Sumit Aherwadkar
AI Engineer
C++
SQL
C
Python
JavaScript
Data Pipeline & Feature Engineering: 4/10
Experimentation & Evaluation: 4/10
Active 2 days ago
+91 (7887) 942088 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
LLM engineering practitioner (senior-level) focused on building multi-agent RAG pipelines and enterprise LLM integrations. The strongest proven skill is designing and implementing hybrid RAG + SQL workflows with RBAC and evaluation tooling, evidenced by the FastAPI-based API, NL->SQL pipeline (csv_query.py) and the RAG indexer/fallback (rag_module.py). There is limited or no evidence of custom model training, GPU/quantization optimizations or centralized experiment tracking in public code.
Technical skills
C++
SQL
C
Python• Senior
JavaScript• Middle
C++
TensorFlow C++
Python
FastAPI
Beautiful Soup
Requests
Databases
MySQL
SQLite
DuckDB
AI/ML
CNN
Groq
Hallucination
LangChain
LangGraph
LLM
RAG
Streamlit
Transfer Learning
Keras
NumPy
Pandas
Scikit-learn
TensorFlow
Frontend
Chart.js
Vite
Bootstrap
DevOps
Git
Rest API
Timeline
Built a LangGraph-based multi-agent AI research system using Groq LLM and Tavily Search to automate information retrieval, analysis, and report generation. Implemented structured output validation and error handling to improve the reliability of AI-generated reports. Developed a Streamlit interface for monitoring the research workflow and presenting generated results.
Built an RBAC-enabled enterprise AI knowledge assistant using FastAPI, LangChain, ChromaDB, BM25, Groq LLM, and Streamlit. Developed secure document and SQL-based knowledge retrieval with role-based access control. Implemented an AI evaluation dashboard to measure faithfulness, relevance, hallucination rate, confidence, retrieval performance, and response latency.
Research Intern – Deep Learning & Medical AI
•
Junior
Pune Institute of Computer Technology
•
Internship
Worked on applying deep learning and machine learning methods to medical AI tasks using Python with TensorFlow/Keras. Built model training and experimentation workflows including evaluation and performance comparison. Performed data preprocessing and debugging while iterating on model optimization to improve prediction quality.
Python
TensorFlow
Keras
Intern – AI & Cloud Technologies
•
Junior
Remote Edunet Foundation
•
Internship
Developed Python-based applications as part of hands-on training focused on machine learning and cloud-related solutions. Applied structured problem-solving to implement and debug features for practical use cases. Completed tasks within an internship learning program in collaboration with AICTE and IBM SkillsBuild.
Pythonsince 2025
Pune Institute of Computer Technology
Bachelor's Degree •
Electronics and Telecommunication
Senior AI/ML Engineer
Confidence: Medium LLM Engineer
LLM engineering practitioner (senior-level) focused on building multi-agent RAG pipelines and enterprise LLM integrations. The strongest proven skill is designing and implementing hybrid RAG + SQL workflows with RBAC and evaluation tooling, evidenced by the FastAPI-based API, NL->SQL pipeline (csv_query.py) and the RAG indexer/fallback (rag_module.py). There is limited or no evidence of custom model training, GPU/quantization optimizations or centralized experiment tracking in public code.
Model Architecture & Training
2/10
How well models are designed and trained
Mostly integration of hosted LLMs and embedding models; no custom architectures or training loops are present.
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Solid data handling and ingestion pipelines for CSVs and documents, DuckDB integration and CSV-to-document conversion, but no advanced feature engineering at production scale.
Evidence
Sumit431511/FInSight-AI/app/rag_utils/csv_query.py: translate_nl_to_sql(), _create_duckdb_connection(), ask_csv() - NL->SQL and DuckDB usage
Sumit431511/FInSight-AI/app/rag_utils/rag_module.py: load_file() converts CSV/MD to documents for embedding
Sumit431511/Creditwise-Loan-Approval/project.ipynb: data cleaning, imputation and feature engineering cells
Experimentation & Evaluation
4/10
How results are measured and tested
There is an evaluation pipeline and metrics instrumentation for RAG responses, plus tests; experiment tracking is lightweight and not centralized (no W&B/MLflow evidence).
Evidence
Sumit431511/FInSight-AI/app/rag_evaluator/metrics.py: build_metrics() and related metric functions
Sumit431511/FInSight-AI/app/rag_evaluator/online_evaluator.py and offline_evaluator.py: LLM-based evaluation calls
Sumit431511/FInSight-AI/app/rag_evaluator/dashboard.py: get_dashboard_data() exposed to UI
MLOps & Deployment
3/10
How models are shipped to production
Has a deployable FastAPI service, persistence, indexer and basic fallback behaviors; limited evidence of production-grade deployment, CI/CD or drift monitoring.
Computational Efficiency
3/10
How efficiently computing resources are used
Some attention to reliability and efficiency: parallel scraping, retries with backoff and a DuckDB-backed SQL engine; no GPU/quantization or low-level optimization work.
Evidence
Sumit431511/ResearchMind-AI-Research-Agent/orchestrator.py: collect_scraped_sources() uses ThreadPoolExecutor and as_completed for parallel scraping
Sumit431511/ResearchMind-AI-Research-Agent/tools.py: @retry decorators using tenacity for robust web calls
Sumit431511/FInSight-AI/app/rag_utils/csv_query.py: DuckDB in-memory fallback handling
Research Depth & Innovation
2/10
Depth of research and new ideas
Thoughtful multi-agent pipeline and critic/revision loop design but no original research, novel algorithms or paper-level reproductions.
Expertise
AI / LLM Engineering (Agents)• Junior
RAG• Middle
Industries
Financial Services• Middle
Technologies
Python• Senior
SQL
C++
MySQL
LangGraph
Rest API
LangChain
DuckDB
Groq
FastAPI
Scikit-learn
Beautiful Soup
Transfer Learning
TensorFlow
Pandas
NumPy
Keras
Git
SQLite
LLM
RAG
TensorFlow C++
Streamlit
Requests
Hallucination
CNN
Groq• mentioned only
LangGraph• mentioned only
Recommendations
- Develop enterprise RAG systems and LLM-based agent pipelines (search, reader, writer, critic) with role-based access controls.
- Implement robust ingestion and NL->SQL data-access layers for internal document + structured-data question answering.
- Build evaluation and monitoring tooling for LLM outputs (online/offline evaluators, dashboards, logging) and extend to drift detection and A/B testing.
- Prototype production deployments of LLM services (containerization, CI/CD, observability and cost/latency budgets).
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Frontend Developer
Confidence: Medium UI Engineer
Frontend UI engineer (middle level) focused on polished, hand-crafted browser interfaces and dashboards. The strongest proven skill is designing and implementing interactive clinical and telemetry UIs with concrete artifacts like the NeuraScan doctor dashboard (dashboard.html, index.html) and the canvas-based robotics map (Insight.IO map.js). Public code lacks automated tests, formal componentization (framework components), and advanced server-state machinery such as request cancellation or documented cache invalidation.
UI Component Architecture
4/10
How interface parts are built
Modular UI logic implemented with clear function boundaries in plain JS and reusable CSS tokens, but no component framework or formal design-system components.
Evidence
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/dashboard.html: renderCharts / renderTable / paintTable functions
Sumit431511/Insight.IO-Dashboard/js/map.js: drawMapOnCtx / getPts (canvas rendering split into drawing functions)
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/index.html: showRes (UI mapping for prediction response)
Responsive & Cross-browser
4/10
Works on all screens and browsers
Responsive layouts use media queries and flexible grids; touch and pointer input handling present, but no explicit RTL or comprehensive i18n support.
Evidence
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/index.html: multiple @media rules and responsive grid (scan-grid, stats-grid)
Sumit431511/Insight.IO-Dashboard/js/map.js: touch event handlers (touchstart, touchmove, touchend) for pan
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/dashboard.html: responsive charts and container queries-like sizing for charts
Performance Optimization
5/10
Speed of the interface
Noticeable performance-minded choices (canvas optimizations, requestAnimationFrame, debouncing, Chart.js instance reuse) but no formal benchmarking or build-time bundle tuning shown.
Evidence
Sumit431511/Insight.IO-Dashboard/js/map.js: _pts caching and regeneration only on resize to avoid per-frame allocation
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/dashboard.html: animCount uses requestAnimationFrame; renderCharts destroys Chart.js instances before re-creating
Sumit431511/Know-Your-Heros/js/main.js: debounce(...) used to throttle search input
Accessibility & Semantics
4/10
Usable for everyone
Basic accessibility and focus handling are present (keyboard handlers, modal focus), some aria/semantic gaps and no automated a11y tooling in repo.
Evidence
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/dashboard.html: openNotesModal / closeNotesModal set focus to textarea after opening
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/login.html: Enter key handling to submit forms and visible spinner with accessible text
Sumit431511/Know-Your-Heros/hero.html: uses alt attributes for hero images and provides fallback loading / error UI
State Management & Data Flow
4/10
Managing data in the app
Server-state discipline is pragmatic: a central authFetch wrapper, Promise.all usage and optimistic UI updates exist, but no advanced cancellation, retry backoff, or state machine patterns.
Evidence
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/index.html: authFetch wrapper used consistently for API calls
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/dashboard.html: loadAll uses Promise.all for parallel endpoints and handles 401 to logout
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/dashboard.html: toggleReview implements optimistic UI update when PATCH may 404
UX & Visual Polish
5/10
Look and feel quality
Strong visual polish, many UX affordances (loading states, preview, spinners, animated bars, charts, heatmap and tooltips) though some patterns are present for demo rather than for full production edge cases (undo, offline sync).
Evidence
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/index.html: preview-box, btn-scan loading state, res-card with animated bars and heatmap display
Sumit431511/Neurascan---Brain-stroke-detection-system/frontend/dashboard.html: alert-banner, api-err, animated stat counters (animCount)
Sumit431511/Insight.IO-Dashboard/css/map.css and js/map.js: polished map view with scanline, robot marker and overlays
Expertise
HTML & CSS• Middle
Industries
Health Care• Middle
Robotics• Middle
Technologies
JavaScript• Middle
Chart.js
Bootstrap
Vite
Recommendations
- Develop clinician-facing dashboards and interactive medical imaging front-ends that require clear data-to-UI mapping, charts, and PDF export features.
- Build real-time telemetry and operations dashboards with custom canvas visualizations and responsive controls for robotics or IoT interfaces.
- Prototype interactive, static marketing or informational sites using Bootstrap/Vite and progressive enhancement techniques.
- Harden existing apps by adding request cancellation (AbortController), explicit race handling, and small state machines for modals / upload flows.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: