Manish Sharma 📊
Lead AI Engineer | Retrieval, Search & Ranking, Evals, Agent Memory | Multi-agent systems built end to end and shipped on AWS, GCP & Kubernetes | Research @ IISc (Speech, Vision)
Greater Bengaluru Area
4+ years of work experience
Aug 20251 yr 3 mosPresent
Bangalore Urban, Karnataka, India
Leading a team of engineers building production agentic AI systems on AWS.
• Multi-agent orchestration: designed the orchestration layer of a multi-agent platform on AWS + Strands (task routing, modular agent workflows, reusable orchestration patterns) and a no-code drag-and-drop DAG builder for composing agent pipelines.
• Retrieval-grounded agents: built a quality-events intelligence system that retrieves similar historical events with multimodal evidence from a vector store as context for new checks; owned the backend architecture and AWS/EKS deployment.
• Evals and regression gating: built the eval harness for a Strands + AWS Bedrock decision-tree agent: golden datasets, tool-call trajectory graders, span-level tracing of tokens, cost and latency, a human-calibrated LLM-as-a-Judge, pass@k / pass^k consistency checks, and a regression gate that blocks releases below baseline.
• Online evals: turned reviewer decisions into production labels by comparing AI-suggested vs human-final answers per decision node, giving a continuous AI-human agreement metric with no extra instrumentation.
• Memory and context engineering: role-scoped short-term memory in DynamoDB, long-term episodic memory in a vector database for grounding, and audit trails for every override.
• Data infrastructure: Databricks pipelines that unify multiple enterprise sources into one repository.
• Ran 10+ technical interviews (ML, LLMs, system design) and a company-wide hackathon.
Stack: Python, AWS (Bedrock, EKS, DynamoDB), Strands Agents, vector databases, Databricks.
Jan 20241 yr 7 mosJul 2025
India
Core ML team building LLM and retrieval systems over construction product documents.
• Retrieval at scale: model-number-to-datasheet recommendation across 5M documents; 80% lower latency with DynamoDB caching and Gemini 2.0 Flash via OpenRouter.
• Fine-tuning: fine-tuned Llama 3.3 70B Instruct on H100s (Modal) with a custom Alpaca-format dataset for attribute extraction.
• Extraction: family-name extraction with Gemini 2.5 Flash, raising recall from 89% to 96%; extended header and column detection from 4 to 8 columns at 97% accuracy, cutting manual processing by 30%.
• LLM-as-a-Judge: parallel GPT-4o and Gemini 2.5 Flash judges that triage datapoints for annotation, bringing human annotation down to ~74%.
• Multimodal RAG: BGE text and CLIP image embeddings on FAISS over a mixed image and text knowledge base; 0.94 MRR in retrieval and 0.96 accuracy in generation.
• Infrastructure: migrated the AI Kubernetes workload to AWS EKS, with the load balancer behind API Gateway and isolated staging and production.
Stack: Python, PyTorch, Hugging Face, Gemini, OpenAI, FAISS, AWS (EKS, DynamoDB, API Gateway), Modal.
Nov 20221 yr 1 moNov 2023
Remote
Document AI for enterprise extraction APIs.
• Layout-aware extraction: designed and shipped Document KV and Table extractors using LayoutLM, BROS and YOLO for both fixed-template and unstructured documents.
• Productionized these models across 10+ custom client APIs generating $80K-$100K MRR.
• Chat-AI: built in-product LLM question answering and support on LangChain with Pinecone retrieval.
• GPT-4 KV extractor for LLM pre-annotation, cutting annotation time from a full day to ~2 hours.
Stack: Python, PyTorch, Hugging Face Transformers, LayoutLM, BROS, YOLO, OCR, LangChain, Pinecone, OpenAI.
Aug 20211 yr 1 moAug 2022
India
Building Digital Therapeutics for cancer patients using AI.
• Built and deployed OCR Extraction and Medical Mapping Flow for Medical Documents, Reports and
prescriptions using Amazon Textract, RxNORM, UMLS and SciSpacy. The Entire pipeline takes 8-10
seconds
• Architected and deployed the Core-Carepathway Algorithm and API’s using MedXN and Knowledge
Graph from scratch.
• Built Customized Symptom Checker Algorithm using Neo4j Aura and Neo4j Community with Cypher to
build the Entire Knowledge Graph Database and Relationships Mapping
Jun 20211 yr 3 mosAug 2022
Bangalore Urban, Karnataka, India
I, and the team will be working under Prof. Prasanta for collecting and building high end Automatic Speech Recognition model's for farmers which can help them understand about their needs.
Manish’s AI Stack
Education
Research Assistant, Speech and Vision
Indian Institute of Science (IISc)
20211 yr2022
Electronic and Communication Engineering, B Tech
Nitte Meenakshi Institue Of Technology
20184 yrs2022
Join Manish and 10k+ other professionals on
Tal is the app where bosses meet engineers & builders directly, without middlemen
Download Tal Boss