Devesh Jaluka
AI/ML Engineer focused on RAG and FastAPI development, Claude Certified Architect.
Hyderabad
2+ years of work experience
Mar 20269 mosPresent
Hyderabad, Telangana, India
Architected and shipped a production agentic NL2SQL analytics platform over a high-dimensional, multi-million-record
longitudinal data warehouse; designed a multi-stage LangChain workflow for question decomposition, structured query
planning, scope validation, self-correcting SQL generation, execution, and adaptive analytical outputs, achieving 98% SQL
accuracy on a curated golden evaluation set.
– Engineered the analytics platform’s production reliability and data layer through PostgreSQL migration and parity
validation, connection pooling, read-only execution, multi-layer SQL guardrails, query bounds, and structured tracing; built a
2K+ case regression suite and a standalone MCP server for secure, tool-based schema discovery and analytical querying.
– Designed and shipped a production multimodal RAG system for complex presentation-heavy knowledge sources,
implementing 3-stage cascading retrieval with progressive metadata relaxation, vector fallback, and cross-encoder reranking,
achieving 0.75 MRR@10 on retrieval evaluations.
– Reduced the RAG system’s end-to-end P50 latency by 67% (90s → 30s) by redesigning the inference path with non-blocking
async execution, persistent connection reuse, concurrent model racing with early cancellation, token-level SSE streaming, and
adaptive context compression for large retrieved contexts.
– Built a reusable GenAI evaluation and API benchmarking framework to compare production AI systems across
curated datasets, combining fuzzy matching, embedding-based semantic similarity, and LLM-as-a-Judge with repeated-run
consistency analysis to quantify accuracy, semantic agreement, response stability, and detect reliability regressions.
-Service Delivery Partner:- Wipro
Nov 20251 yr 1 moPresent
Hyderabad, Telangana, India
May 20257 mosNov 2025
Remote
•Developed a modular agent-based context aware AI pipeline using LangGraph, LangChain, RAG Quadrant, OpenAI Embeddings, Perplexity API for research , and Azure OpenAI to dynamically generate high-quality PowerPoint presentations.
•Designed and implemented a custom decision graph using LangGraph to coordinate multiple specialized agents across a generation workflow.
•Applied advanced prompt engineering and retrieval techniques to enhance content accuracy, relevance, and structure.
•Built a chatbot generation system leveraging LangChain, RAG architecture, and web scraping to create custom knowledge-based conversational agents.
•Integrated web content via scraping tools like BeautifulSoup and Playwright, processed it into vector databases using FAISS, and embedded responses with Azure OpenAI.
•Fine-tuned the Salesforce BLIP vision-language model on the MIMIC-CXR dataset to improve chest X-ray captioning performance for Indian patients.
•Incorporated CLIP embeddings, domain-specific data augmentation, and custom captioning techniques to achieve near state-of-the-art results.
•Utilized tools such as PyTorch, Hugging Face Transformers, Grad-CAM, and TensorBoard for model training, evaluation, and interpretability.
Jul 20246 mosDec 2024
Developed a CNN-LSTM model to predict dynamic land cover changes using 39 years of satellite imagery data for Chatham, Massachusetts. Preprocessed annual satellite images to 256×256 resolution and applied image normalization for efficient model training. Contributed to the publication demonstrating the model’s ability to capture spatial and temporal patterns for environmental management.
Aug 20234 mosNov 2023
Remote
Devesh’s AI Stack
Education
Bachelor of Technology, Computer Science
Vellore Institute of Technology
20214 yrs2025
Junior Collage, Science and Mathematics
Delhi Public School, Ranchi
20182 yrs2020
Join Devesh and 10k+ other professionals on
Tal is the app where bosses meet engineers & builders directly, without middlemen
Download Tal Boss