SU/dev

AI/ML Engineer | Developer Portfolio

RAG pipelines.
SQL that answers questions.
Systems built to ship.

I build retrieval-grounded assistants, Text-to-SQL analytics, and computer vision services — connected into production-ready full-stack products.

See Live Demo Contact
01 Flagship Project
Live & Deployed

Enterprise RAG & SQL
Copilot

A production platform that routes any natural-language query to the right engine — vector retrieval for documents, validated T-SQL generation for relational data — then returns grounded answers with inline visualizations and an explainability trace.

What it does

  • Multi-agent query routing (RAG vs SQL decided at runtime)
  • Dynamic bar, line & pie visualizations from SQL results
  • Explainability panel — shows which source answered the query and why
  • Supabase pgvector + AdventureWorks SQL dataset
Launch Demo View Source
FastAPILangChainpgvectorSupabase ReactTypeScriptDocker

System Architecture

router_service.py intent classifier
↙     ↘
rag_service
sql_service
↓
explainability_service.py + VisualizationCard

Frontend components

ChatInterface.tsx ExplainabilityPanel.tsx VisualizationCard.tsx MessageBubble.tsx
02 Experience

January 2026 – Present

AI/ML Intern

E2E Rosh IT Solutions

Anonymized Industry Client Project

Computer Vision Safety Platform

  • Implemented driver fatigue detection with PERCLOS scoring using MediaPipe facial landmarks.
  • Built worker proximity monitoring, oil leak detection, and pose-estimation-based fall alert modules.
  • Developed FastAPI services for real-time inference results through REST APIs.
PythonFastAPIMediaPipePose EstimationREST APIs
03 RAG & LLM Systems

Assistant patterns selected by risk and data shape.

From retrieval-grounded knowledge bots to validated SQL analytics — each approach chosen to match the data and the stakes.

Flagship

01 | RAG

Retrieval-Augmented Generation

Grounded answers from enterprise records and academic resources — the system retrieves evidence before it generates, reducing hallucination at the source.

  • Chunking strategies & embedding models
  • Supabase pgvector semantic retrieval
  • Context assembly & citation faithfulness
  • Permission-aware authenticated access

Built: ERP Chatbot · VTU Academic Bot · Enterprise Copilot

Flagship

02 | Text-to-SQL

Natural Language to Validated Query

Natural language translated to T-SQL across multi-table joins, with validated execution and dynamic chart output — no manual SQL required.

  • Schema grounding & few-shot prompting
  • Validated query generation with error recovery
  • Bar, line & pie chart rendering
  • AdventureWorks: sales, HR, production joins

Built: Analytics Copilot · Adventure Works SQL Bot

03 | LLM Orchestration

Generative Assistants

Flexible conversations with structured outputs, controlled tool calls, and multi-agent coordination.

  • System prompting & few-shot examples
  • Guardrails, evals & output schemas
  • LangChain agent orchestration

04 | Semantic Search

Intent & Similarity

Route questions by meaning when users phrase the same request differently.

  • Embeddings & cosine similarity
  • NER and confidence thresholds

05 | Rule-Based

Controlled Workflows

Deterministic dialogue paths for fixed processes and high-control interactions.

  • Decision trees
  • Keyword and regex routing
  • Validated input slots
04 Computer Vision

Pixels, landmarks,
tracks and decisions.

Vision skills applied in industrial safety monitoring and real-time sports performance analysis.

01

Detection

Identify safety events, workers, objects, and scene signals frame by frame.

Object Detection
02

Landmarks

Facial landmarks and PERCLOS measures for fatigue monitoring.

MediaPipe
03

Pose

Skeleton-based action recognition for falls and exercise form.

MoveNet | Pose Estimation
04

Temporal Analysis

Tracking, velocity, and sequential event modelling from movement features.

Tracking | LSTM

Mobile Vision Project

ScoutX Sports Performance Analysis

Flutter app using TensorFlow Lite MoveNet for live joint tracking, rep counting, and posture-correction feedback.

FlutterTensorFlow LiteMoveNet
05 Architecture Concept

Football Match Analytics Copilot

Applying the full RAG + SQL + vision stack to a searchable match-analysis workflow.

Videoframes
Detection + Trackingplayers / ball
Pose + Velocitymovement events
SQL Storemetrics
RAG Storereports / clips
Copilotgrounded answers

Vision Evaluation

mAP, tracking IDF1, keypoint accuracy, and velocity error.

Assistant Evaluation

SQL execution accuracy, retrieval relevance, and citation faithfulness.

Stack

Python, FastAPI, TensorFlow / PyTorch, SQL, RAG, Node.js, and Flutter.

Open To Work · Fresher · Bangalore

Looking for a team building
RAG, LLM, or AI products.

I am actively looking for full-time roles and internships at AI startups and product companies. I ship fast, learn faster, and already have production-deployed work to show.

Start A Conversation