Brijesh Singh

AI/ML engineer and independent researcher · Hyderabad, India

I build production ML systems and study agentic RAG architectures.

Education

Bachelor of Business Administration in Business Analytics Expected May 2026
ROOTS Collegium, Hyderabad
Specialization: Finance

Research Experience

Retrieval Degradation in Multi-Turn Agentic RAG Systems

Brijesh Singh. Manuscript in preparation, 2026.

Status: arXiv submission pending endorsement.

This paper examines how retrieval quality degrades as reasoning turns accumulate in a production agentic RAG deployment. Drawing on 150 sessions and 550 turn-level observations instrumented from Math Professor AI, the study documents a statistically significant 6.6% decline in retrieval quality (p < 0.0001) and identifies context length as the dominant predictor (Pearson r = −0.283). Three contributing mechanisms are characterized — context length interference, keyword drift, and attention head saturation — and a lightweight turn-aware confidence estimator is proposed (AUROC = 0.634). The findings carry direct implications for how agentic RAG systems should be evaluated and designed.

  • Empirical study of retrieval quality degradation across reasoning turns in a production agentic RAG deployment (150 sessions, 550 turn-level observations instrumented from Math Professor AI)
  • Documented statistically significant decline (6.6%, p < 0.0001) and identified context length as the dominant predictor (Pearson r = −0.283); proposed lightweight turn-aware confidence estimator (AUROC = 0.634)
  • Characterized three contributing mechanisms (context length interference, keyword drift, attention head saturation) with implications for agentic RAG evaluation and system design

Technical Skills

Languages
Python, SQL, TypeScript, JavaScript
ML / AI
Random Forest, XGBoost, RAG, Agentic AI, LLMs, Stable Diffusion, LoRA
Backend
FastAPI, REST APIs, Celery, Redis
Frontend
React, Next.js, TypeScript, Tailwind CSS
Cloud / DevOps
GCP, Docker, PostgreSQL, Vercel

Key Projects

Math Professor AI — Agentic RAG Education Platform

Google Gemini SDK, TypeScript, React, Vite, Python

  • Built agentic RAG system using Google Gemini 2.5 Pro with intelligent query routing between local knowledge base and web search grounding for step-by-step math tutoring
  • Implemented input guardrail pipeline using Gemini 2.5 Flash for math-relevance validation, reducing off-topic queries before LLM processing
  • Designed human-in-the-loop feedback loop enabling users to flag incorrect reasoning steps; corrections feed back into the knowledge base to improve future responses
  • Production deployment served as instrumented platform for empirical research on multi-turn retrieval degradation (manuscript in preparation)

Smart Factory Analytics Platform

Python, FastAPI, scikit-learn, React, Next.js, TypeScript, SQL, Docker

  • Trained Random Forest classifier achieving 97.91% accuracy for predictive maintenance on 103K+ simulated industrial sensor samples
  • Engineered real-time analytics dashboard with failure probability scoring, K-Means anomaly detection across 4 cluster patterns, and automated maintenance alerts
  • Designed FastAPI microservice with ML inference endpoints, SQL ETL pipelines for time-series aggregation, and Docker containerization

SAKURA — Anime Character Consistency System (Architecture & Design)

Python, SDXL, LoRA, IPAdapter, FastAPI, React, PostgreSQL, Redis, Docker

  • Designed character DNA encoding system using 512-dim face embeddings (InsightFace/ArcFace), MediaPipe pose estimation, and CLIP semantic features to preserve anime IP consistency
  • Architected Style Preservation Network with Gram matrix style transfer, line weight analysis, and color harmonization across four weighted style components
  • Specified production pipeline with async FastAPI job queues (Celery/Redis), consistency validation engine, and React/Next.js studio dashboard — in active development

Additional Production Applications

  • FinVision finvision-g.streamlit.app — AI financial analyst automating report parsing and analysis in 10–20 seconds vs. hours manually
  • Stock Analysis Pro stock-analysispro.streamlit.app — Real-time trading platform with 6+ technical indicators, backtesting engine, and AI pattern recognition
  • Gemini AI Assistant geminiflow.streamlit.app — Multi-modal document analysis with Gemini API integration

Open Source & Certifications

Open Source
Merged contribution to Alibaba-NLP/DeepResearch (Issue #130) — improved configuration handling for the deep research agent framework.
Certifications
IBM Data Science Professional Certificate (2024)