I'm Anudeep Sri Bathina, an AI Architect leading agentic AI platforms, multimodal RAG, and governed data systems, from architecture through production operations.
11+ years
AI, data & platform engineering
16 engineers
Architecture + delivery leadership
1,500+
Developers served by public GenAI
Agentic AI · multimodal RAG · governed lakehouse · observability
Owns architecture and delivery for AT&T’s $4M competitive-intelligence platform, leading a 16-person team across agentic AI, multimodal RAG, governed data, and production operations for pricing, strategy, and sales teams.
2026–Present
Public GenAI · RAG · MCP · developer platform
Launched AT&T’s first public-facing GenAI product for 1,500+ API developers. The LangGraph and RAG platform cut integration time 80%, resolved 99% of queries at sub-2-second latency, sustained 99% uptime, and earned the RISE Award.
2024–2026
Azure ML platform · team leadership · MLOps
Led an 11-person team delivering a Databricks and Spark ranking platform for Unilever UK; raised Precision@10 from 62% to 85%, reduced stockouts 18%, and cut pipeline runtime 40%.
2019–2021
Distributed ML · product engineering · containers
Independently built and scaled Strait, a proprietary Big Data and ML platform using PySpark, Dask, predictive models, and containerized delivery.
2019–2019
Cloud migration · data engineering · analytics
Migrated 300GB of airline data to Azure SQL, improved SLA compliance 18%, reduced manual pipeline effort 60%, and tripled query performance for 200GB of healthcare data.
2015–2019
Selected production outcomes
From orchestration and tool boundaries to evaluation, observability, and production delivery. Each capability is tied to working evidence.
LangGraph- and Fleet-based agents with explicit state, tool boundaries, deep research, and human review.
Agent tracing, groundedness checks, responsible-AI guardrails, privacy boundaries, and explicit failure behavior.
Production service boundaries, model routing, typed APIs, identity, containers, and scalable delivery patterns.
Governed lakehouses, distributed processing, data quality, and compliant data products that feed production AI.
Selected credentials
Research sharpens the evaluation practice. Teaching makes AI practical, while career mentorship helps people move forward.
M.S. Computer Science
University of Massachusetts Dartmouth · 2023
B.Tech Electrical & Electronics Engineering
VIT University · 2015
RISE Award
AT&T · production GenAI delivery
Extra Mile
Capgemini · data and ML delivery
4.76 / 5
UT Austin · learner rating
Mentor testimonials
“Anudeep is incredibly insightful, listening carefully and offering technical yet straightforward comments that are truly beneficial.”
“Extremely insightful and valuable discussion. Anudeep's depth of knowledge in Data and AI is evident, and his willingness to openly share his expertise is commendable.”
“His tailored advice on skills, job applications, and interviews was practical and insightful, leaving me confident and motivated.”
“His strategic guidance and ability to simplify complex AI and career paths into clear, actionable steps were incredibly helpful.”
“An exceptional session, making complex Agentic AI concepts easy to understand. His motivating approach inspired me to take bold steps.”
Three complementary examples show how I frame boundaries, make consequential decisions, and leave evidence behind for the next reviewer.
Applied AI / retrieval
Hospital staff need fast policy answers without crossing department boundaries or presenting unsupported guidance as fact.
Real ClinIQ product screenshot from the public repository; the interface uses synthetic demonstration data.
My contribution
Designed and built the reference implementation across the product, API, retrieval graph, safety controls, and evaluation fixtures.
Use case
Nurses, administrators, and compliance reviewers working with synthetic policy data.
Evidence-backed outcome
The public repository includes an authenticated workspace, FastAPI service, LangGraph retrieval flow, synthetic evaluation corpus, RBAC matrix, and PHI-masking tests.
Constraints
Decision / consequence
Use department-scoped retrieval indexes instead of one shared knowledge pool.
The boundary is easier to reason about and test, but it adds index-management and routing work.
AI agents / developer platform
Code migrations require coordinated analysis, planning, and validation without letting autonomous changes outrun human review.
My contribution
Architected and built a Fleet-powered agent platform for codebase analysis and migrations, with CLI, API, MCP surfaces, checkpoints, and sensitive-pattern scanning.
Use case
Engineering teams assessing and applying code migrations.
Evidence-backed outcome
The repository uses Fleet to coordinate the bounded workflow across CLI, FastAPI, MCP tools, and UI surfaces. Generated migrations still require human review.
Constraints
Decision / consequence
Use Fleet to coordinate bounded agents while keeping dry-run as the default and requiring an explicit apply step backed by a rollback checkpoint.
Agent handoffs and human review add latency, but make each proposed change traceable, inspectable, and recoverable.
Computer vision / evaluation
Recognizing groundfish across image collections is difficult when underwater datasets are limited and visually inconsistent.
My contribution
Graduate researcher and thesis author; developed and evaluated a cross-database recognition pipeline.
Use case
Fisheries and computer-vision researchers exploring automated species recognition.
Evidence-backed outcome
The UMass Dartmouth research reports 94.10% mAP and 92.14% classification accuracy on its evaluated workload. These are thesis results, not production benchmarks.
Constraints
Decision / consequence
Separate localization with YOLOv8 from classification with ResNet-50.
Two stages can be evaluated independently, but errors from detection can propagate into classification.
Additional systems
Model choice is one decision. Scope, evaluation, failure behavior, and operating boundaries determine whether the whole system can be trusted.
Architecture
Scope retrieval, identity, and data movement before adding more model capability.
Evidence: ClinIQ
Evaluation
Test unsupported questions, access rules, masking, and groundedness, not only happy-path answers.
Evidence: ClinIQ evaluation pack
Delivery
Use bounded Fleet agents, dry-runs, checkpoints, and human review when automation can alter code or policy.
Evidence: ShiftIQ
Observability
Track latency, cost, answer quality, and agent behavior as production signals, not afterthoughts.
Evidence: AT&T + InferIQ
System walkthrough · ClinIQ
01 / Ask
A staff member asks a policy question. The interface can request clarification before retrieval when patient or department context is ambiguous.
Decision: separate department indexes make the access boundary explicit, at the cost of more index management.