Skip to content
View bsurya27's full-sized avatar

Organizations

@IIIT-Pune

Block or report bsurya27

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
bsurya27/README.md

hey, I'm Surya 👋

I'm an AI engineer. I build systems around models -- retrieval pipelines, agentic frameworks, finetuned VLMs. From research to production.

My work spans the full stack -- finetuning vision-language models, building eval pipelines, designing hybrid retrieval architectures, and shipping production backends.


what I've built

On The Porch -- Multimodal civic intelligence agent for Dorchester, Boston
Hybrid SQL + RAG with a 3-mode query router (95% routing accuracy) and retrieval caching. Cut response latency from 15-20s to 2-3s.
LangChain Chroma MySQL Gemini FastAPI

Boston Housing Violations -- End-to-end civic ML pipeline (BU Spark!)
16K+ violation records, multi-dataset integration, 90% accuracy violation prediction, temporal anomaly detection, geospatial heatmaps.
Scikit-learn Folium Pandas KMeans

Spatial Reasoning in VLMs -- Multimodal ML research
Fine-tuned LLaVA-1.5-7B with LoRA for spatial reasoning tasks. Turns out explicit CoT prompting actually hurts performance here. Interesting find.
PyTorch HuggingFace PEFT LoRA LLaVA

Anime Recommender -- Agentic hybrid recommendation system
16M+ user ratings across 28K anime entries scraped via MyAnimeList. SVD-based collaborative filtering + content similarity, all wrapped in an agent tool layer.
Surprise SentenceTransformers Pandas OpenAI

Restaurant Reservation Chatbot -- LLM tool-using agent with human-in-the-loop
Natural language booking requests routed into structured backend ops, with an admin approval workflow keeping humans in control.
FastAPI SQLite Gradio OpenAI


currently building

AI Research Curator -- Two-agent system that learns what you care about and gets better at it every day. One agent runs daily, scrapes the web, scores content against a preference vector in embedding space, and writes you a briefing. The other captures how you engage with that briefing and feeds signals back to drift the vector. Cold start to personalized in a few cycles.
AWS EC2 EventBridge S3 ChromaDB OpenAI Anthropic Streamlit FastAPI


stack I reach for

Python PyTorch HuggingFace LangChain FastAPI
SQL Vector DBs RAG LoRA/PEFT Agentic Systems AWS


reach out

LinkedIn · Email

Pinned Loading

  1. Data-Science-Project Data-Science-Project Public

    Jupyter Notebook

  2. Recommender-Systems Recommender-Systems Public

    Jupyter Notebook

  3. Spatial-Reasoning Spatial-Reasoning Public

    Forked from MultimodalSpaceResearchers/Spatial-Reasoning

    Jupyter Notebook

  4. ViT-CycleGAN ViT-CycleGAN Public

    Forked from junyanz/pytorch-CycleGAN-and-pix2pix

    BU Spring 25 IVC project

    Jupyter Notebook

  5. Query-Assist Query-Assist Public

    Forked from buttaRahul/Query-Assist

    An NLP model which can convert natural language verbal specifications into sql queries and executes them against a database

    Jupyter Notebook