AI Engineer Portfolio
I am Neeraj Kumar Singh, an Applied AI / LLM engineer with 4+ years across NLP, explainable AI, model interpretability, production-oriented RAG systems, and backend AI tooling.
I work at the intersection of applied LLMs, trustworthy AI, and production engineering. My recent work includes LLM attribution, retrieval evaluation, agent workflows, and Python/PyTorch/FastAPI tooling that turns research ideas into usable systems.
Core stack
Python, PyTorch, FastAPI, LangGraph, LangChain, Hugging Face, PostgreSQL, Redis, ChromaDB, Elasticsearch, OpenTelemetry, and RAGAS.
Dec 2025 - Mar 2026
Explainable retrieval-augmented generation system with hybrid search, retrieval attribution, and streaming FastAPI endpoints.
Combines ChromaDB, Elasticsearch, RRF, BGE re-ranking, DeBERTa NLI faithfulness checks, RAGAS, OpenTelemetry, Redis, and SSE delivery.
Open project2026
Public research assistant with Chainlit UI, LangGraph orchestration, LiteLLM model access, and typed search tools.
Uses query-kit provider search, PostgreSQL persistence, Redis caching, and streamed progress updates for research workflows.
Open project2026
Provider-agnostic Python CLI and library for searching public research sources with normalized text and JSON output.
Covers ACL Anthology, arXiv, PubMed, Semantic Scholar, and OpenReview with deduplication, filtering, and sync/async APIs.
Open project2026
LangGraph workflow with typed tools, structured outputs, retrieval attribution, verifier/retry loops, and human review.
Evaluates faithfulness, context precision and recall, factual correctness, latency, and tool-call behavior for trustworthy RAG.
Open project2024 - 2026
Model-agnostic interpretability framework for relevance tracing across MLPs, LLMs, and custom deep networks.
Benchmarked against SHAP, LIME, and Integrated Gradients; extended in production work for transformer and MoE attribution.
Open projectPublished
Indic hate-speech mitigation pipeline for classification, explanation, and intensity-reduction suggestions.
Validated across five Indic languages with BERTScore results in the 0.96-0.99 range for mitigation quality.
Open projectLLM attribution and explainability for transformers, LLaMA-family models, and Mixture-of-Experts architectures
RAG and agent systems with hybrid search, retrieval attribution, verifier loops, and quality metrics
Backend AI delivery with Python, FastAPI, async APIs, tracing, persistence, caching, and deployment-ready tooling
A dedicated publications and talks index is being expanded. Current public activity and technical updates are available through the channels below.
April 2024 - Present
Shipped DLBacktrace v2 for explainability across tabular, vision, NLP, and LLM workloads; expanded transformer and MoE support, compressed attribution traces 17-93x with low error, and built scalable visualization and trace-storage tooling.
September 2021 - February 2024
Contributed to TrustED for trustworthiness evaluation in deep learning; developed explainable Indic hate speech detection and worked on an IC50 prediction web application integrating deep learning models.
2019 - 2021
Percentage: 77.45
2015 - 2019
Percentage: 71.82
For AI engineering roles, collaboration, or technical discussion, reach out by email or LinkedIn.