Sai Abhiram — AI Engineer

I make complex systems feel simple.

AI-powered applications, automated workflows, and cloud systems that help people get meaningful work done.

Get the highlights.

Selected work.

Six systems, each built around the people who use them — with the architecture open for inspection.

01 · Document AI

Alpha Intelligence — Document Workflow Automation

An intelligent healthcare administrative document-processing and routing workflow — from incoming fax to the right staff queue, with human review built in.

01 / system view · 11 components
Document intake to staff queue, with a confidence gate and human review.
confidentuncertain
  • Incoming fax / upload → Document AI extraction
  • Document AI extraction → Patient matching
  • Patient matching → Confidence & validation check
  • Confidence & validation check → Rules-based routing (confident)
  • Confidence & validation check → Human review (uncertain)
  • Human review → Rules-based routing
  • Rules-based routing → Staff work queue
PythonLLM extractionRules enginePostgreSQLAudit logging

Clinic administrative teams handle a constant stream of incoming documents — referrals, forms, and faxes — that must be read, understood, matched to the right patient record, and routed to the right staff queue. Done manually, it is slow, error-prone, and hard to audit.

  • Unstructured documents become structured, routable information
  • Human review boundaries designed in from the start
  • Every routing decision traceable and auditable
Stack

Python · LLM extraction · Rules engine · PostgreSQL · Audit logging · Human-in-the-loop

Private client system

02 · Full-Stack AI

Enterprise Maintenance Platform

A full-stack platform that digitizes maintenance requests — from submission to technician assignment, completion, and closure.

02 / system view · 9 components
A role-based state machine, from submission to closure.
persistedassignedcompletedclosed
  • Request form + photo → Auth & validation boundary
  • Auth & validation boundary → Admin review (persisted)
  • Admin review → Assign technician
  • Assign technician → Technician update (assigned)
  • Technician update → Admin close (completed)
  • Admin close → Requester notified (closed)
Next.jsTypeScriptSupabasePostgreSQLVercel

Maintenance requests lived in phone calls, emails, and paper notes — no single place to submit, track, assign, or close work, and no visibility into status or workload.

  • One system of record for every maintenance request
  • Role-based workflows replace fragmented communication
  • Status visibility for requesters, admins, and technicians
Stack

Next.js · TypeScript · Supabase · PostgreSQL · Vercel · RBAC

Private client system

03 · Generative AI

AI Guest Concierge & Operations Platform

A guest-facing AI concierge that answers everyday questions, helps guests navigate services and shuttles, and turns requests needing staff attention into organized operations workflows — reducing repetitive coordination and helping the team respond more efficiently.

03 / system view · 13 components
One intelligent guest concierge on the front end; better-organized operations on the back end.
questionrequestops· conceptualshuttle
  • Guest mobile app → Intent & role-aware context
  • Intent & role-aware context → Claude concierge answer (question)
  • Claude concierge answer → Answer to guest
  • Intent & role-aware context → Operations / Facilities queue (request)
  • Operations / Facilities queue → Assignment & handling (ops) [conceptual]
  • Intent & role-aware context → Guest shuttle map / status (shuttle)
React NativeClaude APIFirebase RTDBAWSPostgreSQL

Guests relied on phone calls, staff interruptions, and disconnected information to ask routine questions, locate services, request assistance, or find shuttle information. Operations teams had to manually triage repeated inquiries and coordinate requests across fragmented channels, with limited shared visibility into what was pending and who owned it.

  • Guests self-serve routine questions and shuttle info
  • Staff requests arrive structured, in the right queue
  • Human fallback instead of guessed answers
Stack

React Native · Claude API · Firebase RTDB · AWS · PostgreSQL · Node.js · SendGrid

Private client system

04 · AI Automation

Agentic AI Marketing Platform

A 7-agent LangGraph system running marketing across 5+ ad, CRM, and booking platforms.

04 / system view · 9 components
A supervisor routes work to specialists, behind fact-checks and an approval gate.
routedraftverifiedpublishspend
  • Schedule or operator request → Supervisor agent (LangGraph)
  • Supervisor agent (LangGraph) → Copywriting agent (route)
  • Copywriting agent → Fact-check pass (draft)
  • Fact-check pass → Human approval gate (verified)
  • Human approval gate → Ad, CRM & booking platforms (publish)
  • Ad, CRM & booking platforms → Attribution agent · ROAS (spend)
LangGraphClaude APIPythonFastAPIPostgreSQL

Content creation, campaign management, and attribution were manual and fragmented — with no single view of return on ad spend across channels.

  • 5+ platforms orchestrated
  • Unified Google + Meta ROAS view
  • Human-in-the-loop by default
Stack

LangGraph · Claude API · Python · FastAPI · PostgreSQL · Google & Meta Ads APIs

Private client system

05 · Data Analytics

Enterprise Data Warehouse & AI Analytics

14 disconnected platforms unified into one warehouse with a natural-language query layer.

05 / system view · 10 components
Many sources fan into one governed layer that serves both dashboards and NL→SQL.
dbt
  • CRM platforms → Fivetran / Airbyte → BigQuery
  • Booking & property systems → Fivetran / Airbyte → BigQuery
  • Ads & analytics platforms → Fivetran / Airbyte → BigQuery
  • Fivetran / Airbyte → BigQuery → Governed marts & semantic layer (dbt)
  • Governed marts & semantic layer → Claude NL→SQL
  • Governed marts & semantic layer → Executive dashboards (Looker)
BigQuerydbtFivetranAirbyteClaude API

A multi-property wellness and hospitality group ran 14 disconnected platforms. Answering a basic revenue question meant manual exports and reconciliation — executive reporting took days.

  • 14 platforms unified
  • Reporting: days → minutes
  • 11 KPIs via natural language
Stack

BigQuery · dbt · Fivetran · Airbyte · Claude API · Airflow · SQL

Private client system

06 · AI Automation

ARM — Agent Governance & Observability

A governance layer that puts every AI agent under contract — scope, budget, logging, and human approval.

06 / system view · 9 components
Every agent action passes a contract and risk check before it runs.
out of scopeokriskyallowapproved
  • Agent action request → Contract check · scope & budget
  • Contract check · scope & budget → Blocked · outside contract (out of scope)
  • Contract check · scope & budget → Risk classification (ok)
  • Risk classification → Approval inbox (risky)
  • Risk classification → Tool execution (allow)
  • Approval inbox → Tool execution (approved)
  • Tool execution → Tool-call log & run trace
Claude APILangGraphPythonFastAPIPostgreSQL

As agentic systems multiplied, the client had no clear record of what agents did, what it cost, or who approved it — and no way to run AI safely on top of existing business systems.

  • Full audit trail per agent run
  • Human approval for risky actions
  • AI layered over existing systems
Stack

Claude API · LangGraph · Python · FastAPI · PostgreSQL · AWS

Private client system

07 / Behind the engineering

Don't just see what I built. Explore how it works.

Each walkthrough follows one real request through the system — every component, decision, and safeguard explained. Simulated with fictional demo data.

Alpha Intelligence — Document Workflow Automation

Document AI · Generative AI · AI Automation
Follow a Document
One document. Every branch, decision, and safeguard.
Simulated · fictional data
scenarioA routine referral with clear details. It routes automatically.
  • Incoming fax / upload → Intake validation (received)
  • Intake validation → Quarantine & rejection (invalid)
  • Intake validation → Document AI extraction (valid)
  • Document AI extraction → Patient matching (fields)
  • Patient matching → Confidence & validation check (candidates)
  • Confidence & validation check → Rules-based routing (confident)
  • Confidence & validation check → Human review (uncertain)
  • Human review → Rules-based routing (confirmed)
  • Rules-based routing → Staff work queue (routed)
  • Staff work queue → Audit log
  • Human review → Audit log
  • Human review → Approved corrections [planned]
  • Approved corrections → Rules-based routing [planned]
step 1/8
01@ Incoming fax / upload

Fax received

A referral arrives through the single front door.

payload · after this stepfictional
doc_id
doc_7F3A+new
source
fax+new
status
received+new
event trace
… 7 more steps
03 / Expertise

The technologies behind the systems.

Select a capability to see where it was actually used — no percentages, no ratings, just evidence.

AI & Machine Learning

Model selection, evaluation, and reliability for systems that run in production.

how to read this

Every capability on the left is connected to real work. Select one to see the projects where it was used and how.

04 / Experience

Career timeline.

  1. Oct 2024 — Present
    Full-Time

    AI Engineer

    Blue Ridge Preservation · Boone, NC

    Worked directly with a multi-property wellness client, taking projects from the first whiteboard session to production — and mentored newer engineers on the team.

    • Built ARM, a governance and observability layer that puts every AI agent under a contract for scope and budget, logs every tool call, and traces each run end to end
    • Built a 7-agent LangGraph marketing platform across 5+ ad, CRM, and booking systems, with an attribution agent tying spend to revenue for a single ROAS view
    • Pulled 14 disconnected platforms into one warehouse (Fivetran, Airbyte, BigQuery, dbt) — first view of occupancy vs. marketing spend across 7 business lines
    • Cut executive reporting from days to minutes across 11 KPIs with a Claude-powered natural-language query layer
    • Built an identity layer matching guests across 4 CRM, booking, and analytics sources so numbers reconcile to source
    • Shipped a Claude-powered guest concierge in one sprint with real-time GPS streaming on AWS
    Claude APILangGraphPythonFastAPIAWSPostgreSQLBigQuerydbtFivetranReact Native
  2. Jan 2024 — Jun 2024
    Co-Op

    Data Analyst

    Mellowminds · Business Intelligence & Analytics

    Dashboards and reporting on reconciled data — giving leadership a clear view of where customer value actually lived.

    • CRM analytics: modeled customer value on reconciled data — new vs. returning, high-value segments, repeat rate, lifetime revenue
    • Booking analytics: tracked volume, value, property, status, and period to catch demand shifts and cancellation risk early
    • ERP / financial analytics: recognized revenue, booking-vs-ERP variance, payment reconciliation, outstanding transactions
    • Cross-system analytics: connected CRM, booking, and ERP to find root causes — e.g. tracing a revenue drop to fewer repeat bookings
    SQLPythonCRMERPBooking SystemsDashboards
  3. Jul 2021 — Jun 2022
    Data Engineering

    Data Analyst

    Unified Data Foundation · CRM + Booking + ERP

    Built the reconciled CRM/booking/ERP data layer that every downstream dashboard still runs on.

    • Standardized customer, booking, payment, and revenue fields through a clean, match, reconcile, and model pipeline
    • Made customer and transaction identity reliable across systems — matching records, clearing duplicates, single ID per customer and booking
    • Reconciled booking revenue against ERP and flagged mismatches
    • Kept it clean with automated refreshes and data-quality checks, so finance had a source of truth they could rely on
    SQLPythonETLData ModelingData QualityIdentity Resolution
01 / About

Engineering AI that holds up in production.

Portrait of Saiabhiram Velamakanni
Saiabhiram Velamakanni — Boone, NC — open to remote & willing to relocate

I design and develop AI-powered applications, intelligent workflows, cloud infrastructure, and integrated software systems built to solve real-world challenges.

My work sits where models meet real operations: agent orchestration, retrieval over live data, the warehouse and identity layers underneath, and the governance that makes AI auditable. I take projects from the first whiteboard session through deployment and monitoring.

Open to AI Engineering, Machine Learning, and Generative AI opportunities — remote or on-site, and willing to relocate.

Education
M.S. Computer Information Technology
Purdue University
May 2024
B.E. Information Technology
GRIET, Hyderabad
2018 – 2022
3.5+
Years building production AI systems
6
AI & data projects shipped end to end
14
Platforms unified into one warehouse
11
KPIs on natural-language reporting
Focus areas
  • 01
    Artificial Intelligence & Machine Learning
  • 02
    Generative AI & Large Language Models
  • 03
    Retrieval-Augmented Generation (RAG)
  • 04
    AI Agents & Intelligent Automation
  • 05
    Cloud Architecture & AWS
  • 06
    Data Engineering & System Integration
  • 07
    Security & AI Governance
  • 08
    Full-Stack AI Product Development
05 / Assistant

Ask about my work.

A grounded assistant that answers questions about my experience, projects, and stack using only this portfolio's content.

Portfolio assistantgrounded · gemini
Ask me anything about Saiabhiram's experience, projects, or stack — I answer from his portfolio content.
08 / Contact

Let's build what's next.

I'm interested in engineering opportunities where artificial intelligence, software, and infrastructure come together to solve meaningful problems.

Open to AI Engineering, Machine Learning, and Generative AI opportunities — remote or on-site, and willing to relocate.