ss.

Shail Shah.

I build AI systems that work beyond the demo. Stateful agents, reliable retrieval, and the infrastructure that brings it all together.

currently
AI & ML engineer @ Cloudoffis
working with
python / langgraph / vLLM
interested in
agents, retrieval, systems
explore my work linkedin email

Selected work.

Things I've built for the real world.

workflow architecture
Resume.Recover from the last saved state
  1. document
  2. classify
  3. checkpoint
  4. validate
ai infrastructurebuilt at Cloudoffis
AI Infrastructure

Stateful agents

Document workflows that remember where they left off. A checkpoint layer for multi-agent systems, built to recover from failure without starting over.

LangGraphvLLMPython 3.13CheckpointsVLMs
behind the build

The problem

Multi-agent document workflows suffered from unrecoverable failures, requiring expensive full-pipeline reruns and model retraining whenever a new document type was introduced.

The approach

Engineered a stateful multi-agent system on LangGraph with local LLMs/VLMs served via vLLM. Designed an immutable checkpoint-based state management layer for step resumption.

The result

Eliminated full-pipeline reruns, significantly reduced inference costs, and enabled instant document classification for new types without retraining.

workflow architecture
~86%Reduction in hallucinations
  1. query
  2. BM25 + vectors
  3. rank fusion
  4. validate
full-stack ragbuilt at Cloudoffis
Full-Stack RAG

Hybrid retrieval

Exact keyword matching meets semantic search. A hybrid RAG pipeline that combines BM25, vector retrieval, and fuzzy validation for more reliable answers.

RAGBM25Reciprocal Rank FusionChromaDBFAISS
behind the build

The problem

Standard vector-only document search generated hallucinations and missed exact key phrases in financial Q&A tasks.

The approach

Shipped a hybrid RAG retrieval engine combining dense vector search (FAISS/ChromaDB) with sparse Okapi BM25 keyword matching via Reciprocal Rank Fusion (RRF) and fuzzy answer validation.

The result

Reduced LLM hallucinations and improved Q&A retrieval accuracy by ~86% on production document workloads.

workflow architecture
~99%Table extraction accuracy
  1. financial PDF
  2. DotsOCR
  3. HTML table
  4. DataFrame
ai infrastructurebuilt at Cloudoffis
AI Infrastructure

Document vision

Turning complicated financial tables into usable data. A vision-language pipeline that preserves structure from document images to Pandas DataFrames.

DotsOCRvLLMVLMsIBM Granite VisionFlask
behind the build

The problem

Complex financial tables with irregular layouts and embedded graphics failed conventional text-based OCR parsers.

The approach

Built a multimodal vision-parsing pipeline using DotsOCR and IBM Granite Vision 3.3 (VLM → HTML → DataFrame conversion) served via Flask and vLLM.

The result

Achieved ~99% accuracy in tabular data extraction from complex financial PDF documents.

workflow architecture
~10sLess extraction latency per document
  1. document
  2. chunking
  3. parallel inference
  4. extraction
search & databuilt at Cloudoffis
Search & Data

Faster inference

Less waiting, more throughput. Optimized chunking, multithreaded classification, and a migration to Gemma-3 for financial document processing.

Gemma-3vLLMMultithreadingPyTorchLLaMA-2
behind the build

The problem

End-to-end inference latency on financial extraction was bottlenecked by legacy LLaMA-2 models and single-threaded inference paths.

The approach

Redesigned chunking strategies, re-architected classification paths for multithreaded execution, and migrated serving models from LLaMA-2 to Gemma-3.

The result

Cut end-to-end extraction latency by ~10 seconds and improved document classification speed by ~50%.

workflow architecture
End to end.Visibility across ML services
  1. ML services
  2. OpenTelemetry
  3. traces + logs
  4. ELK
distributed systemsbuilt at Cloudoffis
Distributed Systems

Observable systems

Understanding what happens between services. Distributed tracing and structured logs across production ML infrastructure, alongside a faster Python toolchain.

OpenTelemetryELK StackZipkinElastic APMuv
behind the build

The problem

Production ML services lacked distributed tracing, causing difficult debug cycles and slow, flaky CI/CD build pipelines.

The approach

Instrumented OpenTelemetry, Zipkin, Elastic APM, and ELK stack across ML microservices while leading platform upgrades (pip → uv, Python 3.10 → 3.13).

The result

Established complete end-to-end telemetry observability and significantly accelerated CI/CD build speeds and reliability.

01 / 05

Where I build.

From an idea to something people can depend on.

Oct 2023 – Present

AI & ML Software Engineer

Cloudoffis Pvt LtdAhmedabad, India

Building the systems behind production document AI.

  • Designed stateful multi-agent workflows with LangGraph and a checkpoint layer that resumes failed steps.
  • Shipped hybrid RAG with BM25, vector search, rank fusion, and fuzzy answer validation.
  • Built vision-language pipelines for financial tables, served locally with vLLM.
  • Connected production ML services with distributed tracing, OpenTelemetry, and the ELK stack.
  • LangGraph
  • vLLM
  • RAG
  • BM25
  • Gemma-3
  • OpenTelemetry
  • PyTorch
  • Python 3.13
  • DotsOCR
  • ELK

2022 – 2025

B.Tech, Computer Science & Engineering

LJ University

Specialized in Computer Science, Machine Learning, Data Structures, and Software Engineering.

2019 – 2022

Diploma in Engineering

R.C. Technical Institute (GTU)

Foundational engineering, programming principles, system administration, and computer networks.

Skills.

The tools change. The curiosity stays.

Proficiency
Languages & tooling
AI frameworks
Data & retrieval
Infrastructure
Python
FastAPI
vLLM
Hugging Face
Hybrid RAG
BM25
Qdrant
Bitbucket
Flask
uv
Git
LangGraph
Pandas
OpenTelemetry
Linux
AWS
Jenkins
Java
LangChain
PyTorch
scikit-learn
XGBoost
SMOTE
MySQL
Docker
RabbitMQ
Postman
C
C++
spaCy
YOLOv7
MobileNetV3
Postgres / pgvector
MLflow
ELK
Kubernetes
Nginx

Less reading. More tinkering.

Explore the ideas behind the work, right in your browser.

have an interesting problem?

Let's build
something good.

shailmshah@gmail.com
linkedin