Rohit Gusain
LLM & Agents
7+ years exp
7+ years ML exp
Management: 1-2 years
Python
SQL
TypeScript
Node JS
MLOps & Deployment: 5/10
Active 4 days ago
Invite to interview
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
LLM-focused engineer (senior level) specializing in building RAG-enabled conversational assistants and production-facing inference services. The strongest proven capability is engineering an end-to-end RAG/chat service with memory, streaming and tool execution as implemented in SHADON-AI/backend/core (llm_engine_v2.py, rag_engine_v2.py, memory_manager.py and views_v2.py). Public artifacts do not show custom pretraining workflows, systematic experiment tracking (W&B/MLflow), formal performance benchmarking or unit/integration tests for the production components.
Phone
Technical skills
Languages
4
Python
SQL
TypeScript
Node JS
AI/ML
9
spaCy
OpenCV
LangChain
NumPy
Pandas
Scikit-learn
TF-Keras
Keras
OpenAI
Analytics
5
Looker
Power BI
Matplotlib
Seaborn
Plotly
Frontend
3
React.js
Next.js
Zustand
Other
41
Django REST Framework
Redis
FAISS
Rest API
Express
AWS
Claude
LLM
Django
FastAPI
Amazon SageMaker
AWS Bedrock
Gemini
LangGraph
Llama
XGBoost
YOLO
Jupyter Notebook
PyTorch
TensorFlow
Amazon S3
GPT-4
Qwen
Amazon EC2
GitHub
Deep Learning
OCR
Tableau
Docker
Git
Postman
Model Context Protocol
AI Agents
Machine Learning
RAG
Prompt Engineering
Time Series Forecasting
Computer Vision
Multimodal AI
NLP
Function Calling
Timeline
Programme Manager – AI & Digital Transformation
•
Lead
VLCC Healthcare Limited
•
Full-Time
Architected and deployed an enterprise video analytics platform for real-time footfall and workforce monitoring across clinics and retail locations. Built BI reporting by integrating Power BI dashboards with SAP FI/MM data sources to streamline executive KPI reporting. Led the development of an AI-driven multi-platform SaaS workforce management and sales execution system with authentication, attendance, transcription, and multilingual speech processing capabilities.
OpenCV
Python
Power BI
Data Analyst
•
Middle
SISL Infotech Private Limited
•
Full-Time
Designed an automated pipeline to convert PDFs into structured data using OCR and document processing. Built REST APIs to evaluate government schemes by combining time-series forecasting with ranking and analytics components for decision support. Improved data extraction and validation quality through feature engineering and automated checks, increasing structured data accuracy.
Python
OCR
Rest API
Time Series Forecasting
Senior Analyst
•
Senior
GlobalLogic Technologies
•
Full-Time
Analyzed large-scale structured and unstructured datasets to support Google Knowledge Panel and Search Generative Experience initiatives, focusing on search data quality. Developed KPI dashboards and analytics frameworks to monitor trends and identify actionable improvements for product and engineering teams. Automated recurring reporting workflows using SQL-driven validation and process optimization to reduce manual effort and improve consistency.
SQL
Looker
National Institute of Engineering and Technology (NIET)
Master's Degree •
Artificial Intelligence
Software Engineer – AI / NLP
•
Middle
Prakhar Software Solutions Private Limited
•
Full-Time
Designed and deployed an enterprise HR chatbot that supports candidate screening and interview scheduling using NLP components and SQL-backed data workflows. Improved conversational performance by tuning intent classification, entity recognition, and dialogue flows with iterative dataset and model refinements. Delivered continuous evaluation steps to increase chatbot accuracy and reliability for recruitment operations.
Python
SQL
spaCy
Jamia Hamdard
Bachelor's Degree •
Electronics & Communication Engineering
Senior AI/ML Engineer
Confidence: Medium LLM Engineer
LLM-focused engineer (senior level) specializing in building RAG-enabled conversational assistants and production-facing inference services. The strongest proven capability is engineering an end-to-end RAG/chat service with memory, streaming and tool execution as implemented in SHADON-AI/backend/core (llm_engine_v2.py, rag_engine_v2.py, memory_manager.py and views_v2.py). Public artifacts do not show custom pretraining workflows, systematic experiment tracking (W&B/MLflow), formal performance benchmarking or unit/integration tests for the production components.
Model Architecture & Training
3/10
How well models are designed and trained
Good system-level LLM orchestration and prompt/context engineering are present, but there is no evidence of custom pretraining or advanced training pipelines.
Evidence
SHADON-AI/backend/core/ai_services/llm_engine_v2.py: context builder, tool detection and generate/generate_stream orchestration
SHADON-AI/backend/core/ai_services/rag_engine_v2.py: hybrid scoring and retrieval logic using embeddings/FAISS
Data Pipeline & Feature Engineering
3/10
How data is prepared for models
Standard data preparation and feature normalization are used in many notebooks, suited to prototyping and classical ML experiments.
Evidence
Machine-Learning-Projects/bankrupcy_prevention/Bankruptcy.ipynb: CSV/Excel parsing, label encoding and histograms
Machine-Learning-Projects/insaid/Sample_Skeleton.ipynb: preprocessing, normalization and feature pipelines
Experimentation & Evaluation
2/10
How results are measured and tested
Notebook-level evaluation (confusion matrices, classification reports) and basic model comparisons exist, but there is no sign of reproducible experiment tracking, ablations or CI-driven evaluation.
Evidence
Machine-Learning-Projects/musk/musk.ipynb: model training, evaluation metrics and classification_report
Machine-Learning-Projects/bankrupcy_prevention/Bankruptcy.ipynb: confusion_matrix and accuracy reporting
MLOps & Deployment
5/10
How models are shipped to production
Clear engineering for deployment and serving: production Django endpoints, SSE streaming, document ingestion, migrations and caching support indicate practical MLOps and deployment experience.
Evidence
SHADON-AI/backend/core/views_v2.py: Django endpoints, SSE streaming, file upload and memory management endpoints
SHADON-AI/backend/core/migrations/0001_initial.py: production-grade models with indexes and unique constraints
Computational Efficiency
3/10
How efficiently computing resources are used
Attention to inference efficiency is visible (optional FAISS index, sentence-transformers embeddings, threadpooling), but there is no measured profiling, quantization or distributed optimization shown.
Evidence
SHADON-AI/backend/core/ai_services/rag_engine_v2.py: optional FAISS index and embedding handling
SHADON-AI/backend/core/ai_services/memory_manager.py: conditional use of sentence-transformers for embeddings
Research Depth & Innovation
3/10
Depth of research and new ideas
Some thoughtful engineering patterns (context compression, hybrid retrieval, tool execution) show applied research thinking, but there are no novel algorithms, formal ablations or paper-reproduction artifacts.
Evidence
SHADON-AI/backend/core/ai_services/rag_engine_v2.py: ContextCompressor and hybrid semantic+keyword scoring
SHADON-AI/backend/core/ai_services/llm_engine_v2.py: tool detection and execution flow
Expertise
RAG• Senior
LLM• Senior
AI / LLM Engineering (Agents)• Senior
Conversational AI & Chatbots• Senior
Industries
Artificial Intelligence• Senior
Software• Middle
Technologies
Deep Learning
SQL• since 2019 • Senior
LangGraph
Claude
Qwen
OpenCV• since 2025
Model Context Protocol
YOLO
XGBoost
Prompt Engineering
Multimodal AI
Function Calling• since 2026
Computer Vision
AI Agents
NLP
spaCy• since 2019
Llama
AWS Bedrock
TensorFlow• since 2026
Git
PyTorch• since 2026
AWS
Docker
Gemini
LLM
RAG• since 2026
Amazon EC2
Time Series Forecasting• since 2025
Amazon SageMaker
GitHub
Amazon S3
OCR• since 2025
GPT-4
Machine Learning
Recommendations
- Develop production RAG-based conversational assistants and voice agents with the existing Django-based service and memory/RAG stack.
- Implement and harden an inference-serving pipeline (SSE/streaming, caching, rate limiting) and add observability/metrics (Prometheus/Grafana) to the SHADON service.
- Convert notebook prototypes into reproducible training/evaluation pipelines with experiment tracking (W&B or MLflow) and CI tests.
- Add measured efficiency work (profiling, quantization, FAISS tuning, batching) and end-to-end integration tests for memory and ingestion flows.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Senior Data Scientist
Confidence: Medium ML Practitioner
An applied ML practitioner focused on model prototyping and conversational/LLM systems at a solid mid-senior level with production-oriented service code. The strongest proven skill is designing a retrieval-augmented generation and memory-backed conversational architecture as implemented in the Django-based Shadon backend (core/ai_services, memory_manager, rag_engine_v2, llm_engine_v2). Public code lacks rigorous statistical testing, production-grade security hardening and full reproducibility tooling such as automated tests, seeds, or data versioning.
Statistical Rigor
2/10
Correct use of statistics
Statistical analysis is lightweight across notebooks - descriptive statistics and correlations are present, but formal assumption checks, uncertainty quantification, multiple-comparison controls or causal analysis are missing.
Evidence
i-am-rohit/Machine-Learning-Projects/bankrupcy_prevention/Bankruptcy.ipynb: descriptive summaries and value counts without formal tests or uncertainty reporting
i-am-rohit/Machine-Learning-Projects/corona.ipynb: exploratory descriptions and plots but no statistical hypothesis checks or uncertainty quantification
Data Wrangling & Cleaning
4/10
Preparing and cleaning data
Data wrangling is pragmatic - renaming, label encoding, normalization and null checks are used, but there are dangerous shortcuts (LabelEncoder on numeric-like charges, inline file path assumptions) and limited provenance or pipeline safeguards against leakage.
Exploratory Analysis & Visualization
4/10
Exploring and visualizing data
Exploratory plots and visualizations are frequent and useful for quick insight, but most plots lack written interpretation linked to hypotheses and do not progress to formal exploratory findings or actionable summaries.
Predictive Modeling
4/10
Building models that predict
Predictive modeling shows practical competence - multiple classical models and some neural nets are trained and evaluated, but there is limited use of robust CV schemes, time-aware splits where required, calibration checks or comprehensive error analysis.
Evidence
i-am-rohit/Machine-Learning-Projects/bankrupcy_prevention/Bankruptcy.ipynb: training multiple sklearn classifiers and reporting train/test scores and confusion matrices
i-am-rohit/Machine-Learning-Projects/diabetes using tensor flow/diabetes.ipynb: linear and DNN classifiers with TensorFlow estimators and evaluation
Business Insight & Impact
2/10
Turning analysis into business value
Business context and impact analysis are minimal - projects run standard ML workflows but rarely quantify business metrics, costs of false positives/negatives or provide actionable recommendations tied to stakeholders.
Evidence
i-am-rohit/Machine-Learning-Projects/insaid/Sample_Skeleton.ipynb: churn modelling exercise without cost-of-error or business-metric interpretation
i-am-rohit/Machine-Learning-Projects/bankrupcy_prevention/Bankruptcy.ipynb: model evaluation limited to accuracy and confusion matrices without business impact discussion
Reproducibility & Notebook Hygiene
3/10
Clean, repeatable analysis
Reproducibility is partial - there are pinned requirements files for some services and notebooks include pip installs and data paths, but there is limited environment knitting, seed management, modularization, CI/tests or dataset versioning to support robust reproducibility.
Expertise
Data Science• Middle
Streaming• Middle
Industries
Artificial Intelligence• Senior
Technologies
Python• since 2019 • Senior
Redis
LangChain
FAISS
Jupyter Notebook
Scikit-learn• since 2026
Seaborn
Matplotlib
Plotly
Pandas
NumPy
Keras• since 2026
TF-Keras
Django REST Framework
OpenAI
Recommendations
- Develop production RAG-based conversational assistants and LLM orchestration services with Django/REST and LangChain/FAISS integration.
- Prototype and productionize supervised ML models and pipelines where rapid EDA, feature engineering and classical model ensembles are required.
- Implement structured evaluation and MLOps workflows - CI, deterministic seeds, CV pipelines and dataset versioning to raise model robustness.
- Build streaming endpoints for real-time responses and evented LLM streaming with attention to SSE/WS security and backpressure handling.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Frontend Developer
Confidence: Medium App Engineer
App-focused frontend engineer (mid-level) working with TypeScript Next.js clients and a Node/Express backend - strongest at implementing product-facing application flow and server-client data layers. The clearest proven skill is architecting typed client-server data flows and domain services as shown by backend/src/services/beat-session-service.ts and client store/api layers (web-app/src/store/index.ts and web-app/src/lib/api-client.ts). What is not evidenced in public files is a custom component design system, systematic a11y CI or measured performance reports and broad test coverage for the UI.
UI Component Architecture
3/10
How interface parts are built
Component boundaries and architecture exist at application level (store, API client, domain libs) but there is little evidence of a bespoke component library, design-system tokens or advanced composition patterns in the public files.
Evidence
geonexis_new/web-app/src/store/index.ts: store interfaces and scoped getters
geonexis_new/web-app/src/lib/api-client.ts: centralized apiRequest and auth helpers
geonexis_new/web-app/src/services/api.ts: mock AuthService/DataService shape
Responsive & Cross-browser
3/10
Works on all screens and browsers
Responsive layout and cross-browser tooling are present via standard frameworks (Next.js + Tailwind/Tailwind tooling in deps) and meta viewport usage, but no advanced responsive techniques (container queries, fluid type) are documented.
Evidence
geonexis_new/web-app/package.json: includes next, tailwindcss and related front-end tooling
i-am-rohit/iamrohit/index.html: meta viewport and Bootstrap responsive layout usage
Performance Optimization
2/10
Speed of the interface
Some attention to performance is visible through selection of modern libraries (zustand, React) and separation of backend services, but there is no measured performance data or explicit bundle/code-splitting strategy documented in the public files.
Evidence
geonexis_new/web-app/package.json: zustand present for lightweight state
geonexis_new/backend/package.json: service-oriented backend (suggests separation of concerns)
Accessibility & Semantics
2/10
Usable for everyone
Basic accessibility and semantics appear to be considered via use of Radix and standard ARIA-friendly primitives in deps, and presence of alt attributes and sr-only utilities in HTML/CSS, but there is no sign of automated a11y checks or deliberate focus-management in custom widgets.
Evidence
geonexis_new/web-app/package.json: includes @radix-ui dependencies which are a11y-oriented primitives
i-am-rohit/iamrohit/assets/css/bootstrap.css: .sr-only helper and many semantic helper classes
i-am-rohit/iamrohit/index.html: img elements have alt attributes
State Management & Data Flow
4/10
Managing data in the app
Strong server-client state and data-flow discipline is evidenced by explicit store, api-client, typed models and backend services for domain flows; there is indication of thought put into scoped data, services and domain types though advanced patterns like request cancellation, optimistic rollback or state machines are not fully documented.
Evidence
geonexis_new/web-app/src/store/index.ts: AuthState/DataState/UIState and scoped getters
geonexis_new/web-app/src/lib/api-client.ts: apiRequest, loginWithApi, fetchAuthMe signatures
geonexis_new/backend/src/services/beat-session-service.ts: domain service with queueTranscription and session lifecycle
UX & Visual Polish
3/10
Look and feel quality
UX polish and domain UX flows are present with mock-data, role-aware flows and many UI libraries included, but many UI pages appear scaffolded and there is limited evidence of measured perceived-performance techniques or extensive user-research driven refinements.
Evidence
geonexis_new/web-app/src/mock-data/generator.ts: mock database generator for UX
geonexis_new/web-app/src/services/api.ts: createMockDataService and role-aware homePathForRole
i-am-rohit/iamrohit/index.html: theme switcher and interactive UI snippets
Expertise
React• Middle
Modern Web Frameworks• Middle
Industries
Sales & Marketing• Middle
Technologies
TypeScript• Middle
Node JS• Middle
Zustand
Next.js
Express
React.js
Recommendations
- Implement a small demo that isolates custom component design system tokens and documents composition patterns (Storybook) so UI architecture evidence is explicit.
- Add end-to-end and integration tests plus a few performance measurements (LCP/INP) to show real-world performance tuning and regression control.
- Document key async state decisions - request cancellation, optimistic updates or state-machine outlines for non-trivial flows (e.g. beat/session lifecycle).
- Introduce automated a11y checks (axe/lighthouse in CI) and keyboard/focus tests for any custom widgets.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: