Experiences
AI Research EngineerBuilt a native Assamese tokenizer (32K vocabulary, Unigram LM) on a 162M-token corpus with cross-lingual normalization, repairing 11.4% of the dataset and achieving 3.6× fewer tokens per word than Llama 3 and 46.8% better compression than Gemma. Pretrained a 240M-parameter decoder-only LLM on 14.4B curated tokens using PyTorch FSDP across a distributed A40 cluster, transferring hyperparameters from a 31M proxy model and compressing the vocabulary to 8.2K tokens to reallocate ~108M embedding parameters into model capacity.
AI Engineer InternDeveloped an agent-first, multi-turn FastAPI backend for cocktail intelligence supporting web and mobile. Implemented structured agents for intent classification, recipe generation, device control, vision-based ingredient recognition, and action cards using typed Pydantic schemas and JSON APIs. Designed persona-aware routing (emotion, occasion, readiness) and session-based context handling. Optimized latency, quality, and cost through per-agent model selection and fallback strategies. Contributed to structured responses, debugging tools, and system-level enhancements, with roadmap features including persistent memory, tool support, vector databases, and production-ready deployment improvements.
AI Engineer InternBuilt a simple chatbot for greetings and conversation tracking. Also developed a college admission chatbot using PDFs for queries on requirements, deadlines, and processes with Streamlit, Groq, LangChain, and LangChain-Groq, enabling multi-turn interactions and contextual memory.
Undergraduate Researcher (AI)Conducted AI research in semantic intelligence, ontology engineering, knowledge graphs, and large language models. Co-authored multiple conference and journal publications on recommendation systems, document intelligence, semantic reasoning, and knowledge-centric AI. Developed intelligent frameworks integrating deep learning, ontologies, and LLMs for automated knowledge organization, classification, and recommendation across multilingual and domain-specific applications.
Projects
Implemented the complete Transformer architecture from the Attention Is All You Need paper in PyTorch, including multi-head self-attention, positional encodings, encoder-decoder blocks, masking, and beam search. Trained on a custom translation dataset and visualized attention maps to analyze model behavior.
Fine-tuned Llama on English-to-SQL queries using LoRA and efficient fine-tuning, cutting inference errors by 25%. Built a preprocessing pipeline to clean, tokenize, and balance data, boosting training efficiency and SQL accuracy.
Built a Retrieval-Augmented Generation (RAG) pipeline using LangChain, ChromaDB, and an open-source LLM. The key achievements include creating a context-aware text generation system, improving query latency and response relevance by 10% through ChromaDB indexing, and ensuring production readiness by utilizing comprehensive evaluation frameworks.
Publications
Gerard Deepak, Anubrat Bora, MS Roopa, KR VenugopalOSSS is a Web 3.0 framework for automatic ontology generation in environmental journalism. It integrates Bi-LSTM and AdaBoost classification, Pearson correlation–based feature federation, Google Knowledge Graph API and YAGO enrichment, Morista’s Overlap Index for semantic relevance, and Imperialist Competitive Algorithm optimization.
Anubrat Bora, Gerard Deepak and Mohammed Salman SAThis Web 3.0 medical annotation framework targets genito-urinary and urological pathology documents. It uses knowledge encompassment with TF-IDF, generative AI via YaLM-100B for text and caption extraction, and high-density metadata generation through OpenCalais and structural-content feedback integration.
Anubrat Bora, Gerard Deepak and Samiksha ShuklaThis Web 3.0 document recommendation framework for Assamese literature combines IndicBERT classification, TF-IDF, generative semantic intelligence, AxiomiyaBERTa metadata generation, and CoSimRank-based semantic reasoning. It applies the Petraitis Index and Invasive Weed Optimization to refine document recommendations, achieving high precision, recall, and overall recommendation accuracy.
Anubrat Bora and Gerard DeepakThis Web 3.0 video recommendation framework for Assamese content integrates generative semantic intelligence, multilingual knowledge aggregation, ontology-based classification, and SimRank-driven semantic reasoning. It supports intelligent metadata generation, cross-lingual retrieval, and optimized recommendation refinement to deliver relevant and context-aware video recommendations across Assamese and English.
Anubrat Bora, Gerard Deepak and Harshabrat BoraA Web 3.0 ontology synthesis framework for Ambedkar and Peace Studies, combining OpenCalais metadata generation, DistilBERT classification, LSI seed knowledge, and Llama-4-Scout and Mistral-Medium-3 for generative expansion.
Anubrat Bora and Gerard DeepakThis Web 3.0 video recommendation framework integrates hybrid machine intelligence, generative AI, and semantic AI. It generates ontologies, selects linked-similarity features, classifies via logistic regression, applies NPMI thresholds, and optimizes using the Jian-Konrad Index and Elephant Optimization algorithm for relevant, diverse recommendations.
Anubrat Bora and Gerard DeepakThis paper proposes a Web 3.0 microblog annotation framework integrating semantic intelligence with deep learning models. It leverages domain-specific knowledge stacks to enrich metadata, applies RNN for inter-class classification, and refines top categories using XGBoost.
Anubrat Bora and Gerard DeepakThis research presents a strategic framework for classifying criminology and news datasets, enhancing their relevance for expert systems. It integrates LLaMA with XGBoost for efficient categorization and enrichment and uses standard news APIs and the Cricket Algorithm with a diversity index for further optimizations.