Applied AI | Multi-agent systems | RAG | Document intelligence

Building production-grade AI systems that get past the demo.

I am Manish, a Bangalore-based Lead AI Engineer focused on agent orchestration, retrieval systems, evaluation workflows, multimodal evidence, and cloud deployments for enterprise AI products.

Open for contract work and full-time AI engineering roles
Download resume Start a conversation
Manish Sharma

About

Hands-on AI engineering with a builder's bias.

I work on the practical layers that make GenAI useful: workflow builders, retrieval, tool calling, structured evaluation, infrastructure, and the human loops around quality. Recent work spans pharma and life-sciences workflows, course discovery, datasheet intelligence, document extraction, and medical knowledge systems.

I like projects where research, product thinking, and shipping all sit at the same table. The sweet spot: taking ambiguous AI ideas and turning them into reliable systems that teams can run, measure, and improve.

Credentials

Research, education, and engineering credibility.

2021 - 2022

Research Assistant - Speech and Vision

Indian Institute of Science, Bangalore - SpireLabs

  • Worked on OCR and speech-recognition research pipelines with Hindi text and audio datasets.
  • Used PyTesseract, EasyOCR, Librosa, MelSpectrograms, and word-level accuracy evaluation.
  • Built the foundation for later production work in OCR, document AI, and multimodal retrieval.
2018 - 2022

B.Tech - Electronics and Communication Engineering

Nitte Meenakshi Institute of Technology

  • Graduated with 8.98 GPA.
  • Built a strong base in signal processing, systems, algorithms, and applied ML.
  • Moved from speech and vision research into production AI engineering roles.

Experience

Selected work

Aug 2025 - Present UsefulBI

Lead AI Engineer - GenAI

Leading enterprise GenAI platform work across multi-agent orchestration, no-code workflow builders, evaluation systems, quality intelligence, and AWS/EKS production deployment.

Multi-agent systems Workflow builder Agent tracing Agent evaluation harness Quality intelligence AWS/EKS

Gen AI Studio: Designed and implemented the CSR multi-agent orchestration layer using AWS and Strands, enabling modular agent workflows, task routing, reusable orchestration patterns, and enterprise GenAI use cases.

No-code agent builder: Built a generic drag-and-drop workflow canvas for composing configurable agent pipelines, connecting agent nodes, and reusing pipelines across CSR, PLPS, DSUR, and related operational flows.

Agent tracing and evaluation: Built structured multi-agent evaluation workflows to benchmark orchestration quality, tool calls, compliance score, response fidelity, accuracy, and task completion across agentic pipelines.

Engineering harness: Created reusable harness patterns for evaluating multi-agent collaboration, comparing pipeline behavior, experiment tracking, comparative testing, and validating deployment readiness before production release.

IQ Quality Intelligence: Built an AI-powered quality-events intelligence platform using historical quality events as query context, multimodal evidence, vector database retrieval, and similar-event matching for future quality checks.

User, FAM, and QE workflows: Owned end-to-end workflow design, backend architecture, and AWS/EKS deployment for the quality intelligence platform.

AI platform and DP: Architected the SUNY course discovery platform and designed the GQMD unified data repository by integrating Smartsheet Data Vault, GPLM architecture, and Databricks data pipelines.

Leadership: Led 4 engineers on client projects, conducted 10+ AI interviews, organized a company-wide medical science and pharma AI hackathon with judging criteria and prize structure, and curated 50+ AI/ML learning resources.

4 engineers led 10+ AI interviews 50+ AI/ML resources CSR / PLPS / DSUR agent flows
Jan 2024 - Jul 2025 Parspec.io

Machine Learning Engineer - 2

Built production ML and LLM systems for construction-product intelligence, datasheet retrieval, attribute extraction, annotation automation, and multimodal RAG evaluation.

LLM extraction Multimodal RAG Annotation automation AWS EKS

Datasheet intelligence: Built model-number to datasheet mapping across 5M documents using DynamoDB caching and Gemini 2.0 Flash.

Modeling: Fine-tuned Llama 3.3 70B for attribute extraction and improved family-name extraction recall from 89% to 96%.

Evaluation: Built GPT-4o/Gemini LLM-as-judge workflows and a BGE + CLIP + FAISS multimodal RAG pipeline.

80% latency reduction 96% recall 0.94 retrieval MRR
Dec 2022 - Dec 2023 Docsumo AI

Machine Learning Scientist

  • Designed and deployed document KV and table extraction systems using LayoutLM, BROS, and YOLO for fixed and unstructured documents.
  • Integrated ML and deep learning models into 10+ client APIs supporting $80K-$100K MRR workflows.
  • Built Chat-AI with LangChain and Pinecone for product QA and support tasks.
  • Reduced annotation time from one day to about two hours using GPT-4-powered KV extraction.
2021 - 2022 Indian Institute of Science, Bangalore

Research Assistant - Speech and Vision, SpireLabs

  • Worked with the Speech and Vision research group on OCR and ASR data preparation, extraction, and evaluation workflows.
  • Collated and preprocessed large Hindi datasets for OCR and speech-recognition experiments.
  • Used PyTesseract and EasyOCR for text extraction and evaluated OCR output using word-level accuracy.
  • Applied Librosa and MelSpectrogram-based audio processing for ASR experimentation.
Aug 2021 - Nov 2022 Zealth-AI (YC W21)

Machine Learning Engineer

  • Built OCR extraction and medical mapping flows for reports and prescriptions using Amazon Textract, RxNORM, MedXN, and SciSpacy.
  • Created a Neo4j Aura medical knowledge graph with 11K+ relationships and 3K+ nodes.
  • Used Rasa NLU to build a medical AI chatbot for symptom-based care management queries.

Projects

Side builds and experiments

01

RAG - Video-Based RAG System

Query YouTube URLs or uploaded videos, index chunks in Qdrant, retrieve relevant frames and timestamps, and answer with a Streamlit QA interface.

Qdrant RAG Video search Streamlit
GitHub Loom

02

AutoCommit Generator

A fully local commit-message generator using Ollama and Mistral, packaged as a fast terminal workflow in under 100 lines of bash.

Mistral Ollama Local AI Bash
GitHub X post

03

Company Scraper - AI-Powered Agent

A lightweight Relevance AI agent that turns company URLs into clean markdown briefs covering overview, products, features, audience, and integrations.

Relevance AI LLMs Markdown Research
X profile

Skills

Agentic AI, platform engineering, and production ML stack.

Agent engineering

AgentsMulti-agent collaborationAgent orchestrationTask routingTool callingReusable agent nodesWorkflow compositionNo-code agent builderAgent deployment

Agent evaluation

Agent tracingEngineering harnessEvaluation harnessExperiment trackingComparative testingLLM-as-judgeTool-call evaluationCompliance scoringResponse fidelityAccuracy checksTask completion

RAG and retrieval

RAGMulti-agent RAGMultimodal RAGVector searchEmbeddingsSimilar-event retrievalQdrantFAISSPinecone

Models and frameworks

GPT-4oGeminiLlama 3.3 70BPyTorchHugging FaceLangChainLangGraphLlamaIndexStrands Agent

Document AI and NLP

OCRLayoutLMBROSYOLOTable extractionKV extractionSpaCySciSpacyMedXNRxNORMLibrosa

Cloud and deployment

AWSEKSECSBedrockAPI GatewayDockerKubernetesProduction deploymentsStaging/prod partitions

Data pipelines and DP

DatabricksData pipelinesSmartsheet Data VaultGPLM architectureDynamoDBS3GCSMySQLNeo4jHasura

Backend and product systems

PythonFastAPIFlaskAPI designBackend architectureWorkflow buildersAnnotation workflowsUser/FAM/QE flowsLinux

Builder tools

CursorOllamaOpenRouterModal LabsGoogle ColabKaggleGitHubTableauAmplitudeLovable

Contact

Need someone to build, evaluate, or rescue an AI workflow?

I am open to contract projects and full-time positions in applied AI, GenAI platforms, document intelligence, RAG, and agentic systems.

Email Manish Connect on LinkedIn

Bangalore, India | manish.tinkering@gmail.com