DEEPAN SADHUKHAN
AI/ML ENGINEER
000

DEEPAN SADHUKHAN

AI/MLEngineer&publishedresearcherarchitectingautonomousmulti-agentsystems,multimodalvisionarchitectures,andproduction-scaleintelligence,bridgingfrontierresearchwithdeployedsoftware.

LET'S CONNECT
01
MARKETOS.
AGENTIC MARKETING PLATFORM

AI-native marketing operations platform architected around a Supervisor agent orchestrating 16 specialized sub-agents - one-click campaign intent to fully coordinated, 24/7 autonomous execution across channels.

DEEP DUO FOUNDATION
ORG
MARKETOS.
MARKETOS.
AGENTIC MARKETING PLATFORM
DEEP DUO FOUNDATION
ORG
PRODUCT DEV INTERN
ROLE
16 SUB-AGENTS
AGENTS
“

AI-native marketing operations platform architected around a Supervisor agent orchestrating 16 specialized sub-agents - one-click campaign intent to fully coordinated, 24/7 autonomous execution across channels.

MARKETOS
STACK
LANGGRAPHCLAUDE APIKAFKAPGVECTOR
MARKETOS.
MARKETOS.
AGENTIC MARKETING PLATFORM
DEEP DUO FOUNDATION
ORG
PRODUCT DEV INTERN
ROLE
16 SUB-AGENTS
AGENTS
“

AI-native marketing operations platform architected around a Supervisor agent orchestrating 16 specialized sub-agents - one-click campaign intent to fully coordinated, 24/7 autonomous execution across channels.

MARKETOS
STACK
LANGGRAPHCLAUDE APIKAFKAPGVECTOR
02
TRADESIM.
AGENTIC TRADING SIMULATOR

Agentic trading-simulation platform combining live OHLCV market data, RAG-based news retrieval, and technical-indicator computation - LLM-driven analysis produces structured, explainable stock reports with real-time P&L simulation.

IEMA R&D (IEMLABS)
ORG
TRADESIM.
TRADESIM.
AGENTIC TRADING SIMULATOR
IEMA R&D (IEMLABS)
ORG
ML INTERN
ROLE
EXPLAINABLE STOCK REPORTS
OUTPUT
“

Agentic trading-simulation platform combining live OHLCV market data, RAG-based news retrieval, and technical-indicator computation - LLM-driven analysis produces structured, explainable stock reports with real-time P&L simulation.

TRADESIM
STACK
PYTHONRAGLLM ANALYSIS
TRADESIM.
TRADESIM.
AGENTIC TRADING SIMULATOR
IEMA R&D (IEMLABS)
ORG
ML INTERN
ROLE
EXPLAINABLE STOCK REPORTS
OUTPUT
“

Agentic trading-simulation platform combining live OHLCV market data, RAG-based news retrieval, and technical-indicator computation - LLM-driven analysis produces structured, explainable stock reports with real-time P&L simulation.

TRADESIM
STACK
PYTHONRAGLLM ANALYSIS
03
BOL
BACHHAN.
AI VOICE CALL PLATFORM

AI voice call management platform (Retell AI–comparable) with dashboards for campaign management, call-log analytics, AI workflow building, lead management, and knowledge-base integration. Automated GoHighLevel CRM workflows via n8n.

VITANIUMX
ORG
BOL BACHHAN.
BOL BACHHAN.
AI VOICE CALL PLATFORM
VITANIUMX
ORG
40% SALES GROWTH
RESULT
+70% QUALIFIED INBOUND
LEADS
“

AI voice call management platform (Retell AI–comparable) with dashboards for campaign management, call-log analytics, AI workflow building, lead management, and knowledge-base integration. Automated GoHighLevel CRM workflows via n8n.

BOL BACHHAN
STACK
NEXT.JSTYPESCRIPTTAILWINDN8N
BOL BACHHAN.
BOL BACHHAN.
AI VOICE CALL PLATFORM
VITANIUMX
ORG
40% SALES GROWTH
RESULT
+70% QUALIFIED INBOUND
LEADS
“

AI voice call management platform (Retell AI–comparable) with dashboards for campaign management, call-log analytics, AI workflow building, lead management, and knowledge-base integration. Automated GoHighLevel CRM workflows via n8n.

BOL BACHHAN
STACK
NEXT.JSTYPESCRIPTTAILWINDN8N
04
GLYCOCART.
SWIGGY BUILDER CLUB SELECT

Glucose-aware ordering agent for Indians managing PCOS, prediabetes, and CGM-tracked metabolic health. Reads a user's glycemic profile, re-ranks Swiggy's live menu by predicted personal glycemic impact instead of generic GI tables, and places the order through Swiggy's MCP tool schema - a multi-step LLM tool-calling loop, not a static filter.

SWIGGY BUILDER CLUB
SELECTED
GLYCOCART.
GLYCOCART.
SWIGGY BUILDER CLUB SELECT
SWIGGY BUILDER CLUB
SELECTED
HEALTH-TECH AGENT
TYPE
PCOS · PREDIABETES · CGM
USERS
“

Glucose-aware ordering agent for Indians managing PCOS, prediabetes, and CGM-tracked metabolic health. Reads a user's glycemic profile, re-ranks Swiggy's live menu by predicted personal glycemic impact instead of generic GI tables, and places the order through Swiggy's MCP tool schema - a multi-step LLM tool-calling loop, not a static filter.

GLYCOCART
STACK
NEXT.JS 15VERCEL AI SDKCLAUDE 3.5 SONNET
GLYCOCART.
GLYCOCART.
SWIGGY BUILDER CLUB SELECT
SWIGGY BUILDER CLUB
SELECTED
HEALTH-TECH AGENT
TYPE
PCOS · PREDIABETES · CGM
USERS
“

Glucose-aware ordering agent for Indians managing PCOS, prediabetes, and CGM-tracked metabolic health. Reads a user's glycemic profile, re-ranks Swiggy's live menu by predicted personal glycemic impact instead of generic GI tables, and places the order through Swiggy's MCP tool schema - a multi-step LLM tool-calling loop, not a static filter.

GLYCOCART
STACK
NEXT.JS 15VERCEL AI SDKCLAUDE 3.5 SONNET
05
DEEPFAKE
DETECTION.
SIH 2024 · MULTIMODAL FORENSICS

Multimodal deepfake-detection pipeline built for Smart India Hackathon 2024. MTCNN and Azure Computer Vision extract and align faces from raw video, feeding a ResNeXt-50 + LSTM classifier for visual frames alongside Librosa audio-feature analysis, with Kafka-based streaming for real-time inference at scale.

SIH 2024
EVENT
DEEPFAKE DETECTION.
DEEPFAKE DETECTION.
SIH 2024 · MULTIMODAL FORENSICS
SIH 2024
EVENT
96.68%
ACCURACY
MTCNN + AZURE VISION FACE-EXTRACT
PIPELINE
“

Multimodal deepfake-detection pipeline built for Smart India Hackathon 2024. MTCNN and Azure Computer Vision extract and align faces from raw video, feeding a ResNeXt-50 + LSTM classifier for visual frames alongside Librosa audio-feature analysis, with Kafka-based streaming for real-time inference at scale.

DEEPFAKE DETECTION
STACK
RESNEXT-50LSTMLIBROSAKAFKAS3
DEEPFAKE DETECTION.
DEEPFAKE DETECTION.
SIH 2024 · MULTIMODAL FORENSICS
SIH 2024
EVENT
96.68%
ACCURACY
MTCNN + AZURE VISION FACE-EXTRACT
PIPELINE
“

Multimodal deepfake-detection pipeline built for Smart India Hackathon 2024. MTCNN and Azure Computer Vision extract and align faces from raw video, feeding a ResNeXt-50 + LSTM classifier for visual frames alongside Librosa audio-feature analysis, with Kafka-based streaming for real-time inference at scale.

DEEPFAKE DETECTION
STACK
RESNEXT-50LSTMLIBROSAKAFKAS3
06
HILEX
PUBLICATION.
ACCEPTED AT ICDAR 2025

Hierarchical layout extraction system using PyTorch and Roboflow object detectors, benchmarked on a dataset of 1,965 exam papers for automated digital question-bank creation. Co-authored with Utathya Aich, Shinjini Chakraborty, Souradeep Ghosh, and Tulika Saha.

ICDAR 2025
ACCEPTED
HILEX PUBLICATION.
HILEX PUBLICATION.
ACCEPTED AT ICDAR 2025
ICDAR 2025
ACCEPTED
CO-AUTHOR
ROLE
1,965 EXAM PAPERS
DATASET
“

Hierarchical layout extraction system using PyTorch and Roboflow object detectors, benchmarked on a dataset of 1,965 exam papers for automated digital question-bank creation. Co-authored with Utathya Aich, Shinjini Chakraborty, Souradeep Ghosh, and Tulika Saha.

HILEX PUBLICATION
STACK
PYTORCHROBOFLOW
HILEX PUBLICATION.
HILEX PUBLICATION.
ACCEPTED AT ICDAR 2025
ICDAR 2025
ACCEPTED
CO-AUTHOR
ROLE
1,965 EXAM PAPERS
DATASET
“

Hierarchical layout extraction system using PyTorch and Roboflow object detectors, benchmarked on a dataset of 1,965 exam papers for automated digital question-bank creation. Co-authored with Utathya Aich, Shinjini Chakraborty, Souradeep Ghosh, and Tulika Saha.

HILEX PUBLICATION
STACK
PYTORCHROBOFLOW
MORE ON GITHUB

OTHER PROJECTS

BLOODLINK

Full-stack blood bank platform with donor registration, live inventory dashboards, and geolocation-based nearby-bank search on interactive maps. A TensorFlow demand-forecasting model optimizes stock allocation, with automated donor–recipient matching and real-time inventory sync across banks.

REACT.JSEXPRESS.JSMONGODB ATLASREDISTENSORFLOWLEAFLET

ECO-ALERT

Cross-platform hazard-reporting app with geotagged, photo-backed incident reports and live air-quality overlays on an interactive map, backed by a Supabase real-time data layer. AI-assisted heatmap analysis and instant notifications accelerate response across water, air, noise, and chemical-spill incidents.

REACT.JSSUPABASEREST APISLEAFLET

CAMERA PLACEMENT VISUALIZER

3D pond camera-placement optimizer for CCTV coverage planning - a customer-facing Three.js scene with animated water and orbit controls, driven by a greedy placement algorithm that scores camera arrays on coverage, overlap, pixel density, and bird-detectability, then exports the deployment plan as JSON, CSV, and a Markdown report.

JAVASCRIPTTHREE.JS

SURVEYSPARROW

AI-assisted survey builder with a drag-and-drop question editor and a results dashboard. A prompt describing the survey's goal is turned into a full schema - titles, question types, options, rating scales, Likert matrices - by Gemini's structured-output API, ready to edit and publish.

TYPESCRIPTREACTGEMINI APIRECHARTS

RADISH

Vehicle-dynamics simulator built from a from-scratch physics engine - custom engine, transmission, brake, and chassis models - exposed as a Gymnasium environment for reinforcement-learning agents trained with Stable-Baselines3. Renders through a PyQt6 desktop app and a Panda3D 3D viewport with physically-based studio lighting.

PYTHONPYQT6PANDA3DGYMNASIUMSTABLE-BASELINES3PYTORCH

NUTRISENSE AI

Meal-intelligence assistant built for a Google hackathon on a Google ADK multi-agent system - a root agent delegates to a Gemini Vision scanner, a diet-coach agent, a meal-planner agent, and a Places-API agent for nearby healthy options, with Firestore-backed logs and a live progress dashboard.

GOOGLE ADKGEMINI VISIONFASTAPIFIREBASEREACT

BIRD DETECTOR

Real-time bird detection on a Raspberry Pi 4 with an IMX708 camera, running a YOLO26n model exported to NCNN for ARM inference and streaming annotated MJPEG video over Tailscale. Documented failure analysis on small-object detection and SAHI tiling tradeoffs drove iterative accuracy improvements on constrained hardware.

PYTHONYOLONCNNRASPBERRY PIFLASK

KODAK COLORSPACE VLM

Research probe into how color-space representation biases vision-language models. Each image in the Kodak benchmark is decomposed into 17 color-space variants - RGB, grayscale, CMYK and HSV/LAB channels - then run through Gemini Flash and Llama 4 (Scout, Maverick) to compare how scene understanding shifts across representations.

PYTHONGEMINILLAMA 4
RESEARCH

PUBLICATIONS

HiLEx: Image-Based Hierarchical Layout Extraction from Question Papers

Utathya Aich, Shinjini Chakraborty, Deepan Sadhukhan, Souradeep Ghosh, Tulika Saha

  • -Developed a hierarchical document layout extraction system using PyTorch and Roboflow object detectors, benchmarked on a multimodal dataset of 1,965 exam papers to automate digital question-bank creation.
  • -Published at ICDAR (International Conference on Document Analysis and Recognition) - a premier IEEE/IAPR venue for multimodal AI, document understanding, and vision-language research.
ACCEPTED · ICDAR 2025CO-AUTHOR
TECHNICAL PROFICIENCY

SKILL MATRIX

GENERATIVE AI & LLMS
LANGGRAPHAGENTIC AIRAG SYSTEMSPROMPT ENGINEERINGCLAUDE APIOPENAI APIHUGGING FACE TRANSFORMERS
7 SKILLS
MACHINE LEARNING
PYTORCHTENSORFLOWSCIKIT-LEARNOPENCVCNNSLSTMSOBJECT DETECTION
7 SKILLS
DATA & VECTOR INFRA
POSTGRESQLPGVECTORMONGODBREDISCLICKHOUSEKAFKAAWS S3
7 SKILLS
FRONTEND
NEXT.JSREACT.JSTYPESCRIPTTAILWIND CSSSTREAMLITGRADIO
6 SKILLS
BACKEND
PYTHONFLASKEXPRESS.JSREST APISCLOUDFLARE WORKERS
5 SKILLS
EDGE AI & DEPLOYMENT
DOCKERONNXRASPBERRY PIRTSPMQTTFFMPEG
6 SKILLS
TOOLS & AUTOMATION
GIT/GITHUBN8NPOSTMANROBOFLOWTHREE.JSLATEX
6 SKILLS
PROFESSIONAL JOURNEY

EXPERIENCE

DD
DEEP DUO FOUNDATION
MAR 2026 - JUN 20263 MO
PRODUCT DEVELOPMENT INTERN
  • Architected MarketOS: a Supervisor agent orchestrating 16 specialized sub-agents
  • Built event-driven multi-agent system with LangGraph, Claude API, Kafka, pgvector
  • Defined inter-agent schemas, observability pipelines, and non-technical dashboards
IR
IEMA R&D
DEC 2025 - FEB 20262 MO
MACHINE LEARNING INTERN
  • Built TradeSim, an agentic trading-simulation platform with RAG news retrieval
  • Delivered LLM-driven stock analysis with real-time P&L simulation
  • Designed structured, schema-validated LLM output pipelines
V
VITANIUMX
MAR 2025 - DEC 20259 MO
FULL STACK SOFTWARE DEVELOPER INTERN
  • Built Bol Bachhan, an AI voice call management platform, with Next.js and TypeScript
  • Automated GoHighLevel CRM workflows via n8n and rebuilt landing pages
  • Drove 40% sales growth, ~70% more qualified inbound calls, 30% better lead conversion
NI
NEXAWINGS INTERNATIONAL
JAN 2024 - APR 20243 MO
SOFTWARE DEVELOPER INTERN
  • Developed Flask microservices with CRUD endpoints for core product workflows
  • Identified 9 critical backend issues via systematic Postman API testing
  • Scraped and validated ~1,500 high-intent sales leads with BeautifulSoup
II
IEM IEDC
SEP 2023 - NOV 20232 MO
STUDENT RESEARCHER
  • Built a disease-diagnosis RAG chatbot achieving 85% diagnostic accuracy
  • Validated the chatbot against 500+ test users
  • Automated a portal managing 30+ student-project mappings
BACKGROUND

EDUCATION & ACHIEVEMENTS

EDUCATION

BTech in Computer Science, Specialization in AI & Machine Learning

Institute of Engineering and Management

Salt Lake, Kolkata, West Bengal, India

2023 - 2027
ACHIEVEMENTS & ACTIVITIES
  • -Led interdisciplinary teams in national-level hackathons and innovation programs: Startup Mahakumbh, MSME IP-Yatra, IDE Bootcamps, Google Solution Challenge - building experience in rapid AI prototyping under competitive conditions.
  • -Proactively explores emerging AI tools, LLM frameworks, and automation platforms, designing end-to-end workflow automations that eliminate repetitive processes.
  • -Continuously tracks applied AI research and arXiv preprints to stay current with fast-moving developments in LLMs, multi-agent systems, and AI infrastructure.