Skip to content
View ArjunPramod's full-sized avatar
💭
Active
💭
Active
7 followers · 7 following

Block or report ArjunPramod

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.

Content in all repositories owned by your account will be closed.
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
ArjunPramod/README.md

Arjun Pramod

AI Engineer

LLM Orchestration Agentic Systems Applied Computer Vision


Email LinkedIn GitHub



I build production LLM and agentic systems — routing pipelines, retrieval layers, and multi-provider orchestration that hold up under real traffic and real constraints. Currently an AI Engineer at Nesa Software, working across LangGraph, MCP, and CPU-bound computer vision at the edge.



 Projects

SinroSphere AI
Multi-Tenant Analytics Agent
Jul 2026 — Present

▸ Natural-language analytics agent spanning 12 modules — structured routing, RBAC, guardrails, clarification handling

▸ Local BGE embeddings constrain LLM context via semantic retrieval, cutting input token usage by 80%

▸ Multi-provider orchestration across Groq, Gemini, OpenAI, Ollama — failover, SSE streaming, rolling memory, redaction

LangGraph LangChain MCP BGE Groq Gemini Ollama

Intellicam
Multi-Camera Vision System
Mar 2026 — Present

▸ Multi-camera CV pipeline — YOLO detection, pose estimation, and tracking for real-time behavioral analytics

▸ Zone/blind-spot detection, tamper detection, ArcFace + FAISS face recognition — 3 concurrent streams at 14 FPS, CPU-only

▸ Quantized with OpenVINO + NNCF (INT8/FP16) — model pooling, RTSP recovery, runtime watchdogs

YOLO OpenVINO NNCF ArcFace FAISS RTSP

DomainTune
LLM Fine-Tuning for Extraction
Sep 2026

▸ Qwen2.5-0.5B fine-tuned via QLoRA — 4-bit quantization, LoRA, completion-only loss masking for OCR receipt extraction

▸ Diagnosed overfitting via loss curves, fixed with noise augmentation + regularization — accuracy 0% → 68%

▸ Shipped as a containerized FastAPI service with automatic GPU/CPU inference detection

Qwen2.5 QLoRA PEFT Transformers FastAPI Docker

 Stack

LLM & GenAI

OpenAI Anthropic Gemini Groq RAG FAISS pgvector LoRA/QLoRA/PEFT Hugging Face

Agentic AI

LangGraph LangChain MCP Multi-Agent Systems Tool Calling

Eval & Safety

Guardrails Prompt-Injection Detection LLM Evaluation Regression Testing

Vision

PyTorch TensorFlow YOLO OpenCV Transfer Learning Grad-CAM Pose Estimation

MLOps

FastAPI Docker AWS MLflow CI/CD Git

 Recognition

▸ Best Paper Award, ICAIN 2025 — lead author, ESADN: Enhanced Spatial Attention Network for Road Accident Detection
▸ Gold Medalist, ProductKraft Expo 1.0


Building with intent, one system at a time.

Pinned Loading

  1. DomainTune DomainTune Public

    Fine-tuning Qwen2.5-0.5B-Instruct with QLoRA for structured receipt data extraction from noisy OCR text, with zero-shot and few-shot evaluation baselines.

    Jupyter Notebook

  2. rag-text-qa-langchain-huggingface rag-text-qa-langchain-huggingface Public

    Retrieval-Augmented Document QA system using LangChain, FAISS, and FastAPI to answer questions grounded in custom documents with source citations. Dockerized and deployed on AWS EC2.

    Python 1

  3. Customer-Retention-Decision-Platform Customer-Retention-Decision-Platform Public

    End-to-End Customer Retention Decision Platform that predicts churn, explains risk, segments customers, and recommends actions through production APIs and dashboards.

    Jupyter Notebook