Introduc)on

Techathon 6.0
Building tomorrow’s solu8ons for today’s India

Techathon is more than a compe88on—it is a launchpad for India’s brightest young minds to drive
change through bold ideas and breakthrough innova8on. Since its incep8on in 2020, Techathon has
grown into one of the country’s most an8cipated campus contests, inspiring students to push
boundaries and reimagine possibili8es.

Now in its sixth edi8on, this year’s challenge invites par8cipants to explore the fron8er of Agen)c AI,
a transforma8ve leap from automa8on to autonomous, goal-driven systems. The theme sets the
stage for visionary solu8ons that could redefine industries and accelerate India’s digital future.

Stay tuned as the ac8on unfolds!

Problem statements

Challenge I: Pharmaceu)cals

About the business: A leading mul8na8onal generic pharmaceu8cal company, with significant
business in the US and a broad product porKolio, seeks to diversify beyond the highly compe88ve,
low-margin generics market. The company aims to develop value-added, innova8ve products by
repurposing approved molecules for new indica8ons, alterna8ve dosage forms, or different pa8ent
popula8ons—targe8ng unmet medical needs.

Problem statement: Iden8fying such opportuni8es requires extensive literature reviews, oNen taking
two to three months and involving mul8ple itera8ons to uncover viable product concepts. To
accelerate this process, the company plans to adopt an Agen8c AI solu8on that integrates with
various online sources and subscrip8on-based databases.

This AI-driven tool will enable users to interac8vely explore poten8al innova8on cases, significantly
reducing research 8me and increasing throughput. By enhancing the speed and quality of early-stage
product evalua8ons, the company aims to strengthen its pipeline with differen8ated offerings that
deliver greater clinical and commercial value.

Goal: Teams must design an Agen8c AI solu8on where the Master Agent:
• Can be linked to various regulatory websites, clinical trial websites, scien8fic journals and paid
databases (subscrip8ons provided by the client), along with any internal databases of the client
• Features a user interface that allows users to input prompts for finding informa8on from the
web, analyzing market data and summarize scien8fic journals
• Generates a summary report of the searches and save the report in an archival system

Key deliverable: A 5 slider PPT showcasing the end-to-end journey from the ini8al prompt of finding
a molecule, iden8fying its unmet needs, checking for ongoing clinical trials, exploring its probable
use in other diseases, and determining if any patents have been filed, leading to the development of
an innova8ve product story.

Agen)c AI roles
1. Master Agent (conversa8on orchestrator)
 • Interprets user queries and breaks them into modular research tasks
• Delegates tasks to domain-specific Worker Agents
• Synthesizes responses from Worker Agents into coherent summaries with references
• Responds with forma_ed text, tables, charts or PDF reports as needed

2. Worker Agents
a. IQVIA Insights Agent
• Queries IQVIA datasets for sales trends, volume shiNs and therapy area dynamics
• Outputs: Market size tables, CAGR trends, therapy-level compe88on summaries

b. EXIM Trends Agent

• Extracts export-import data for APIs/formula8ons across countries
• Outputs: Trade volume charts, sourcing insights, import dependency tables

c. Patent Landscape Agent

• Searches USPTO and other IP databases for ac8ve patents, expiry 8melines and FTO flags
• Outputs: Patent status tables, compe88ve filing heatmaps, PDF extracts of relevant patents

d. Clinical Trials Agent

• Fetches trial pipeline data from ClinicalTrials.gov or WHO ICTRP
• Outputs: Tables of ac8ve trials, sponsor profiles, trial phase distribu8ons

e. Internal Knowledge Agent

• Retrieves and summarizes internal documents (e.g., MINS, strategy decks, field insights)
• Outputs: Key takeaways, compara8ve tables or downloadable briefing PDFs

f. Web Intelligence Agent

• Performs real-8me web search for guidelines, scien8fic publica8ons, news and pa8ent
forums
• Outputs: Hyperlinked summaries, quota8ons from credible sources, guideline extracts

g. Report Generator Agent

• Formats the synthesized response into a polished PDF or Excel report
• Outputs: PDF summaries with charts/tables, downloadable links in-chat
ClinicalTrials.gov

Data and system assump)ons

• Synthe8c queries: The team will simulate at least 10 strategic ques8ons that pharma planners
might ask (e.g., “Which respiratory diseases show low compe88on but high pa8ent burden in
India?”)
Mock data sources:
• IQVIA mock API: Returns market size, growth and compe8tor data per therapy area
• EXIM mock server: Simulated export or import volumes of APIs or formula8ons
• USPTO API clone: Mock API for patent filings, expiry 8melines and innova8on trends
• Clinical trials API stub: Simulated access to ongoing trials and their sponsors
• Internal documents repository: Synthe8c PDFs of past strategy decks and field reports
• Web search proxy: Simulated web results for real-8me signals and references
• Report generator: Converts chatbot responses into a downloadable PDF summary
• File upload (op8onal): Upload mock internal document PDFs to be summarized by the agent

Agen8c AI Framework (CrewAI/LangGraph)

1. Master Agent: Conversa8on orchestrator
• Understands porKolio planning queries from the user
• Decomposes ques8ons and allocates subtasks to worker agents
• Gathers, summarizes and formats findings into cohesive responses

2. Worker Agents
Agent responsibili8es output type
• IQVIA Insights Agent: Fetches market size, growth, compe8tor data (tables and graphs)
• EXIM Trade Agent: Summarizes import-export trends by molecule (tables, bullet insights)
• Patent Landscape Agent: Lists relevant patent filings, expiry and FTO risks (patent tables, risk
flags)
• Clinical Trials Agent: Extracts trial pipeline informa8on by indica8on or MoA (trial summary
tables)
• Internal Insights Agent: Summarizes uploaded internal PDFs (key takeaways, PDF extracts)
• Web Intelligence Agent: Performs mock web search for guidelines, RWE, news (text responses
with links)
• Report Generator Agent: Converts responses to PDF report (PDF download link in chat)

Submission format: 5 slider PPT

Challenge II: Banking, Financial Services, and Insurance (BFSI) Tata Capital

About the business: A large-scale Non-Banking Financial Company (NBFC) with a presence across
India offers personal loans, home loans, auto loans and more. To increase revenue from exis8ng
customers, the NBFC aims to sell personal loans to prosects and its exis8ng customers through a
web-based chatbot interface.

The chatbot will serve as a digital sales assistant, where a Master Agent (Agen)c AI
Controller) coordinates mul8ple Worker AI agents to handle the end-to-end loan sales process —
from conversa8on and verifica8on to credit evalua8on, and approval and genera8ng a sanc8on
le_er.

Problem statement: The NBFC wants to improve its sales success rate for personal loans by using an
AI-driven conversa8onal approach. The solu8on must simulate a human-like sales process, where
the Master Agent handles customer conversa8ons, engages customers in a personalized manner and
collaborates with mul8ple Worker AI agents to complete the loan process.

Goal: Teams must design an Agen)c AI solu)on where the Master Agent:
1. Chats with customers landing on the web chatbot via digital ads or marke8ng emails
2. Understands the customer's needs and convinces them to take a personal loan
3. Orchestrates mul8ple Worker AI agents to complete all tasks—verifica8on, underwri8ng and
sanc8on le_er genera8on—before closing the chat.

Key deliverable: A 5-slides PPT showcasing the end-to-end journey from the ini8al chat to sanc8on
le_er genera8on.

Agen)c AI roles
1. Master Agent (main orchestrator):
o Manages the conversa8on flow with the customer
 o Hands over tasks to Worker Agents and coordinates the workflow
o Starts and ends the conversa8on
2. Worker Agents:
o Sales Agent: Nego8ates loan terms, discusses customer needs, amount, tenure and
interest rates
o Verifica)on Agent: Confirms KYC details (phone, address) from a dummy CRM
server
o Underwri)ng Agent:
§ Fetches a dummy credit score (out of 900) from a mock credit bureau API
§ Validates eligibility:
§ If the loan amount ≤ pre-approved limit, approve instantly
§ If ≤ 2× pre-approved limit, request a salary slip upload. Approve only
if expected EMI ≤ 50% of salary
§ Reject if > 2× pre-approved limit or credit score < 700
o Sanc)on LeYer Generator: Generates an automated PDF sanc8on le_er if all
condi8ons are met

Data and system assump)ons

• Synthe)c customer data: Teams must create dummy data for at least 10 customers with
details like name, age, city, current loan details, credit score and pre-approved personal loan
limit
• Offer mart server: A mock server or API hos8ng pre-approved loan offers
• CRM server: Dummy customer KYC data
• Credit bureau API: Mock API to fetch credit scores
• File upload: Simulated salary slip upload (dummy PDF or image)
• Teams may make any reasonable assump8ons as long as the solu8on feels realis)c

Submission format
• 5 Slider PPT

Tips for par)cipants

• Keep the chatbot conversa)onal and persuasive, just like a human sales execu8ve
• Focus on orchestra)on logic—how the Master Agent decides which Worker Agent to trigger
• Demonstrate edge cases (e.g., loan rejec8on or the need for addi8onal salary slip
verifica8on)

Challenge III: Automo)ve [Hero + M&M]

Use Case: Autonomous Predictive Maintenance and Proactive Service Scheduling with
Manufacturing Feedback loop

About the Business:

A leading automotive OEM and service network in India provides aftersales maintenance services
to a large customer base across metros and tier-2 cities. The business aims to increase customer
retention, reduce vehicle breakdowns, optimize service center utilization, and improve
manufacturing quality by proactively predicting maintenance needs, autonomously scheduling
service appointments, and feeding insights back to the manufacturing team.
To achieve this, the company plans to deploy a web-based Agentic AI system acting as a Master
Agent orchestrating multiple Worker AI agents to handle end-to-end predictive maintenance,
customer engagement, service scheduling, and manufacturing quality improvement—using real-
time vehicle data, historical maintenance logs, and CAPA/RCA records.

The Problem Statement

Business Problem:

The company wants to improve vehicle uptime, enhance customer experience, and drive
product quality improvements by:

1. Proactively predicting mechanical failures before they occur.

2. Autonomously scheduling service appointments to minimize unplanned downtime.

3. Leveraging RCA/CAPA insights from maintenance and manufacturing logs to improve

design and reduce recurring defects.

Goal:

Design an Agentic AI solution where a Master Agent orchestrates multiple Worker AI agents to
autonomously:

1. Continuously analyze real-time vehicle sensor data and historical maintenance logs
using vehicle telematics

2. Predict upcoming mechanical issues using advanced diagnostics and failure prediction
models.

3. Proactively contact vehicle owners with personalized maintenance recommendations

primarily via voice-based agents, with mobile app notifications as a secondary channel. .

4. Forecast general service demand from maintenance history and vehicle usage patterns
to optimize service center workloads and appointment planning.

5. Manage appointment scheduling by coordinating service center availability and

customer preferences.

6. Track service progress until completion and follow-up for customer feedback.

7. Perform RCA/CAPA-driven analysis by cross-referencing predicted failures with historical

maintenance and manufacturing defect records to suggest preventive actions, best-
practice solutions, and feed insights back to manufacturing teams for quality
improvement.

8. Ensure security and compliance by implementing UEBA (User and Entity Behaviour
Analytics) for Agentic AI to monitor autonomous agent interactions, detect anomalies,
and prevent unauthorized actions during orchestration. (Refer TIPS at the bottom for
UEBA Understanding )

Key Deliverable:

A 5 slides PPT showcasing:

• Continuous vehicle monitoring and predictive failure detection.

 • Forecasting general service demand and autonomous scheduling based on vehicle usage
and maintenance patterns.
• Persuasive customer engagement via voice agent.
• RCA/CAPA-based insights generation and feedback to manufacturing for quality
improvement.
• UEBA in action – detecting abnormal agent behavior or preventing unauthorized access.

Agentic AI Roles

1. Master Agent (Main Orchestrator):

• Monitors vehicle health data streams and overall conversational flow.

• Coordinates Worker Agents in diagnosis, customer outreach, scheduling, and
feedback collection and insight feeding to manufacturing team
• Ensures all agent interactions comply with security policies through UEBA-based
anomaly detection
• Initiates and ends customer interactions.
2. Worker Agents:

• Data Analysis Agent: Continuously analyzes streaming vehicle telematics and

sensor data plus maintenance history to detect early warning signs and forecast
likely maintenance needs or service demand

• Diagnosis Agent: Runs predictive models to assess probable component failures

and assigns priority levels.

• Customer Engagement Agent: Initiates personalized conversations with vehicle

owners via chatbot to explain predicted issues and recommend service.

• Scheduling Agent: Checks service center capacity, proposes appointment slots,

and confirms bookings with customers.

• Feedback Agent: Follows up post-service to capture customer satisfaction and

update vehicle maintenance records.

• Manufacturing Quality Insights Module : Automatically generates actionable

insights for the manufacturing team by analyzing predicted failures and historical
CAPA/RCA data to improve product design and reduce recurring defects.

Data & System Assumptions

• Synthetic Vehicle Data: Data for 10 example vehicles including sensor readings, usage
patterns, maintenance history, and diagnostic trouble codes.

• Telematics API: Mock real-time sensor data feed.

• Maintenance Records Server: Dummy database of historical repairs and service visits
(can leverage open-source automotive datasets from Kaggle, UCI Repository,
HuggingFace, etc.).

• Service Center Scheduler: Mock API to retrieve available appointment slots and confirm
bookings.
 • Customer Interaction Layer: Simulated voice-based virtual agent as the primary interface
for owner communication, supplemented by app notifications for reminders and
confirmations

• Security Layer: UEBA integrated to monitor Master and Worker Agents for anomalous or
malicious behaviours (e.g., unauthorized API calls or unexpected workflow changes).

• Forecasting & RCA Data: Historical maintenance, manufacturing CAPA, and RCA records.

Submission Format: A 5 slider PPT

Tips for Participants

• Emphasize persuasive and human-like chatbot/Voice agent conversations explaining the

vehicle's condition and convincing owners to book services.

• Showcase how the Master Agent coordinates real-time data analysis, predictive
modeling, service demand forecasting, customer engagement, and appointment
scheduling seamlessly, while integrating manufacturing feedback loops.

• Demonstrate edge cases like declined appointments, urgent failure alerts, or multi-
vehicle fleet scheduling, and recurring defects. Show how RCA/CAPA analysis informs
better decision-making in these cases.

• What is UEBA - UEBA uses advanced analytics and machine learning to establish
behavioural baselines for users and entities (including AI agents) and detect anomalies
that indicate potential threats or unauthorized activities.

Example of UEBA: “In the predictive maintenance system, UEBA can monitor the
Master Agent and Worker Agents for unusual activities—for instance, if the
Scheduling Agent suddenly tries to access vehicle telematics data (which it
normally doesn’t need), UEBA will flag this as anomalous behaviour and trigger an
alert.”

• Show how predicted failures and RCA/CAPA patterns are fed back to manufacturing
teams to improve product design, reduce recurring defects, and close the loop between
aftersales and production.

Challenge IV: FMCG [Asian paints]

Executive Summary

This challenge focuses on optimizing the B2B RFP response process for a leading industrial
manufacturer using Agentic AI. Participants must design a multi-agent system that automates RFP
identification, product matching, and pricing estimation to improve response speed and accuracy.
‘About the Business

A large Industrial products manufacturing firm (client) with a business across Fast moving electrical
goods (FMEG), Wires and Cables has seen considerable growth over the last 5 years. The growth has
been driven by large Business to business (B2B) Request for proposals (RFPs) and tenders from
infrastructure build out in India. The client wishes to further develop the B2B business by increasing
the number of RFP qualifications and responses.

Context on B2B RFPs

Typically, large projects – e.g., government projects are launched by government departments or
public sector undertakings (PSUs). These projects are executed by lumpsum turnkey (LSTK) project
executors (PSUs or private enterprises). The LSTK project executors in raise RFPs for material
supplies – e.g., wires, cables, other materials as RFPs to be supplied by Original Equipment
Manufacturers (OEM) vendors.

A typical RFP consists of technical scope of supply (quantity of wires / cables), technical specification
(for wires and cables), test requirements and acceptance tests to be conducted before accepting the
delivery of wires and cables (refer sample RFP). OEM vendors bid for these tenders and lowest
priced tender is awarded the contract. The client is one of the large wires and cables OEM in India,
who regularly bid for such RFPs.

B2B RFP response process

• Once the RFP is received to the client Sales Team, they qualify the RFP based on the date for
submission, past experience and product coverage.
• Qualified RFPs are passed on to the Product Technical Team to match the client product
SKUs to RFP product requirements. Refer sample of product specs.
• The Product Technical Team finalizes the client product SKUs, which are the closest
matches.
• This is then worked on by the Pricing Team, which estimates the price of the products and
additional costs required to meet the testing and acceptance test requirements at project
site.
• The Sales Team consolidates the inputs from Product Technical Team and Pricing Team to
submit the RFP response to the LSTK project executor.
• In case the client product SKUs has a low match for RFP product requirement, an internal
request for new made to order SKUs is prepared to meet the RFP product requirements.

Bottlenecks in the B2B RFP Response

• The entire RFP response process is manually driven – with manual handoff between each
team.
• In many cases the Sales Team may not be aware of the release of RFPs by LSTK project
executors on time. These are typically released on LSTK project executor’s website, received
on emails. Delay in initiating response to RFPs would certainly mean lower chances of
winning.
• The technical matching of RFP requirements to Product SKUs is manually done. This requires
knowledge of technical standards, specifications, and product knowledge.
• While standards, specifications and product details are well documented, the technical team
is often concurrently managing multiple RFPs and takes a lot of lead time to turn around
 their response. The Pricing Team can work on the price only after technical team’s
responses.

The Problem Statement

Business Problem:

The goal is to scale RFP responses using AI.

The client wants to further drive of the B2B business channel. Given that their past growth has been
on B2B RFPs, they have now significant credentials and see their team having the “right to win” in
this segment. Given the rapid growth in the past few years, they are now facing bottlenecks in RFP
response process which are impacting their ability to respond to RFPs on time.

Upon analyzing the past wins and losses on large RFPs, the client has identified the following.

- 90% of the wins correlated to RFPs which were received and actioned on time.
- 60% of the wins correlated to RFPs where adequate time was provided for the technical
team to match product requirements with RFPs.
- Technical product SKU matching RFP requirements takes the most time in the process.
- Delays in RFP submission significantly reduced the chances of winning.

The client wants to improve the number of RFPs responses (per year) and the timely response by
using an Agentic AI approach. The solution must simulate the B2B RFP response process through a
Sales agent and Technical Agent which automates the process of identifying an RFP on time and
maps the right SKUs with the RFP.

Goal:
Teams must design an Agentic AI solution where the Sales Agent:

1. Scans a set of predefined URLs to identify the RFPs which are due to be submitted in the
next 3 months.
2. Summarizes the requirement of the RFP in terms of products to be shared with Technical
Agent
3. Summarizes the testing and acceptance requirement to be shared with the Pricing Agent

Key Deliverable:
A) showcasing the end-to-end journey from RFP identification to collating the RFP response from
technical and pricing agents.

Agentic AI Roles

1. Main Agent (Main Orchestrator)

• Prepares a summary of an identified RFP to be shared with Technical and Pricing Agent
• Summary shared with technical and pricing agent needs to be contextual to their role.
• Receives the response from the Technical and pricing agent to consolidate the overall
response of the RFP.
• The overall response of the RFP needs to contain the OEM product SKUs suggested, their
price and the costs for tests required in the RFP.
• Starts and ends the conversation.

2. Worker Agents:
o Sales Agent:
§ Identifies the RFPs which are due for submission in the next 3 months.
§ Scans identified web URLs to summarize RFPs with their due dates.
§ Identifies 1 RFP to be selected for response and sends this to the Main
agent.
o Technical Agent:
§ Receives the summary of RFP and RFP document from the main agent.
§ Summarizes the products in Scope of supply.
§ Recommends top 3 OEM products which match each of the product in
Scope of supply, show a “Spec match” metric (in %) for each of OEM product
recommendations.
§ Recommendation of top 3 OEM products come from a repository of product
datasheets. Refer sample of product specs.
§ The spec match metric should reflect the closeness with which the
recommended OEM product matches the RFP product specs considering
that all the required specs have an equal weightage.
§ Prepares a comparison table of RFP spec parameters requirements, spec
values for Top 1,2 and 3 OEM product recommendations for each RFP
product.
§ Select the top OEM product which closely matches the RFP products for all
items in scope of supply based on the spec match metric.
§ Final table of products in Scope of Supply and recommended OEM product
SKUs
§ Send this table to main agent and the pricing agent.
o Pricing Agent:
§ Receives the summary of the tests and acceptance tests to be done for the
product from the main agent.
§ Receives the product recommendation table from the technical agent.
§ Pricing agent assigns a unit price for each product based out of a dummy
pricing table and price for each test based out of a dummy services price
table.
§ Consolidates the total material price and services price for every product in
scope of supply.
§ Send the consolidated price table to Main agent.
Data & System Assumptions

• RFP Data – Teams can download the RFPs from internet. 2 sample RFPs have been provided
as a reference.
• OEM product data sheets – Teams can download the cables datasheets for any of the
leading cable manufacturer websites – a few sample URLs have been provided as a
reference.
• Product and Tests pricing data:
Teams must create synthetic data for price of products, type of tests and the price of each of
the tests.
• Scanning RFP from websites – The teams can stand up sample web pages which show the
RFPs for the sales agent to scan and summarize.

Submission Format: 5 slider PPT

Tips for Participants

• Provide outputs in structured summary (bullets rather than long paragraphs) and structured
tables.
• Focus on the logic for matching RFP requirements with OEM product recommendations.
• Detail out the logic for scoring Spec match metric for RFP products and OEM products.
• Emphasize more on the process of matching requirements rather than accuracy of matching.
Challenge V: Retail [ABFRL]

About the business: A leading retail brand seeks to revolu8onize its sales experience by deploying an
AI-driven Conversa8onal Sales Agent that seamlessly operates across online and physical channels.
This Agent must emulate a top-8er sales associate—guided by natural, personalized dialogue—while
orchestra8ng specialized Worker Agents to handle tasks from product discovery through checkout
and post-purchase support.

Problem statement: Customers face fragmented experiences when moving between online
browsing, mobile app shopping, messaging apps, and in-store interac8ons. Limited bandwidth
among sales associates leads to missed up-sell and cross-sell opportuni8es. The goal is to increase
Average Order Value (AOV) and conversion rates by offering a unified, human-like conversa8onal
journey that an8cipates needs, provides tailored recommenda8ons and facilitates sales across all
channels.

Goal: The team should design an Agen8c AI solu8on where the Sales Agent:

1. Engages customers via web chat, mobile app, WhatsApp/Telegram, in-store kiosk or voice
assistant

2. Understands preferences and context (e.g., past purchases, store loca8on, current
promo8ons) and guides customers toward relevant products

3. Coordinates mul8ple Worker Agents to handle inventory checks, personalized

recommenda8ons, promo8ons lookup, payment processing and order confirma8on—
culmina8ng in a seamless purchase or booking (e.g., “reserve in-store for try-on”)

Key deliverable: 5 slider PPT

A customer ini8a8ng a conversa8on on one channel (e.g., mobile app)

• The Sales Agent adap8ng the conversa8on when the customer switches to an in-store kiosk
or messaging app

• End-to-end orchestra8on: recommenda8on → inventory check → payment →

digital/physical fulfillment → post-purchase follow-up

Agen)c AI roles

Role Responsibili)es

Sales Agent • Manages mul8-channel conversa8on flow

• Routes tasks to Worker Agents
• Handles context switching and session con8nuity

Worker Agents
 • Recommenda)on Agent • Analyzes customer profile, browsing history and seasonal
trends
• Suggests products, bundles and promo8ons

• Inventory Agent • Checks real-8me stock across warehouses and stores

• Offers “ship to home,” “click & collect,” or “in-store availability”

• Payment Agent • Processes payments via saved cards, UPI, giN cards or in-store
POS
• Manages payment failures and retries

• Fulfillment Agent • Schedules delivery or reserve-in-store slots

• No8fies logis8cs or store staff for pickup orders

• Loyalty and Offers Agent • Applies loyalty points, coupon codes and personalized offers
• Calculates final pricing and displays savings

• Post-Purchase Support • Handles returns/exchanges, tracks shipments and solicits

Agent feedback

Data and system assump)ons

• Synthe)c customer profiles: ≥ 10 customers with demographics, purchase history, loyalty

8er, device preferences

• Product catalog API: Mock endpoint with SKUs, categories, a_ributes, pricing, images

• Inventory server: Simulated real-8me stock levels for online and mul8ple store loca8ons

• Payment gateway stub: Dummy API for authoriza8ons, captures and declined transac8ons

• Loyalty and promo)ons service: Mock rules engine for loyalty points and 8med promo8ons

• POS integra)on: Simulated in-store terminal interac8ons for barcode scan and payment

Submission format: 5 slider PPT

Tips for par)cipants

• Omnichannel consistency: Maintain session con8nuity when moving between channels

(e.g., chat → in-store kiosk)

• Sales psychology: Use persuasive, consulta8ve language—ask open ques8ons (“What

occasion are you shopping for?”), suggest complementary items (“These shoes pair well
with…”) and handle objec8ons gracefully

• Edge-case demonstra)ons: Show how the Agent recovers from scenarios like payment
failures, out-of-stock items or a request to modify an order

• Modular orchestra)on: Keep Worker Agents loosely coupled so new capabili8es (e.g., GiN-
wrapping Agent) can be added easily
Challenge VI: IT/BPM [Firstsource]

Provider Data Valida)on and Directory Management Agent for Healthcare Payers

Business Context
Healthcare payers struggle with maintaining accurate provider directories, with studies showing
80%+ of provider entries contain errors like incorrect addresses, phone numbers, professional
details, license details. Manual valida8on processes are 8me-intensive, requiring staff to call
providers, verify creden8als, and update mul8ple systems. This creates frustra8on among members
when they can't reach providers, regulatory compliance risks, and wasted opera8onal resources. A
simplified AI solu8on focused on automa8ng basic provider data valida8on and directory updates
can demonstrate significant value while being feasible for hackathon development using publicly
available data sources.

1) Business Problem - Current Challenges, Desired Outcomes

Current Challenges:
• Provider directories contain 40-80% inaccurate contact informa8on causing member
frustra8on and access issues
• Manual verifica8on processes requiring staff to call hundreds of providers monthly for basic
updates
• Mul8ple data entry points crea8ng inconsistencies between online directories, mobile apps,
and printed materials
• Regulatory requirements demanding frequent provider data updates with limited automa8on
capabili8es
• Time-consuming creden8al verifica8on processes delaying provider network addi8ons by
weeks or months
• Member complaints about outdated provider informa8on leading to unsuccessful
appointment a_empts
Desired Outcomes:
• Automate provider data valida8on through intelligent web scraping and API calls
• Reduce manual verifica8on 8me through AI assistance
• Achieve target provider contact informa8on accuracy through con8nuous automated
valida8on
• Create unified provider data management reducing inconsistencies across member-facing
plaKorms
• Demonstrate reduc8on in provider directory maintenance costs through intelligent automa8on

2) Goal
Develop a simplified Agen8c AI system that automates basic provider data valida8on using publicly
available sources, demonstrates intelligent data quality improvement, and showcases the poten8al
for full-scale provider data management automa8on with synthe8c and public data sources.
Provider data includes
• Demographics: Name, contact informa8on
• Professional Details: Special8es, licenses, cer8fica8ons
• Network Affilia8ons: Insurance networks, affilia8ons with other providers or groups
• Services Offered: Clinical focus, appointment availability
• Loca8on and Facili8es: Addresses, medical imaging facili8es

3) Key Deliverable
Demonstra)on Scenario: Automated valida8on and upda8ng of 200 provider profiles using publicly
available data sources
• Input: Sample provider dataset with names, addresses, phone numbers, special8es, and basic
creden8al informa8on. Must include scanned pdf (unstructured data) formats.
• Process: AI agent automa8cally validates contact informa8on via web scraping, checks
creden8als against public databases, iden8fies inconsistencies, and flags providers needing
manual review
• Output: User interface to show updated provider profiles with confidence scores, ac8ons
status reports, priori8zed list of providers requiring human a_en8on, and generate
communica8on email.
• Timeline: Complete valida8on cycle in under 30 minutes versus tradi8onal manual work

4) Agen)c AI Roles (Suggested / Illustra)ve)

Data Valida)on Agent:
• Performs automated web scraping of provider prac8ce websites to verify current contact
informa8on and services
• Cross-references provider informa8on against public databases including NPI registry and state
licensing boards
• Conducts intelligent phone number and address valida8on using publicly available verifica8on
services
• Generates confidence scores for each data element based on source reliability and cross-
valida8on results

Informa)on Enrichment Agent:

• Searches public sources for addi8onal provider informa8on including educa8on, board
cer8fica8ons, and special8es
• Analyzes provider websites and online profiles for updated prac8ce informa8on and service
offerings
• Iden8fies poten8al network gaps by analyzing geographic distribu8on and specialty coverage
• Creates standardized provider profiles with enriched data from mul8ple public sources

Quality Assurance Agent:

• Compares provider informa8on across mul8ple sources to iden8fy discrepancies and
inconsistencies
• Flags providers with suspicious or poten8ally fraudulent informa8on for manual review
 • Tracks data quality metrics and generates reports on valida8on success rates and common
error pa_erns
• Priori8zes providers for manual verifica8on based on member impact and data confidence
levels

Directory Management Agent:

• Generates updated provider directory entries in mul8ple formats (web, mobile app, PDF)
• Creates automated alerts for providers requiring immediate a_en8on or manual verifica8on
• Produces summary reports showing valida8on results, data quality improvements, and
recommended ac8ons
• Manages workflow queues for human reviewers with priori8zed tasks and suppor8ng
documenta8on

5) Data & System Assump)ons

Publicly Available Data Sources:

• NPI Registry (CMS): Free API access for provider basic informa8on, creden8als, and prac8ce
loca8ons
• State Medical Board Websites: Public license verifica8on and disciplinary ac8on informa8on
(can be scraped)
• Hospital/Health System Websites: Provider directory pages with current prac8ce informa8on
and contact details
• Google My Business/Maps API: Prac8ce loca8on verifica8on, phone numbers, and pa8ent
review data
• Medicare Provider U)liza)on Database: Public claims data showing provider special8es and
prac8ce pa_erns

Synthe)c Data Genera)on:

• Provider Profile Generator: Create realis8c provider datasets with names, addresses,
special8es, and creden8al informa8on
• Valida)on Scenario Creator: Generate common data quality issues like outdated phone
numbers, moved prac8ces, and creden8al changes
• Member Impact Simulator: Create synthe8c member complaint data related to provider
directory accuracy
• Network Coverage Generator: Generate geographic and specialty distribu8on data for
network adequacy analysis

6) Submission Format: 5 Slider PPT

7) Tips for Par)cipants

• Guardrails First: Enforce content modera8on/PII redac8on and grounded (SOP/KB) before
enabling sensi8ve ac8ons (rejec8on).
• Fast Wins: Target top 5-8 high-volume intents/proper8es for AHT reduc8on.
 • Edge-Case Rigor: Demonstrate behavior for ambiguous/handwri_en content, fuzzy match,
missing data, and source failure scenarios.
• Modular Orchestra)on: Keep Worker Agents loosely coupled so new programs (clients) and
tasks can be added safely.

8) Target KPIs (for pilot)

• Valida)on Accuracy: 80%+ success rate in iden8fying outdated provider contact informa8on
• Processing Speed: Complete valida8on of 100 providers in under 5 minutes versus hours of
manual effort
• Informa)on extrac)on: Achieve 85%+ accuracy during informa8on extrac8on from
unstructured documents/scanned PDFs with 95% right confidence score
• Processing Throughput: Handle 500+ provider valida8ons per hour through automated
pipeline

9) Example Flows to Implement (Pick 2–3)

Flow 1: Automated Provider Contact Informa)on Valida)on

Trigger: Daily batch processing of 200 provider profiles for contact informa8on accuracy
Process Steps:
1. Data Valida8on Agent extracts provider prac8ce informa8on from synthe8c dataset including
names, addresses, and phone numbers
2. Agent performs web scraping of provider prac8ce websites and Google My Business lis8ngs to
verify current contact informa8on
3. Cross-valida8on against NPI registry API to confirm provider iden8fica8on and basic prac8ce
details
4. Quality Assurance Agent compares informa8on across sources and generates confidence
scores for each data element
5. Directory Management Agent creates valida8on report showing confirmed updates,
discrepancies, and providers needing manual review
6. Automated priori8za8on of providers for human verifica8on based on member impact and
data confidence levels

Flow 2: New Provider Creden)al Verifica)on and Onboarding

Trigger: 25 new providers applying for network inclusion with basic creden8al documenta8on
Process Steps:
1. Informa8on Enrichment Agent extracts provider informa8on from applica8on forms and
searches NPI registry for verifica8on
2. Automated lookup of provider licenses through state medical board websites and public
creden8al databases
3. Data Valida8on Agent performs background research on provider educa8on, board
cer8fica8ons, and prac8ce history
4. Quality Assurance Agent cross-references informa8on across mul8ple sources and iden8fies
any red flags or inconsistencies
5. Automated genera8on of provider profiles with enriched informa8on and confidence ra8ngs
for creden8aling decisions
 6. Directory Management Agent creates summary reports for creden8aling commi_ee with
recommenda8ons and suppor8ng documenta8on

Flow 3: Provider Directory Quality Assessment and Improvement

Trigger: Weekly quality assessment of en8re provider directory database (500 providers) for accuracy
and completeness
Process Steps:
1. Quality Assurance Agent analyzes all provider profiles to iden8fy missing informa8on, outdated
data, and poten8al inconsistencies
2. Data Valida8on Agent performs selec8ve verifica8on of providers iden8fied as high-risk for
accuracy issues
3. Informa8on Enrichment Agent a_empts to fill data gaps through public source research and
web scraping
4. Automated genera8on of data quality metrics showing improvement trends and areas needing
a_en8on
5. Directory Management Agent creates priori8zed ac8on lists for staff including specific
providers and data elements requiring manual update
6. Produc8on of execu8ve dashboard showing overall directory quality scores and improvement
recommenda8ons

11) Minimal Tech Stack (Reference/Illustra)ve)

• Programming Language: Python with libraries including Beau8fulSoup, Requests, Pandas, and
Scikit-learn for web scraping and data processing
• AI/ML Framework: OpenAI API (free 8er) or Hugging Face Transformers for natural language
processing and data matching
• Data extrac)on: VLM based data extrac8on for maximum performance
• Database: SQLite or PostgreSQL for local provider data storage and valida8on tracking
• Web Framework: Flask or FastAPI for crea8ng simple dashboard and API endpoints
• NPI Registry API: Free CMS API for provider verifica8on and basic informa8on lookup
• Web Scraping Tools: Selenium or Beau8fulSoup for automated provider website data
extrac8on
• Google Maps API: Free 8er for loca8on verifica8on and contact informa8on valida8on
• State License APIs: Public APIs where available or web scraping for license verifica8on
• Task Scheduling: Python Cron jobs or Celery for automated valida8on cycles
• Data Processing: Pandas and NumPy for data manipula8on and quality analysis
• Confidence Scoring: Custom algorithms using data source reliability and cross-valida8on
results
• Report Genera)on: Python libraries like ReportLab or Matplotlib for crea8ng valida8on reports
and dashboards