portfolio · v2026

JAYDATTA
BADEAI Engineer, Agentic AI & LLM Systems

AI Engineer|Agentic AI|LLM Inference|RAG Systems

AI Engineer and Data Scientist with 2+ years of experience designing, developing, and deploying Generative AI and machine learning solutions on Azure and AWS. Building agentic AI systems with LangChain, LangGraph, AutoGen, and the OpenAI Agents SDK.

01

About

I turn messy enterprise problems into measurable AI outcomes — RAG pipelines, multi-agent systems, VLM document extraction, and the FastAPI plumbing that keeps them alive in production.

BasedPune, India
FocusAgentic AI · RAG · LLM Ops
CloudsAzure · AWS
StackPython · FastAPI · LangGraph
02

Work

Mar 2026 — Present
Remote

Lawstronaut

AI Engineer — Individual Contributor
  • Building a legal document structuring pipeline that slices laws and long-form legal documents into their logical hierarchy — titles, chapters, sections, subsections, clauses — so retrieval and drafting operate on real legal units, not arbitrary fixed-size chunks.
  • Preserving parent–child relationships across the extracted hierarchy so every component stays addressable by its exact position in the source document.
  • Self-hosting open-weight LLMs on GPU with vLLM — continuous batching and paged attention — to run document-scale inference in-house, taking per-token API cost and third-party rate limits out of the pipeline's hot path.
  • Engineering the pipeline for scale: batched GPU inference, async orchestration, and stage-wise processing (segment → summarize → extract keywords) so large documents flow through without blocking.
  • Building a multi-agent quality-control system on LangGraph — specialised agents validate organizational records for completeness, consistency, and rule compliance, with the workflow modelled as an explicit state graph so every verdict is inspectable instead of one opaque LLM call.
PythonvLLMGPU InferenceLangGraphLangChainFastAPIAzure OpenAIAzure AI SearchAzure AI Foundry
Nov 2025 — Feb 2026
Pune, India

Airtel

AI Engineer — Assistant Manager
  • Led AI development for B2B products, embedding practical AI solutions that improved operational efficiency across enterprise teams.
  • Implemented a KYC document extraction pipeline using Vision-Language Models for end-to-end document understanding and structured data extraction.
  • Built AI-driven RPA pipelines automating end-to-end B2B order processing, significantly reducing manual intervention.
PythonVLMsOpenCVRPAAzure AI Services
Aug 2024 — Oct 2025
Pune, India

EY Global Delivery Services

AI Engineer — Associate Analyst
  • Architected a scalable multi-agent AI system using AutoGen — orchestrating autonomous agents for retrieval, semantic analysis, and conversational intelligence in enterprise sales.
  • Developed agentic RAG pipelines integrating Azure OpenAI with hybrid keyword + vector search via Azure AI Search.
  • Shipped FastAPI services on Azure App Service with GitHub Actions CI/CD and a ReactJS frontend.
  • Automated sales intelligence workflows end-to-end, reducing manual reporting effort by ~80% and cutting time-to-insight from hours to minutes.
  • Designed a RAG-based medical policy comparison chatbot, reducing manual review time by 60%.
  • Built a VS Code extension with a TypeScript backend and GPT-4o integration to automate test-case generation, boosting developer productivity by 40%.
FastAPIAutoGenAzure OpenAIAgentic RAGAzure AI SearchCosmosDBReactJS
Mar 2024 — Aug 2024
Remote

FilersKeepers.co

Junior Data Scientist
  • Built an automated classification and mapping system to assign structured labels to legal record descriptions in the LegalTech space.
  • Benchmarked SVM, Random Forest, BERT, DistilBERT, and LLaMA for hierarchical text classification.
  • Deployed a scalable pipeline on Azure ML Studio, reducing manual effort by 70%.
PythonScikit-learnBERTDistilBERTLLaMAAzure ML
03

Projects

AnchorAG

Three retrieval strategies behind one interface — with a verifier that grounds every generated claim.

View Code
  • Built a RAG system where an LLM router picks between naive dense, hybrid (dense + Postgres full-text fused with Reciprocal Rank Fusion, then a bge-reranker-v2-m3 cross-encoder), and agentic multi-hop retrieval planned with LangGraph.
  • Added a deterministic grounding layer that verifies each generated claim against its cited chunk and labels it grounded / unsupported / uncited / synthesized — flagged claims are never presented as fact.
  • Self-hosted bge-m3 embeddings on pgvector (HNSW + GIN full-text in one Postgres), cutting per-token cost to generation only; answers stream over SSE with a live ROUTE → RETRIEVE → GENERATE → VERIFY trace.
  • Benchmarked all three modes on a 50-query golden set (150 runs): hybrid leads context precision (0.72), agentic leads recall (0.96), and citation accuracy holds at 1.000 across modes even after tripling the corpus to 1,201 papers.
  • Cut the latency tail 5.7x (hybrid p95 63.7s → 11.1s) via reranker sequence-length and candidate-pool tuning, off-event-loop compute, model warm-up, and embedding/response caches.
PythonFastAPILangGraphpgvectorbge-m3Cross-Encoder RerankModal A10GSSE

ClauseLens MCP

A remote MCP server that turns any MCP-capable AI into a contract analyst.

View Code
  • Built a production MCP server (FastMCP over Streamable HTTP, deployed on Railway) that Claude, Cursor, or any MCP client connects to with just a URL — no API key, no sign-up.
  • Exposes sharp tools instead of a second brain: an SSRF-hardened fetch_document that pins DNS to defeat rebinding and re-validates every redirect hop, an offset-exact segment_clauses splitter, verify_spans, and a 15-category risk taxonomy plus severity-rubric resources.
  • Anti-hallucination guardrail: the model must prove every quote appears verbatim at exact character offsets before display; anything unverified is dropped.
  • Ships an analyze_contract prompt that injects a 7-step workflow and forces the AI to judge risk from a chosen side (tenant vs. landlord, contractor vs. client) with severity and confidence scores.
  • Zero server-side LLM calls, zero document retention — stateless in-memory processing with IP rate limiting, Pydantic-typed tool I/O, and 44 tests running in CI on Python 3.12 and 3.13.
PythonFastMCPModel Context ProtocolRailwaytrafilaturaPydanticGitHub Actions
04

Stack

Languages
PythonSQL
Frameworks & Libraries
LangChainLangGraphAutoGenOpenAI Agents SDKvLLMPyTorchTensorFlowScikit-learnFastAPI
AI/ML & GenAI
LLMsLLM InferenceTransformersRAGAgentic AIMulti-Agent SystemsMCPNLPDeep LearningPredictive Modeling
Cloud & DevOps
AzureAWSAzure AI ServicesAmazon BedrockDockerGitCI/CD (GitHub Actions)
Databases & Vector Stores
CosmosDBFAISSChromaDBMySQLSQLite
Tools & Utilities
PandasNumPyMatplotlibSeabornOpenCVStreamlitReactJSMCP
certifications
  • Microsoft Certified: Azure AI Engineer Associate (AI-102)
  • Microsoft Certified: Azure Data Scientist Associate (DP-100)
  • Google Data Analytics Professional Certificate
05

Education

2020 — 2024
Savitribai Phule Pune University
B.E. — Artificial Intelligence & Data Science
CGPA 9.24
2018 — 2020
Fergusson College, Pune
HSC — Science
71.38%
05

Say Hi

Got an AI problem
worth solving?

Book a call for AI strategy & consulting engagements — I'm also open to full-time, part-time, and contract work.