❯ who is adhith sankar?

I build LLM systems, and the evals that keep them honest.

A GenAI engineer with 2+ years of production experience building LLM, RAG and multi-agent systems for enterprise clients. He takes products from whiteboard to deployment, and measures whether they actually work.

GenAI Engineer · LLM, RAG & Agentic AI Systems Coimbatore, India Open to relocating anywhere in South India
retrieved context · top 3

The short version, retrieved.

chunk_01source: experience

2+ years taking GenAI products from whiteboard to deployment at Entelegen Analytics LLP: architecture, backend APIs, containerised deployment on SAP AI Core, GCP and AWS, and client handover.

chunk_02source: strengths

RAG pipelines, LangGraph multi-agent orchestration, and turning vague business problems into production AI systems people actually use.

chunk_03source: current focus

Evaluation. Shipping an LLM system is the easy half; knowing whether the last prompt or model change made it worse is the hard half. That is what I build now.

experience · trace view

Shipped to production, span by span.

Every project below went past the demo stage. Open a span to see what I built and with what.

trace_id: career · 4 child spans · status ok

Data Scientist · Entelegen Analytics LLP

Aug 2024 – Present — Official title Data Scientist; the work is GenAI engineering end to end.
AGENT › Agentic Purchase Requisition Automation handed over to production
  • LangGraph multi-agent system that automates purchase requisition workflows: each agent owns one step, the orchestration layer manages state and handoffs, and new agents plug in without changing the core graph.
  • Flask + WebSocket backend for a real-time chat interface, plus a parallel n8n low-code version the client can modify without engineering support.
  • Ran client demos and iterated on agent logic and UX through to production handover.
LangGraphLangChainFlaskWebSocketn8nDocker
RAG › Contract Document Intelligence: RAG Q&A 100+ contracts processed
  • RAG pipeline over enterprise contract PDFs: text, tables and images extracted, embedded with Sentence Transformers and indexed in FAISS on SAP Blob Storage.
  • A locally hosted Mistral model served via Ollama, so sensitive contract data stayed inside the client's infrastructure.
  • Ingestion and inference run as separate Dockerised services on SAP AI Core; used by the client team for contract review.
FastAPIFAISSSentence TransformersOllama · MistralSAP AI CoreDocker
RETRIEVAL › UNSPSC Commodity Recommendation Engine live on a customer-facing site
  • Maps free-text purchase descriptions to UNSPSC commodity codes, replacing a manual classification step: cosine-similarity retrieval over the embedded taxonomy, with a Mistral layer for ambiguous inputs.
  • Shipped as a Dockerised Flask REST API on SAP AI Core, then migrated to AWS Fargate as usage grew; live on the client's procurement site across multiple organisations.
FlaskEmbeddingsCosine similarityMistralSAP AI Core → AWS Fargate
RAG › NCERT Knowledge Assistant deployed for a Chennai school
  • Contributed to a RAG learning assistant over NCERT textbooks, deployed for a leading Chennai school.
  • Built the ingestion pipeline (text and image extraction, images to Cloud Storage, embeddings with metadata in Qdrant) and FastAPI/WebSocket services for chat Q&A with image citations and a quiz mode.
FastAPIQdrantGemini APIMongoDBKnowledge graphsGCP Cloud Run

Machine Learning Intern · TechAnalogy

Jun 2021 – Feb 2022

Integrated NLP/ML models into web services; built DataCLI, a Python (Click) CLI that automates preprocessing.

Data Associate Intern · ExpandAI

Aug 2021 – Nov 2021

Annotated training data and ran object-detection model experiments.

Student Research Assistant · SRM University

Nov 2021 – 2022

Co-authored 'Coronary Artery Disease Prediction using ML Algorithms' (IEEE ICESIC 2022).

flagship project · evaluation

RAG Hallucination Cascade

Catching made-up answers from a RAG system without paying an LLM judge to check every answer: free rules first, a small trained classifier next, and the judge only for what the first two can't settle.

pipelinebar = share of answers that reach the layer
L0
Input guardrailfree

6 rules for prompt injection and jailbreaks, checked on the question

L1
Rulesfree · ~0.07 s

Every answer. Flags hard evidence: a guess after a refusal, a number the excerpts never mention

L2
Classifierfree · ~0.14 s

Logistic regression on 12 cheap signals; only clears answers, never blocks

L3
LLM judge~$0.007 · ~3 s

Checks every claim against the retrieved excerpts · sees 35% of answers

35%
of answers sent to the judge
vs 100% judging everything
92%
of hallucinations caught (11 of 12)
rules alone: 50%
$2.60
per 1,000 answers
vs $7.08
0.19 s
typical check time
vs 2.65 s
SystemSent to judgeCaughtCost / 1kTypical time
Judge every answer100%100%*$7.082.65 s
Cascade35%92% (11/12)$2.600.19 s
Cascade, length rule instead of classifier45%92%$3.270.10 s
Rules only0%50%$00.06 s

40 held-out answers, 12 of them hallucinations, with thresholds chosen on training data first. 95% intervals: 20–50% sent to the judge, 71–100% caught. *Partly by construction: the test reference uses the judge's label where the human labeller agreed with it.

what didn't work

The classifier didn't beat a length rule

Test ROC-AUC 0.71 vs 0.80 for answer length, intervals overlapping. Its value is cost: about 10 points fewer judge calls at the same recall.

Hallucinations looked more like the contract, not less

They reuse its wording while misreading it, so word overlap was useless (AUC 0.52). Embeddings measure topic, not truth.

The judge was checked against people

60 answers labelled blind: 67% agreement (κ 0.25) at first, mostly from inconsistent refusal labels. Reported as is, not hidden.

The guardrail catches phrasings, not intent

19 of 24 test attacks blocked; all 5 reworded attacks written afterwards got through.

PythonFAISSQwen3 embeddingsMistral 7B (local)scikit-learnClaude as judgeStreamlitCUAD dataset
Read the code on GitHub ↗

Self-hosted n8n automation server

A personal automation server in Docker with persistent workflows: the same low-code tool I shipped to a client as an alternative to a hand-built agent.

n8nDocker Compose
stack · nearest neighbours

The tools I reach for, clustered.

Hover a cluster to see its neighbours. Placed by hand, not by an embedding model.

LLMs & agents Retrieval Evaluation Backend Cloud & deploy LangGraph LangChain Multi-agent systems Prompt engineering Ollama Mistral Gemini API Claude API NLP RAG FAISS Qdrant Sentence Transformers Semantic search Knowledge graphs LLM-as-judge Human agreement (κ) Cascades scikit-learn Guardrails Python FastAPI Flask WebSocket REST APIs MongoDB n8n Git Docker SAP AI Core GCP Cloud Run Cloud Storage AWS Fargate 5 clusters · 33 skills

LLMs & agents

LangGraphLangChainMulti-agent systemsPrompt engineeringOllamaMistralGemini APIClaude APINLP

Retrieval

RAGFAISSQdrantSentence TransformersSemantic searchKnowledge graphs

Evaluation

LLM-as-judgeHuman agreement (κ)Cascadesscikit-learnGuardrails

Backend

PythonFastAPIFlaskWebSocketREST APIsMongoDBn8nGit

Cloud & deploy

DockerSAP AI CoreGCP Cloud RunCloud StorageAWS Fargate
credentials

Paper trail.

Certification

Google Cloud Professional Data Engineer

Aug 2026 – Aug 2028

Education

B.Tech, Computer Science and Engineering

SRM Institute of Science and Technology · 2019–2023 · CGPA 8.9/10

Publication

Coronary Artery Disease Prediction using ML Algorithms

IEEE ICESIC 2022

contact · tool call

Building something with LLMs? Let's talk.

# open to GenAI / LLM / AI Engineer roles
await adhith.contact({
  email:    "adhithsagit0510@gmail.com",
  github:   "github.com/CodeSavitar",
  based_in: "Coimbatore, India",
})