BHARATHKUMAR // AI.SYS
THE WORK LOOP // PERSONA.OS1
⚡ AI/ML ENGINEER STUDENT
RLCD.OS+
RAG EVAL
faithfulness
calibration

CHERALA
BHARATHKUMAR

Made to work, not just to demo

B.Tech CSE (AI & ML) student building end-to-end AI systems, from RAG pipelines and autonomous agents to evaluation, APIs, and deployment. I care about things that are tested, measured, and actually usable.

EXPLORE PROJECTS ↗ GITHUB 💻 LINKEDIN 💼
// 01 - PRODUCTION AI SYSTEMS

Built Systems & Agents

P.01

Production-RAG-Application

Fully offline contract RAG system with source attribution and metrics-gated CI/CD.
THE PROBLEM

Evaluating RAG applications on 500+ legal contracts was inconsistent, slow, and expensive, leading to vague uncited answers and model hallucinations.

WHAT I BUILT

Built a BM25 + vector search RAG pipeline over 510 CUAD contracts with cross-encoder reranking (top 20 to top 5). Raised Ragas faithfulness to 1.00 and answer relevancy to 0.71, gated by GitHub Actions CI workflow with a zero-cost local Ollama LLM.

PythonLangChainChromaDBOllama (Llama 3.1 8B)RagasGitHub ActionsStreamlit
THE WORK LOOP // RAG PIPELINEVIEW REPO ↗
510 CUAD Contracts (Legal Documents)
▼
BM25 + ChromaDB Vector Search (Top 20)
▼
Cross-Encoder Reranker (Top 20 → Top 5)
▼
Ollama Llama 3.1 8B Local Generation
▼
Ragas Faithfulness 1.00 Gate
P.02

Autonomous Research & Report Agent

Multi-agent research and synthesis engine with adversarial critic loops and human approval.
THE PROBLEM

Autonomous LLM agents commonly derail from tool-call budget exhaustion, broad topic reasoning drift, and unverified source hallucination.

WHAT I BUILT

Engineered a multi-agent LangGraph workflow (Planner → Researcher → Writer → Critic) where the Critic agent cross-examines citations against live Tavily scrape outputs. Bounded sub-tasks prevent reasoning drift; includes human approval gates.

LangGraphLangChainGroq (Llama-3.3)Tavily APIPythonStreamlit
AGENT TOPOLOGY // LANGGRAPH 4-NODE LOOPVIEW REPO ↗
User Research Prompt & Scope
▼
Planner Agent (Sub-Task Decomposition)
▼
Researcher Agent (Tavily Tools Fetch)
▼
Writer Agent (Markdown Draft)
▼
Adversarial Critic Loop (Verify & Gate)
P.03

VisionInspect AI — Industrial Defect Detection

Deep learning transfer learning defect classifier with severe class-imbalance resolution.
THE PROBLEM

Extreme class imbalance in manufacturing datasets caused standard classification models to default to the majority normal class, yielding an unacceptable 36% defect recall.

WHAT I BUILT

Fine-tuned ResNet50 across 3 MVTec AD defect categories using weighted cross-entropy sampling and gradient clipping. Increased minority-class defect recall from 36% to 91% (87–98% overall test accuracy) and shipped as a Dockerized FastAPI service.

PyTorchResNet50FastAPIDockerScikit-learnMVTec AD
CV PIPELINE // FASTAPI & RESNET50VIEW REPO ↗
Industrial Camera / Image Upload
▼
FastAPI Preprocessing & Normalization
▼
ResNet50 Fine-Tuned Backbone
▼
Weighted Imbalance Sampling
▼
Defect Classified (Recall 91%)
P.04

Faithfulness Guard Loop Evaluator

Real-time RAG evaluation loop gating hallucinated or uncited LLM responses.
THE PROBLEM

Production LLM systems silently output convincing assertions with zero factual grounding when retrieved context is ambiguous or incomplete.

WHAT I BUILT

Designed an automated evaluation guard loop running Ragas metrics at generation time. Low-confidence or ungrounded statements are automatically caught and rerouted for additional retrieval passes or filtered with clear fallback notices.

RagasLangChainLLM EvaluationPrompt EngineeringPython
EVAL LOOP // RAGAS METRIC GATINGVIEW REPO ↗
Generated LLM Response Candidate
▼
Ragas Faithfulness Metric Extractor
▼
Threshold Evaluation (Score >= 0.85)
▼
Pass to Client UI / Route Fallback
P.05

Pharmaceutical Inventory Analysis & POS Vision

Pharmaceutical analytics dashboard cutting reporting turnaround by ~80%.
THE PROBLEM

Manual multi-store medicine stock audit processes caused persistent stockouts, inventory shrinkage, and expired stock write-downs.

WHAT I BUILT

Automated ETL analytics pipeline using Python and SQL to ingest and analyze multi-thousand medicine inventory transactions. Built interactive Streamlit dashboards highlighting supply velocity and optimal reorder levels, slashing manual reporting time by 80%.

PythonSQLPandasSeabornStreamlitData Analytics
DATA PIPELINE // SQL & PYTHON ANALYTICSVIEW REPO ↗
Raw POS & ERP Inventory Records
▼
SQL Cleaning & Batch Deduplication
▼
Pandas Reorder Velocity Modeling
▼
Streamlit Dashboard (~80% Time Saved)
// 02 — STACK

What I build with.

LANGUAGES
PYTHON SQL
FRAMEWORKS
LANGCHAIN LANGGRAPH LLAMAINDEX FASTAPI STREAMLIT
AI / ML
GENERATIVE AI LLMS RAG PROMPT ENGINEERING AGENTIC AI MULTI-AGENT SYSTEMS HUMAN-IN-THE-LOOP LLM EVALUATION (RAGAS) ML PIPELINES LOCAL LLM (OLLAMA)
LIBRARIES
NUMPY PANDAS SEABORN PYTORCH TENSORFLOW KERAS HUGGING FACE SCIKIT-LEARN
DEVOPS
DOCKER GIT GITHUB GITHUB ACTIONS
● STATUS: OPEN TO ROLES

Let's Build Reliable Systems

Seeking AI/ML Engineer internship or full-time opportunities. Specialized in RAG pipelines, autonomous agents, and production evaluation.

✉ bharathkumarcherala@gmail.com
GLIDER 1.1 // SIGNAL