AI EngineerFinal-year CSE · VIT BhopalOpen to 2027 roles

I build LLM systems,
then measure them.

RAG pipelines, document extraction, and the eval harnesses behind them. In my internship, I cut extraction time on 20-page drawings from 7 minutes to 90 seconds.

View work Download resume Ask my AI ↗

github.com/noviciusss linkedin.com/in/spsamar huggingface.co/noviciusss

89.2%
answer correctness · 40-question eval · DoCopilot: 2.86s
avg latency · DoCopilot: 7m → 90s
PDF extraction, 20 pages · internship project: 92%
pass rate · 50-case eval · ContextCore

stack:LangGraphQdrantFastAPIPostgreSQLAzureMCP

// CASE_FILES

Featured Projects

Evaluated AI/ML architectures with clear problem statements, production builds, and verifiable metrics.

FLAGSHIP PROJECT — CASE_FILE // 0140-QUERY EVAL

CASE_FILE // 01

DoCopilot — Multi-Tenant RAG

Multi-tenant RAG with 89.2% correctness and 2.86s average latency on a 40-question eval harness.

Correctness89.2%

Relevance90.5%

Avg Latency2.86s

✦Eval: 40 questions · LLM judge Llama-3.1-8B judging Qwen3-32B

// PROBLEM

Document Q&A across multi-format business files suffered from hallucinated context, absent source attribution, and lack of strict multi-tenant boundary isolation.

// BUILD

Engineered a production-ready multi-tenant RAG platform with JWT authentication and Qdrant payload-filtered retrieval. Features SHA-256 upload deduplication, Celery + Redis async ingestion, hybrid search (MiniLM dense vectors + BM25 with Reciprocal Rank Fusion), and cross-encoder reranking from top-20 down to top-5. Implemented SSE token streaming, source-cited grounded answers, PII/regex guardrails, per-IP rate limiting, and CI/CD via GitHub Actions (pytest → Docker → ACR → Azure Container Apps).

FastAPINext.jsQdrantHybrid Search (RRF)Cross-Encoder+2

Code

Live

Analysis

CASE_FILE // 02

Argus — Multi-Agent Research Engine

Autonomous research supervisor coordinating 4 specialist agents with 3 depth tiers and LangSmith tracing.

Depth Tiers~20/45/90s

Specialists4 + Supervisor

Critique LoopsMax 3 Loops

✦Max 3 critique loops · 30-min auto-finalize · SQLite/Postgres auto-switch

// PROBLEM

Compiling deep research reports across heterogeneous web sources (Tavily, ArXiv, Wikipedia) was manual and unverified, requiring repetitive human oversight to filter noise and reject weak summaries.

// BUILD

Built a LangGraph supervisor agent orchestrating 4 specialists (planner, researcher, critic, writer) routed dynamically via Command(goto). Features 3 configurable depth tiers (~20s / 45s / 90s), isolated Human-In-The-Loop interrupt nodes, maximum 3 critique feedback loops before 30-minute auto-finalization, SSE streaming, SQLite/Postgres auto-switching checkpoint persistence, and complete execution tracing in LangSmith packaged with Docker.

LangGraphFastAPILangSmithDockerTavily+1

Code

LiveHosted on Render — cold start may take ~30s

CASE_FILE // 03

ContextCore — Stateful Memory Agent

LangGraph CLI agent with 3-tier memory persistence achieving 92% pass rate across 50 eval cases.

Eval Pass Rate92%

Memory Stores3 Tiers

ProtocolFastMCP

✦Eval: 50 cases · 92% pass rate · Hardened tool schemas

// ARCHITECTURE // 3-TIER MEMORY

1. PostgreSQL

Execution Checkpointing & State History

2. Qdrant

Semantic Recall & Vector Similarity

3. MongoDB

User Profile & Preference State

→Routed via FastMCP tools into LangGraph agent executor

// PROBLEM

CLI agents lose state across sessions, and standard single-vector memory stores hallucinate user profile data or overwrite active execution states.

// BUILD

Designed an intent-routed LangGraph CLI agent (task / memory-recall / chat) backed by a custom FastMCP server on PostgreSQL. Implements three distinct memory layers: Postgres checkpointing for exact step states, Qdrant for semantic recall, and MongoDB for persistent user profiles. Features astream_events token streaming, async background memory persistence, and an LLM-judge test suite with 92% pass rate across 50 test cases.

FastMCPLangGraphPostgreSQLQdrantMongoDB+1

Code

INTERNSHIP PRODUCTION WORK — AmberFlux EdgeAI — CASE_FILE // 04LEAD SHEETS

CASE_FILE // 04INTERNAL

GFS-AI — Document Intelligence Pipeline

Architectural drawing extraction pipeline cutting intake from ~7 min to ~90 sec (~4.7×) on 20-page drawings.

Extraction Speed7m → 90s

Confidence>0.85 Conf.

Doc ScaleUp to 400p

✦AmberFlux EdgeAI · Production Lead Sheets · Up to 400 pages

// EXTRACTION_PIPELINE // HEURISTIC + VISION FALLBACK

PDF Intake (≤400p)

Docling & Regex heuristic parsing

Conditional Dispatch

GPT-5 vision only on missing fields

LangGraph Aggregation

Schema validation & retry guardrails

// PROBLEM

Intake validation for multi-page architectural CAD/PDF drawings took ~7 minutes with frequent timeout errors and high risk of missed lead-sheet specifications.

// BUILD

Architected a high-throughput document intelligence layer combining non-AI heuristic checks (regex & Docling) with conditional GPT-5 vision fallback — invoking vision only for missing fields. Integrated concurrent batch dispatch (ThreadPoolExecutor + asyncio), retry handling, schema-enforced JSON outputs, and LangGraph-routed aggregation with page-level guardrails and drawing-page exclusion.

GPT-5 VisionThreadPoolExecutorasyncioDoclingLangGraph+1

Internal (AmberFlux)

// EXPERIENCE

Experience

May 2026 – Present · Remote (Hyderabad, Telangana)

AI/ML Engineering Intern

AmberFlux EdgeAI Private Limited

May 2026 – Present·Remote (Hyderabad, Telangana)

// Latency reduction7m → 90s (~4.7×)

// Document rangeUp to 400p

// Vision extraction>0.85 conf.

•Architected a vision extraction pipeline for multi-page architectural PDFs using GPT-5 — cut latency from ~7 min to ~90 sec on 20-page drawings (~4.7×) 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 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, drawing-page exclusion, and LangGraph-routed aggregation.

TECH:GPT-5 VisionThreadPoolExecutorasyncioLangGraphDoclingPython

// STACK

Skills

Technical competencies structured across core engineering layers — verified across active systems and evaluation benchmarks.

// Languages

PythonC++TypeScriptSQL

// ML / DL

PyTorchTensorFlowTransformersPEFT / LoRAScikit-learn

// LLM & Agents

LangGraphLangChainRAG SystemsTool-CallingFastMCP

// Retrieval

QdrantFAISSHybrid Search (BM25 + Dense)RRF FusionCross-Encoder Rerank

// Backend & Web

FastAPIPostgreSQLMongoDBNext.jsRedisConcurrency (asyncio)

// Delivery & DevOps

DockerAzure Container AppsGitHub Actions CI/CDpytestLangSmith

// Evaluation & Observability

LLM-as-a-JudgeRagasLangSmith TracingMLflowW&B

// QUERY_INTERFACE

Ask me about my work

Grounded Q&A over my resume, internship, and project write-ups.
Lightweight version: BM25 keyword retrieval over a static corpus + Groq. DoCopilot's full hybrid pipeline (dense + BM25, RRF, cross-encoder rerank) is in the case file above.

Try asking:

TERMINAL // bm25-groq-runtime

CORPUS: RESUME_V8.61

❯ ask --about samarth

Ask any question about my technical experience, architecture decisions, or internship at AmberFlux.

// BACKGROUND

About

I'm a final-year CSE student at VIT Bhopal and an AI/ML engineering intern at AmberFlux EdgeAI, where I own the AI vision-extraction layer of a document pipeline for architectural drawings. I like the unglamorous parts of AI systems: retrieval quality, evaluation harnesses, failure handling, and cost-aware routing.

Outside work I build LLM systems end to end (RAG, tool-calling, MCP), measure them with LLM-judge evals, and ship them with CI/CD. I'm looking for AI/LLM engineering roles starting 2027.

// FOCUS

RAG & LLM Agents

// EVALUATION

LLM-as-a-Judge

// TARGET

2027 AI Engineering

Download Resume (PDF)

// CREDENTIALS

Education & publications

B.Tech in Computer Science and Engineering

VIT Bhopal University

2023 – 2027Bhopal, Madhya Pradesh

// VERIFIED_ACADEMIC_CGPA8.61 / 10

// SELECTED_COURSEWORK

Data Structures & AlgorithmsOperating SystemsDatabase Management SystemsComputer NetworksCloud ComputingSoftware Engineering

Published Fine-Tuned Models (Hugging Face)

FLAN-T5 Dialogue Summarizer

LoRA Fine-Tuning · SAMSum Dataset

// 49.01 ROUGE-1 · 72.25 BERTScore F1 · 42.51 METEOR

Parameter-efficient fine-tuning updating only 2% of parameters while matching full fine-tuning performance.

Model

RoBERTa Banking77 Classifier

Intent Classification · 77 Categories

// 93.7% Accuracy · 93.6% Macro-F1

Fine-tuned with AdamW and mixed precision on 13K customer banking queries.

Model

Certifications:

•Applied Machine Learning in Python (University of Michigan (Coursera), 2025)

•Google IT Support Professional Certificate (Google Career Certificates, 2026)verify ↗

// ARCHIVE

Fine-tuning & earlier work

Earlier experimental pipelines, PEFT fine-tuning runs, and machine learning foundations.

FLAN-T5 Dialogue Summarizer

MODEL

LoRA 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.

// STACK: PYTHON · LORA · TRANSFORMERS · GRADIO · HF SPACES

Code Live

RoBERTa Banking77 Classifier

MODEL

Fine-tuned RoBERTa-base on Banking77 dataset (77 intents, 13K queries) with AdamW and mixed precision. Achieved 93.7% accuracy and 93.6% macro-F1.

// STACK: PYTORCH · TRANSFORMERS · CUDA · PYTHON

Code Live

// Earlier Web Projects: Project Loom (Next.js, Sanity CMS, NextAuth full-stack project sharing platform).

github.com/noviciusss/projectloom ↗

// INTERACTIVE PIPELINE HARNESSSIMULATION — illustrative numbers

RAG Pipeline compiler

Interactive simulation demonstrating how retrieval components chain together to optimize accuracy and latency.

// Available Blocks:

// Current Layout:

No blocks selected. Click blocks on the left.

// ACTIVITY_LOG

Activity

Public engineering activity tracked across GitHub open-source repositories and LeetCode algorithmic problem solving.

GitHub Profile

Public Repos

Followers

// HANDLE@noviciusss

LeetCode Algorithmic

Easy

Medium

Hard

// TOTAL_SOLVED180 Problems

// CONNECT

Contact

Hiring for AI/LLM engineering roles or internships? I reply within 24 hours.

samarthsin2006@gmail.com

Phone

+91 9452026413

Location

Pratapgarh, U.P., India

Direct inbox: samarthsin2006@gmail.com

I build LLM systems,then measure them.

Featured Projects

DoCopilot — Multi-Tenant RAG

// PROBLEM

// BUILD

Argus — Multi-Agent Research Engine

// PROBLEM

// BUILD

ContextCore — Stateful Memory Agent

// PROBLEM

// BUILD

GFS-AI — Document Intelligence Pipeline

// PROBLEM

// BUILD

Experience

AI/ML Engineering Intern

Skills

Ask me about my work

About

Education & publications

B.Tech in Computer Science and Engineering

// SELECTED_COURSEWORK

Published Fine-Tuned Models (Hugging Face)

FLAN-T5 Dialogue Summarizer

RoBERTa Banking77 Classifier

Fine-tuning & earlier work

FLAN-T5 Dialogue Summarizer

RoBERTa Banking77 Classifier

RAG Pipeline compiler

Activity

Contact

I build LLM systems,
then measure them.