I build AI agents
that actually work.
Retrieval pipelines, multi-agent graphs, and evals — the parts most people skip. I build them, then measure them, before I ship.
- CORRECTNESS
- 89.2%
- AVG_LATENCY
- 2.86s
- TURNAROUND
- 30–90s
- PIPELINE_SPEED
- 7m→90s
Ask me anything
A live RAG query interface grounded in resume text, project write-ups, and case studies. Same retrieval + guardrail pattern as DoCopilot — BM25 over a static corpus, Groq LLM for generation.// powered by rank-bm25 + groq
# Corpus: resume · projects · experience · skills · setup · interests · sports · media
// Answers are grounded in the corpus above. Off-topic questions are redirected. Source tags indicate which document each answer draws from.
The developer
AI/ML Engineer & Systems Developer
I am Samarth Pratap Singh, a BTech Computer Science student at VIT Bhopal University (CGPA 8.57) specializing in Generative AI, RAG systems, and LLMOps. I focus on building robust retrieval pipelines and multi-agent graphs where performance is measured and verified.
My expertise includes designing hybrid search indexes, implementing reciprocal rank fusion (RRF) rerankers, fine-tuning LLMs with PEFT/LoRA, and automating systems validation using LLM-as-a-Judge frameworks. I seek to build scalable AI applications with clear, documented results.
Experience
AI/ML Engineer Intern
•Architected a vision extraction pipeline for multi-page architectural PDFs using GPT-5 — cut latency from ~7 min to ~90 sec on 20-page documents via concurrent batch dispatch (ThreadPoolExecutor + asyncio), with structured output enforcement and retry handling.
•Built a cover/dimension intake pipeline combining non-AI heuristic extraction (regex/Docling) with conditional GPT-5 fallback — vision invoked only for missing fields — achieving >0.85 confidence and near-complete field extraction on real architectural lead sheets.
•Designed a job aggregation layer consolidating multi-source document outputs into structured JSON for documents up to 400 pages, with page-level validation guardrails and LangGraph-based routing.
Skills
Technical competencies structured by architectural layer. Evaluated and deployed across active projects.
RAG Pipeline compiler
Click the pipeline blocks below to construct a retrieval-augmented generation pipeline. Compile and run to evaluate accuracy and speed index.
Education
Bachelor of Technology in Computer Science and Engineering
Specializing in Artificial Intelligence, Machine Learning, and Cloud Computing.
// SELECTED_COURSEWORK
Activity Log
Real-time activity logs. Integrates live webhook tracking from GitHub profile contributions and LeetCode solve counts.
Featured Projects
Detailed case files for production-grade AI/ML architectures. Every project is measured against exact performance indices — no aesthetic fluff, only verifiable metrics.
DoCopilot — RAG Document Q&A
// PROBLEM
Baseline keyword-search retrieval was producing relevant answers in only ~57.7% of multi-format document queries, suffering from hallucinated context and zero citation trace.
// BUILD
Implemented a full-stack RAG pipeline. Utilizes hybrid search (BM25 + dense vectors) in Qdrant with Reciprocal Rank Fusion (RRF), cross-encoder reranking via hosted inference API, and source-grounding citation filters. Added PII redaction and prompt-injection detection guardrails.
Argus — Multi-Agent Research Engine
// PROBLEM
Compiling research reports from Tavily, ArXiv, and Wikipedia was manual and time-consuming, requiring human iteration to trace sources and reject low-quality summaries.
// BUILD
Built an autonomous research supervisor graph with LangGraph containing 5 specialist agents (planner, researcher, critic, writer, supervisor). Features cyclic routing, SQLite checkpointing for failure recovery, and end-to-end tracing in LangSmith. The critic agent rejects low-quality drafts and re-routes before writing.
ContextCore — Stateful Memory Agent
// PROBLEM
Standard agents lose task/note context across sessions, and typical memory approaches hallucinate user preferences when querying vector databases directly.
// BUILD
Designed a FastMCP server exposing structured note/task tools integrated with LangGraph. Employs a dual-memory layer: Postgres for exact execution states, MongoDB for profile states, and Qdrant semantic recall. Intent router directs queries based on semantic similarity.
GFS-AI — Document Intelligence Pipeline
// PROBLEM
Intake validation for multi-page (up to 400p) architectural PDF documents took up to 7 minutes with high rates of timeout errors and missing dimension listings.
// BUILD
Architected a vision extraction pipeline combining non-AI heuristic regex/Docling checks with GPT-5 fallback. Integrated concurrent batch dispatch (ThreadPoolExecutor + asyncio) and fixed canvas resizing bugs to improve field detection. Built at AmberFlux EdgeAI.
Historical Archive
Earlier experimental pipelines, PEFT notebook fine-tuning runs, and core full-stack foundations. Archived for log completeness.
FLAN-T5 Dialogue Summarizer
ARCHIVEDLoRA fine-tuned FLAN-T5-base on SAMSum dataset (14.7K dialogues), updating only 2% of parameters. Achieved 49.01 ROUGE-1 · 72.25 BERTScore F1 · 42.51 METEOR.
RoBERTa Banking77 Classifier
ARCHIVEDFine-tuned RoBERTa-base on Banking77 dataset (77 intents, 13K queries) with AdamW and mixed precision. Achieved 93.7% accuracy and 93.6% macro-F1.
Project Loom
ARCHIVEDFull-stack project sharing board featuring Next.js SSR/ISR, auth via NextAuth, and automated schema content delivery with Sanity.io headless CMS.
Active Development
AgentGuard
An AST-based static analysis and observability CLI/GitHub Action for agentic AI codebases (LangGraph, CrewAI, AutoGen, MCP). Automates cyclic graph checks, checks FastMCP tool schema alignment, and implements three detection rules for missing checkpoint handlers. Backed by a full pytest validation suite.
Receipts
Milestones, verified credentials, and publications tracked across academic and engineering pursuits.
Comprehensive 5-course program covering troubleshooting, networking, operating systems, system administration, and security. Credential ID: whvAjzYf
Applied Machine Learning in Python
Completed specialization in supervised/unsupervised learning, feature engineering, model evaluation, and scikit-learn for practical ML applications.
Published Fine-Tuned Models on Hugging Face
FLAN-T5 Summarizer with reproducible evaluation achieving 49.01 ROUGE-1 and 72.25 BERTScore F1 on SAMSum dataset
VIT Bhopal Academic Excellence
Maintained 8.57 CGPA with focus on AI/ML coursework including DSA, Cloud Computing, and Software Engineering
Get in touch
Establish connection for consulting, pipeline audits, or research collaborations. Communication will be answered within 24 hours.