AI/ML ENGINEER · HYDERABAD

Deepak Chaudhary

I build production AI systems end-to-end — from data pipelines and model fine-tuning to autoscaling cloud and edge deployments. Spanning LLMs, RAG, multi-agent systems, recommendation, semantic search, and computer vision.

4+
Years
9
Projects
1.8M+
Users Served
GTC 2026
NVIDIA Demo
PythonPyTorch Hugging FaceLangChain CrewAIFAISS Azure MLDocker KubernetesFastAPI NVIDIA NIMDGX Spark YOLODeepStream XGBoostSentenceTransformers OpenAIPostgreSQL

Projects

Every project is a runnable FastAPI service — clone, install, run, open in browser

Production
01
Semantic Search & Classification
Embedding-based document retrieval with weighted-vote domain classification. FAISS index + SentenceTransformers with retrieval evaluation metrics (MRR, Recall@K).
FastAPIFAISSSBERTPydantic
View Code →
Production
02
Multi-Agent Research Crew
4-agent sequential crew with live SSE streaming — planner, researcher, analyst, writer. Shared context, fallback-enabled LLM loop, real-time trace dashboard.
FastAPISSEOpenAICrewAI
View Code →
Production
03
SBERT Training Pipeline
4-stage fine-tuning pipeline for sentence transformers: data prep → training → evaluation → registry. YAML configs, Pydantic schemas, metrics dashboard.
SentenceTransformersPyTorchYAML
View Code →
04
CLIP Visual Search
Multimodal image-text retrieval with CLIP embeddings. Drag-drop UI, similarity scoring, batch indexing. Supports text→image and image→image search.
CLIPFastAPIFAISS
View Code →
Production
05
Edge Person Tracker + Dwell Analytics
YOLO + IoU tracker with ROI-based dwell-time monitoring, zone occupancy counting, heatmap visualization, and real-time WebSocket streaming. Mirrors production edge CV pipeline.
YOLOv8OpenCVWebSocketDwell Analytics
View Code →
Production
06
Resume Enhancer — Skill-Driven Agent
AI agent that reads operating instructions from editable skill files (skills/*.md) — like Claude's skill system. ATS keyword scoring, before/after previews, multi-backend (HF/Claude), hot-reloadable rules.
FastAPISkill FilesATSPyMuPDF
View Code →
Hackathon
07
ReAct Weather Agent
Pure ReAct (Reasoning + Acting) agent with iterative tool calling. SSE streaming shows Thought→Action→Observation trace in real-time. Domain-gated, supports online and offline modes.
OpenAIReActSSETool Calling
View Code →
Hackathon
08
Energy Forecaster
SARIMA vs XGBoost for 24h-ahead electricity demand forecasting. Full pipeline: data quality → cleaning → feature engineering → anomaly detection → model comparison with MAE/RMSE/MAPE.
XGBooststatsmodelspandasFastAPI
View Code →
Hackathon
09
Multimodal Contrastive Trainer
Dual-encoder (ResNet-50 + BERT-Tiny) contrastive learning with InfoNCE and focal loss. Trains to align image-text pairs. Retrieval eval with Recall@K and Mean Rank.
PyTorchResNet-50BERTContrastive
View Code →

How Demos Map to Production

Simplified, public versions of patterns shipped in production systems

Semantic Search & Classification
Domain mapping — classifying jobs into taxonomy via embedding nearest-neighbour voting
Multi-Agent Research Crew
Multi-agent compliance review with tool use (GTC 2026 demo)
SBERT Pair Trainer
Sentence-encoder fine-tuning for job-talent matching and ranking
Person Tracker + Dwell Analytics
Real-time edge CV on RTSP streams with DeepStream + IoT Hub alerting
ReAct Weather Agent
Tool-calling agents for domain-specific Q&A with structured tool schemas
Energy Forecaster
Time-series forecasting for operational planning with model comparison
Multimodal Contrastive Trainer
Multimodal retrieval & document understanding with contrastive objectives

Highlights

NVIDIA GTC 2026
Multi-agent system demonstrated live at NVIDIA's flagship AI conference
Indian Patent (Granted)
Co-inventor on UAV-based disaster management system
M.Tech, CS — ISI Kolkata
Indian Statistical Institute — one of India's premier research institutions
Platform Scale
AI systems serving 1.8M+ users on a production talent platform