Senior Associate Technology Engineer @ ClarityX

Building Intelligent Systems That Scale

I build multi-agent LLM systems

AI, Geospatial Platforms & GenAI. 5+ years in R&D building production ML, LLM and geospatial systems, from a multi-tenant site-selection platform on AWS to multi-agent LLM analytics over 1M+ records/day.

View Projects Get In Touch
susheel@portfolio: ~
$
97%+ CV Model Accuracy
1M+ Daily Records Processed
<150ms Real-time Latency
55k LOC Platform Backend
PythonFastAPIPyTorchYOLO v8ArcFaceDetectron2Mistral 7BLLaMA 3 70BAWS BedrockLangChainOllamaPostGISRedisQGISPydeckDockerAWSGitLab CI/CDnginxpytestPythonFastAPIPyTorchYOLO v8ArcFaceDetectron2Mistral 7BLLaMA 3 70BAWS BedrockLangChainOllamaPostGISRedisQGISPydeckDockerAWSGitLab CI/CDnginxpytest
01

About Me

I'm Susheel Kumar Pandey, an AI Engineer and Senior Associate Technology Engineer at ClarityX, building production ML, LLM and geospatial systems.

At ClarityX I architected ATLAS, a multi-tenant geospatial site-selection platform on AWS (55k LOC backend, six scoring engines, CI-enforced hexagonal architecture), and a production multi-agent LLM analytics platform over 1M+ daily mobility records.

Before that, at MapmyIndia, I built real-time computer vision for fleets (driver monitoring at 97%+ accuracy under 150 ms), ArcFace identity verification, a multilingual LLM semantic search system, and geospatial AI tooling for satellite imagery and 3D city models.

📍
Location New Delhi, India
🎓
Education B.Tech in CSE, GLA University (2021)
💼
Current Role Sr. Associate Technology Engineer @ ClarityX
02

Technical Expertise

Computer Vision

Real-time detection and face verification, tuned and quantized for edge devices.

YOLO v5/v7/v8 ArcFace Detectron2 Model Quantization Hyperparameter Tuning
97%+ Detection Accuracy
See projects →

LLM & GenAI

Multi-agent systems for intent classification, SQL generation and semantic search.

Mistral 7B LLaMA 3 70B AWS Bedrock LangChain Ollama
35% Search Accuracy Boost
See projects →

Geospatial Intelligence

Scoring engines, spatial analytics and satellite-imagery tooling at city scale.

PostGIS Geospatial Scoring Pydeck QGIS Dimensional Modeling
6 Scoring Engines
See projects →

Backend, Cloud & MLOps

Production APIs, multi-tenant AWS deployments and CI-enforced architecture.

FastAPI Async SQLAlchemy Hexagonal Architecture AWS GitLab CI/CD
10× p90 Latency Cut
See projects →
03

Featured Projects

lowhighIllustrative grid hotspot scoring · hover a cell
Mappls mGIS Data
→
6 Scoring Engines
→
UniversalSchema
→
Multi-tenant API
ClarityX Geospatial Platform

ATLAS: Geospatial Market-Intelligence Platform

Architected and built the full backend of a multi-tenant retail site-selection platform: city prioritisation, micro-hotspot, pincode and catchment analysis, in production on AWS.

55k LOC Python/FastAPI
27 Async API Routers
20k LOC Automated Tests
  • Six scoring engines (grid hotspot, city, pincode, catchment, rural, hybrid) plus a BFSI vertical, driven by versioned YAML weights
  • 18 swappable plugins, hexagonal architecture enforced in CI
  • AWS EC2, RDS, S3, Cognito, Verified Permissions multi-tenant RBAC, Secrets Manager, nginx, 4-stage GitLab CI
  • Key p90 latencies cut by an order of magnitude, resolving a 240 s upstream stall and a 74.7 s p90 tile render

04

Experience

Apr 2026 — Present

Senior Associate Technology Engineer

ClarityX • New Delhi

Building geospatial market-intelligence and LLM analytics platforms in production.

AWSFastAPIPostGISMistral 7BLLaMA 3 70BGitLab CI
Show 8 highlightsHide highlights

ATLAS — Geospatial Market-Intelligence Platform

  • Architected and built the full backend of ATLAS, a multi-tenant retail site-selection platform (city prioritisation, micro-hotspot, pincode and catchment analysis): 55k LOC Python/FastAPI, 27 async API routers, 18 swappable plugins, hexagonal architecture enforced in CI
  • Designed six geospatial scoring engines (grid hotspot, city, pincode, catchment, rural, hybrid) plus a BFSI vertical on Mappls mGIS data, driven by versioned YAML weights and exposed through UniversalSchema, a single JSON contract for all outputs
  • Took the platform to production on AWS (EC2, RDS, S3, Cognito, Verified Permissions multi-tenant RBAC, Secrets Manager) with nginx and a 4-stage GitLab CI pipeline backed by 20k LOC of automated tests
  • Cut key p90 latencies by an order of magnitude via percentile-driven profiling and single-flight caching, resolving a 240-second upstream stall and a 74.7-second p90 tile render

Footfall Analysis LLM Agent Platform

  • Architected and deployed a production multi-agent system: Mistral 7B Intent Agent, LLaMA 3 70B SQL Generator, plus Executor and Summarizer agents, converting natural-language queries into business intelligence for location analytics
  • Built a 3-layer dimensional model (Search, Directions, Navigation) over 1M+ daily mobility telemetry records, with a weighted intent formula (searches × 1.0 + clicks × 2.0 + directions × 2.5) validated against 60–80% actual visit probability
  • Implemented schema-aware SQL generation with automated validation, resolving table-alias conflicts and formula accuracy through a 100+ test-case framework, reducing manual testing by 75%
  • Integrated PostGIS spatial queries with Pydeck heatmaps and point maps; deployed on Streamlit at 15–30 s latency for ~50 concurrent non-technical users
Apr 2024 — Mar 2026

Senior Software Engineer

MapmyIndia • New Delhi

LLM-powered place discovery on MapmyIndia APIs.

LLMsIntent ClassificationSemantic Search
Show 2 highlightsHide highlights

LLM Semantic Search System

  • Built a multilingual intent-based search system using LLMs integrated with MapmyIndia APIs, improving place discovery accuracy by 35%
  • Enabled natural-language queries for place discovery, with semantic understanding of user intent and contextual matching
Mar 2022 — Mar 2024

Software Engineer

MapmyIndia • New Delhi

Production computer vision, CV microservices and geospatial AI.

YOLO v7ArcFaceDetectron2FastAPIDockerQGIS
Show 15 highlightsHide highlights

Driver Monitoring System

  • Developed real-time detection of driver anomalies such as drowsiness and distraction using YOLO v7, achieving 97%+ accuracy with response times under 150 ms
  • Led R&D experiments across YOLO variants with systematic hyperparameter tuning (learning rate, batch and image size, augmentation, confidence and IoU thresholds), balancing accuracy against the real-time latency budget
  • Curated and annotated in-cabin driver datasets covering varied lighting, head poses and occlusions; evaluated with precision, recall and mAP to minimise false alerts
  • Quantized and optimized for on-device edge processing, reducing cloud dependency and improving system reliability

Face Recognition and Verification

  • Implemented an ArcFace-based driver identity verification module with 89%+ accuracy, reliable in low-light in-vehicle conditions
  • Tuned ArcFace hyperparameters (margin, scale, learning rate) and calibrated the similarity threshold to balance false accepts against false rejects
  • Deployed in production fleet management systems, processing 1000+ daily verifications

AI Microservices Platform

  • Built a scalable microservice platform integrating YOLO, ArcFace and Detectron2 models with FastAPI and Docker
  • Handled 200+ concurrent requests with 94% uptime, serving computer vision workloads across multiple client applications

Geospatial AI Tools

  • Created a QGIS plugin with models for land-use classification, change detection and object recognition from satellite imagery
  • Improved geospatial data accuracy by 25%, accelerating map data enrichment workflows

3D Building Automation

  • Automated city-scale 3D building generation from aerial imagery and DEM data, reducing manual modeling effort by 65%
  • Achieved 40% model size reduction through optimization techniques while maintaining visual quality

MLOps and Deployment

  • Improved ML pipelines with asynchronous APIs and batching, cutting processing times by 50%
  • Automated CI/CD workflows, reducing release cycles by 40% and improving model deployment velocity
Sep 2021 — Feb 2022

Trainee Engineer

MapmyIndia • New Delhi

Geospatial imagery analysis and object detection.

YOLO v8QGISData Augmentation
Show 3 highlightsHide highlights
  • Assisted in developing a QGIS-based geospatial analysis plugin for AI-driven satellite and street imagery analysis
  • Built a YOLO v8-based tank detection model achieving 94% precision and 92% recall, optimized for challenging defense scenarios
  • Implemented data augmentation strategies to handle varied terrain and lighting conditions
05

Let's Connect

Interested in AI, GenAI, Computer Vision, or Geospatial projects? I'm always open to discussing new opportunities and collaborations.

Email susheel123pandey@gmail.com
LinkedIn linkedin.com/in/susheel7860
GitHub github.com/susheel7860
Phone +91-7860377399