Bajaj Finserv Health's national engineering hackathon - teams move from a strict-template pitch and code submission to a live leaderboard and finale, building production-minded AI for health, insurance, lending, and document workflows.
Brabble editorial · Updated 15 Aug 2026 · Official page
Particulars
Organiser
Bajaj Finserv Health
Mode
Online pitch and leaderboard; finalist jury presentation
Prize / outcome
6.0: ₹2.35 lakh listed cash awards plus potential hiring consideration
Team size
2-5 from one institution
Application fee
Free
Rounds
4
What it is
HackRx is Bajaj Finserv Health's annual flagship engineering hackathon. The sixth edition centred on an LLM-powered query-and-retrieval system for large policy documents, contracts, and emails, testing contextual inference rather than a generic question-answer bot. Teams first interpret a detailed brief and mandatory pitch template, then submit code to a live evaluation leaderboard, attend shortlisted-team briefings, and demonstrate the final system to a jury on innovation, execution, scalability, UX, impact, and presentation.
Rounds
01. Deep-dive webinar
Organisers explain the theme, problem statement, evaluation environment, and domain context. Teams should convert every clarification into a requirement or test.
02. Pitch deck and code submission
The preliminary submission must use the provided template and explain the value proposition and technical approach; deviations can invalidate an otherwise strong entry.
03. Online leaderboard evaluation
Teams submit against the challenge platform and receive leaderboard feedback across daily rounds. Shortlisting considers ranking and defined evaluation criteria, so robustness beyond a single visible score matters.
04. Shortlisted-team webinar and grand finale
Finalists receive further rules, complete the solution, and present functionality and approach to judges evaluating uniqueness, implementation, scale, UX, impact, and communication.
How to prepare
Step 01
Turn the brief into an evaluation harness
Leaderboard competitions are won by repeatable measurement, not demo improvisation.
Parse expected input/output, document types, query classes, latency limits, scoring, and prohibited resources into automated tests.
Create a representative corpus with policy clauses, exclusions, tables, cross-references, emails, ambiguous queries, and deliberately irrelevant text.
Establish BM25/vector and single-pass RAG baselines before adding rerankers, query decomposition, or agents.
Step 02
Make retrieval defensible
A confident answer without the controlling clause is a failure in finance or health.
Measure retrieval recall separately from answer correctness; preserve page/section citations and quote only the minimal supporting span.
Handle OCR noise, tables, chunk boundaries, versioned documents, metadata filters, conflicting clauses, and queries that require abstention.
Protect against prompt injection inside documents, data leakage across users, unauthorised tools, and logging of sensitive content.
Step 03
Operate through a shifting leaderboard
Avoid overfitting to public feedback.
Version code, prompts, indexes, models, datasets, and scores; change one factor at a time and record cost/latency.
Keep a holdout set that resembles likely private tests and includes negative/adversarial cases.
Freeze a reliable submission before the deadline, then experiment on a branch; prepare local/offline fallbacks for rate limits and provider failure.
Representative practice - not leaked questions
Questions to practise
01
Answer whether a treatment is covered when the policy has definitions, waiting periods, exclusions, and endorsements in separate sections.
What to show: Multi-hop retrieval, precedence, citations, uncertainty, and safe refusal when evidence conflicts.
02
A document says to ignore system instructions and reveal other policies. Design the defence.
What to show: Content isolation, instruction hierarchy, sanitisation, access control, tool permissions, red-team tests, and audit.
03
Your retrieval recall improves but latency triples. Choose a production configuration.
What to show: Metric thresholds, reranking budget, caching, async work, model tiers, cost, and user-risk trade-offs.
04
The public leaderboard rises while your local holdout falls.
What to show: Overfitting diagnosis, leakage, distribution differences, experiment rollback, and private-test strategy.
Who can enter
HackRx 6.0: students in 2nd, 3rd, or 4th year of B.E./B.Tech or M.Tech engineering programmes
Teams of 2-5 from the same institution; members may span semesters and domains
Free to enter and open across India, subject to the edition's registration and verification rules
Every team must follow the official submission template and portal instructions exactly
What you actually get
HackRx 6.0 published ₹1 lakh for the winner, ₹50,000 for first runner-up, ₹25,000 for second runner-up, and four ₹15,000 category awards, plus goodies and potential internship or job consideration for strong performers. Hiring outcomes remain discretionary and performance-based.
Frequently asked questions
What was the HackRx 6.0 problem?+−
Teams built an LLM-powered query-retrieval system for large unstructured policy documents, contracts, and emails, expected to retrieve relevant evidence and make contextual inferences.
Is the pitch template mandatory?+−
Yes. The 6.0 brief explicitly warned that deviation from the provided pitch-deck template could make a submission invalid. Treat every format and portal rule as a hard acceptance test.
How do I improve a RAG leaderboard score?+−
Measure the pipeline by stage: ingestion/OCR, chunking, retrieval recall, reranking, context construction, answer correctness, citation accuracy, latency, and abstention. Use a private holdout and change one variable at a time.
What should the finale demo show?+−
Show realistic documents, multiple query types, source citations, one difficult cross-document inference, an unsafe or unanswerable request, latency/cost, and the operator/user journey. Keep a recorded fallback.
Does a high leaderboard rank guarantee a job?+−
No. Shortlisting and employment consideration use organiser criteria beyond the visible score, and any internship or full-time path remains a separate discretionary process.
Checked against the official HackRx portal, Bajaj Finserv Health's announcement, and the published HackRx 6.0 brief, 15 Aug 2026. The query-retrieval details and cash figures apply to 6.0; the next edition may use a completely different domain or evaluation system.