 2 domains (Image Generation vs. Image Editing), 3 input/control modalities (Text, Text + Image, Region-Based Control), and 3 adaptation strategies (Training-Based, Training-Free, Test-Time Fine-Tuning).
• Analyzed architectural trade-offs and performance via quantitative metrics (FID, CLIP Score, sampling speed) while addressing ethical considerations in compute cost and bias.
• Synthesized 10+ cross-cutting control and adaptation techniques (Classifier-Free Guidance, ControlNet, PEFT, test-time optimization) to formulate best-practice recommendations and future research directions.

• Deriving and coding the Flash Attention algorithm from first principles based on Umar Jamil's tutorial, reimplementing all operations in Python with Triton.
• Translating each mathematical step into custom Triton kernels to optimize GPU memory access patterns and computation throughput.
• Integrating these kernels into a PyTorch workflow and validating numerical correctness against reference implementations.
• Benchmarking prototype performance on NVIDIA GPUs to assess latency and throughput improvements over standard attention modules.

• Currently implementing Stage 2 data-engine pipelines: orchestrating a Python ETL workflow to ingest Notion snapshots and crawled web resources, apply heuristic and LLM-based quality scoring, and store processed documents in MongoDB for downstream retrieval.
• Leveraging ZenML for reproducible pipeline orchestration and Opik to benchmark RAG feature extraction and retrieval efficacy.
• Will fine-tune a Llama 3.1 8B summarization model in Stage 4 using Unsloth-managed experiments tracked via Comet, to generate concise, high-fidelity knowledge digests.
• Will evolve the system in Stage 6 by integrating agentic RAG inference with Smolagents and building an interactive online interface—roadmap to adapt based on user feedback and emerging best practices.

Projects

• Processed 260,000+ images using PaddleOCR and Qwen-2VL-2B, fine-tuning the latter on 50,000 images by freezing non-critical layers and applying LoRA adapters with 4-bit quantization and FP16 precision. Evaluated performance based on multi-language accuracy and text orientation handling; ranking us 12th out of 75,000+ participating teams.
• Utilized A100 and T4 GPUs for parallel execution, employing Liger kernels, gradient accumulation, and cosine learning rate scheduling to maximize throughput while minimizing computational overhead.
Report

• Composed a few-shot dataset by overlaying 5–10 Amazon Berkeley product cutouts per Panchayat scene—yielding thousands of unique training samples
• Augmented data with 4 methods (rotation, scaling, translation, color jitter) to quadruple scene diversity and bolster model generalizability
• Fine-tuned the DeVIT few-shot ViT (shot = 10) via meta-learning on custom pipelines (61% Python) for real-time, low-data object detection
• Co-ordinated with team to integrate a Flask (Python) backend & Next.js (TypeScript) frontend prototype to serve model inference seamlessly in Prime Video streams
Demo Video / GitHub

Hackathons

• Researched and architected an ensemble trading agent blending PPO, A2C, and SAC DRL algorithms—achieving a 20% higher annualized return compared to individual agents
• Implemented dynamic risk management with drawdown caps and volatility scaling—delivering a 1.5 Sharpe ratio and 30% lower max drawdown
• Conducted rigorous backtesting (2010–2021 training; 2021–2024 validation), validating a consistent 25% alpha over DJI and S&P 500 benchmarks
• Deployed in live markets since 2024—sustaining 99% uptime and outperforming major indices by 15% in NSE execution

• Analyzed 500+ hand‑drawn curve samples, extracting 10+ geometric features (e.g., curvature, length, closure) to drive algorithm selection and parameter tuning
• Regularized shapes by implementing detection and fitting for 5 curve classes (lines, circles, ellipses, rectangles, polygons), achieving 96% classification accuracy on a 1,000‑curve test set
• Inpainted 150+ occluded curves with a partial‑convolution network in partialconv/, reaching a mean IoU of 0.88, and deployed the end‑to‑end pipeline as a Streamlit app for real‑time demonstration
Certificate

• Built RESTful endpoints in Flask to orchestrate generative AI tasks via Google Gemini‑Pro, embedding a safety layer to guarantee valid JSON outputs across all routes
• Built a modular Next.js/React frontend with real‑time AI response streaming, responsive Tailwind layouts, and React hook state management—achieving sub‑1 s update latency for seamless user interaction
• Deployed end‑to‑end CI/CD on Vercel with custom middleware (rate‑limiting + CORS) to ensure secure, scalable inference services during the hackathon showcase

Research Summits

• Engaged with 5+ venture capitalists to validate Sapphire's strategic vision and incorporate direct investor feedback
• Presented our AI-trading framework to Ashwani Bhatia (SEBI Board Director), securing regulatory insights and endorsement of our risk-management protocols
• Delivered a 20-minute pitch to a panel of 15+ academics, industry leaders, and regulators, spotlighting live trading performance and DRL innovations
• Clinched the Best Product Showcase award out of 200+ national entries, distinguishing Sapphire as India's leading AI-powered trading solution

Exams

JEE MAINS 2022: 99.28 percentile
WBJEE 2022: 99.72 percentile
JEE ADVANCED 2022: 98.02 percentile
KVPY SA 2021: Qualified