Hari Prasath Ulaganathan

Analysts who only build charts create a queue. I try to remove one.

Every point below is something I have already shipped in production; the evidence sits in the case studies further down.

What the tools I built are worth in a year

Time and cost taken out of one department's operation by two tools now running in production.

990 hrs

of manual work removed a year: about 3 hours a day across scheduling and budget entry

$36,250

cost avoided a year: $50/day in scheduling plus $1,500/month in curtailed expenses

30,401

shift records recovered from 85 spreadsheet tabs

181

employees covered by the resulting schedule analysis

Scheduling runs every day, so it is annualized over 365 days; budget entry over a 260-day working year. Daily and monthly figures are conservative baselines; the yearly numbers follow from them.

Five problems, and what changed

Most recent first. Each follows the same shape: the problem someone actually had, what I built, and what was different afterwards.

Grand Colorado on Peak 8  ·  2025–now

Housekeeping scheduling, automated end to end

PythonOR-ToolsStreamlitSupabasePower AutomateSharePointTeams

github.com/hari-94/BGV-Scheduling →

2+ hrssaved daily

$50saved a day

$18,000+a year

0manual steps

41colleagues using it

$0a month to run

Problem

Every morning, two people divided the day's rooms (100+ on a busy day, across three buildings that don't all connect) among 15–30 housekeepers by hand, then typed each assignment into HotSOS, the hotel's housekeeping system, one room at a time. It was slow, one wrong entry meant a missed or double-booked room, and workloads were hard to keep fair.

Engine

A Python engine packs each day's rooms into balanced workloads under the department's hard rules: a time cap per person, one apartment to one housekeeper, never two buildings that don't connect. Rooms are ordered by walking route from the real floor plans, and a Google OR-Tools (CP-SAT) solver checks each building for the fewest people and the tightest fit.

App

A Streamlit and Supabase app with roles for managers, inspectors, housekeepers and housepersons. Each person sees their own rooms on their phone, in English or Spanish, with a done-by time for each room paced to finish by 3:30 PM and early arrivals first. Managers reassign rooms by drag and drop and follow the floor live.

Automated

Power Automate saves the front desk's Arrival Report from email. At 5 AM an agent on the office PC pulls the day's room report and the staff schedule from SharePoint, builds the schedule, writes the day's sheet back to SharePoint and sends every assignment to HotSOS. At 7:30 the notes go to the team's Teams chats. A second PC takes over if the office PC is down.

Forecast

Every two hours the agent reads bookings for today and the next 21 days and turns them into how many housekeepers and inspectors each day needs, split by Full Clean and Daily Service, so a week can be staffed before it arrives.

Reports

A performance report counts rooms cleaned and inspected per person. A timing view shows when each room was started, finished and inspected, and flags late starts and rooms left open.

Outcome

The schedule is ready before anyone arrives, with no manual steps from email to HotSOS. At least 2 staff hours a day come back (about $50 a day), and it runs on free tiers plus tools the department already had. In the last 30 days, 41 colleagues used it: 28 housekeepers, 10 inspectors and 3 managers.

What it solved

• Ready at 5 AM. The schedule builds itself before anyone arrives, and nobody types it into HotSOS.
• Fair workloads. The same rules for every person, every day.
• Less walking. Over the stored days, Daily Service lists kept inside one building went from 112 to 153, and inspector rounds crossing buildings fell from 245 to 192.
• Finishes on time. 95% of room lists are paced to finish by 3:30 PM.
• Staffed ahead. A three-week staffing forecast from live bookings.
• Measured at last. Start, finish and inspection times for every room, per person.
• Easy to use. Rooms on each person's phone, in English or Spanish.

rules engine → team app & logins → 5 AM automation → forecast & reports

How it grew, one manual task at a time: 300+ updates and about 30,000 lines of code in the current version.

Grand Colorado on Peak 8  ·  2025–now

Budget management system

StreamlitSupabasePostgreSQLReportLab

github.com/hari-94/gc8-budget-tracker →

1 hrsaved daily

$1,500expenses cut / mo

$18,000a year

Problem

Departmental budget tracking lived in manual spreadsheets, an hour of data entry a day, where the same expense could be entered twice and nobody noticed until reconciliation.

Built

A full-stack budget application on a relational database with vendor management and database-enforced duplicate-expense prevention, a live dashboard showing month-over-month variance and over-budget alerts, and automated PDF reporting for stakeholders. I owned deployment through GitHub and Streamlit Cloud.

Outcome

Removed an hour of manual Excel entry a day and gave the department visibility that curtailed expenses by at least $1,500 a month. The data now validates itself and the monthly report generates itself.

What it solved

• One source of truth. The whole budget lives in one app instead of scattered spreadsheets.
• No double counting. The same expense can't be entered twice.
• Clear accountability. It shows who changed what, and when.
• No surprises. Spending against budget is visible at a glance, with a warning before it runs over.

Grand Colorado on Peak 8  ·  2025–now

Labor schedule data recovery

PythonopenpyxlETLclustering

30,401records recovered

85sheets consolidated

181employees covered

Problem

Years of housekeeping scheduling history sat trapped in a single multi-year Excel workbook of roughly 85 sheets: unqueryable, and therefore unused.

Built

A Python/openpyxl ETL pipeline that cleaned and transformed the workbook into 30,401 structured shift records across 181 employees, then ran quantitative trend, clustering, travel and workload-balance comparisons across them.

Outcome

Scheduling policy could be argued from measured history rather than recollection.

What it solved

• Years of data made usable. Scattered scheduling sheets turned into one clean dataset.
• Patterns made visible. Who works when, and how evenly the work is shared.
• Better planning. Staffing decisions are now based on facts, not memory.

ThoughtSpot  ·  2022–2023

Analytics pipeline modernization

dbtSnowflakePythonSQL

github.com/hari-94/dbt_TS →

3model layers

51+data tests

Problem

Reporting ran on ad-hoc SQL scripts with no version control and no tests, and recurring performance reports were assembled by hand every cycle.

Built

Pulled four operational sources into Snowflake with Python on Databricks, then modeled them into layered dbt models (staging, intermediate and mart) with validation and data-quality checks embedded in the transformations. Automated the recurring KPI and performance reports in Python against the curated marts.

Outcome

Transformations became version-controlled, documented and testable, and manual report assembly came out of the analytics workflow entirely.

CMA-CGM  ·  2020–2021

Self-service BI enablement

LookerLookMLSnowflakeSQL

Problem

PMO and IT infrastructure teams needed ticketing and resource-allocation numbers on demand, and every request routed through an analyst first.

Built

LookML-defined Explores and metric definitions over curated Snowflake tables, plus a trace of upstream data-quality discrepancies in the source tables feeding them.

Outcome

Teams self-served their own KPI reporting without analyst involvement, and dashboard metrics stayed consistent with source-of-truth records.

Where I have done this

What I work in