NABARD Hackathon 2026 · AI/ML cash-flow forecasting

See rural-enterprise stress before a missed repayment.

A consented 26-week cash-flow range turns fragmented financial, market and climate signals into an explainable alert—and a human-approved response.

01 · Configure a consented case

Three challenge-aligned enterprise types; one deep dairy SHG story.

Fixture data · no PII
Women-led SHGenterprise type
12members / owners
Anand, Gujaratdistrict rail
26 weeksforecast horizon
Case Manager
Bounded, tool-using agent

8 allow-listed tools · 2-turn / 6s budget · deterministic numbers · verifier · kill switch · durable replay key

Ready · checking runtime…

02 · Forecast the range, not a false point estimate

P10 / P50 / P90 net cash flow by week; visible uncertainty and stress window.

Model artifact loading

Sakhi Dairy SHG · weekly net cash flow

₹ thousands · future 26 weeks · lower values mean less repayment headroom

Stress begins W8
P10–P90 uncertaintyP50 expectedP10 / P90stress window
Temporal split: 3 rolling origins × 8 validation weeksInterval coverage —Quantiles are ranges, not guarantees

03 · Compare a response, then stop at the human gate

Counterfactual scenarios support judgement; they are not causal promises.

No autonomous execution

Intervention simulator

Choose one approved playbook option.

What-if
₹7.6killustrative P10 weekly uplift after stress begins
—revised lowest weekly P10

Simulation changes a deterministic scenario curve. It does not estimate causal treatment effect and cannot change an account.

Maker–checker gate

Separation of duties is part of the product.

Locked
✓ Maker prepared→○ Independent checker→○ Customer action
Run the review to create an action card

No customer has been contacted and no account has changed.

Policy P-04 · same actor cannot prepare and approve · override requires reason · withdrawal and kill switch checked again before action.

04 · Inspect the harness—and the evidence behind it

Visible tool events, stable replay fingerprint, numerical fidelity check and release metrics.

Offline-ready
AUDITRun the review to generate a replay fingerprint.

05 · Built as interoperable public-interest infrastructure

DPG-readiness is demonstrated through reusable contracts and governance—not asserted as certification.

⌁

Open & reusable

Apache-2.0 code, OpenAPI, consent/action schemas, model card and data card.

◉

Privacy by design

Granular purpose consent, minimum fields, withdrawal, TTL and no raw PII in the public demo.

⇄

DPG-aligned rails

AA, LokOS, ULI, AGMARKNET and IMD are adapter contracts; live status is never overstated.

⌾

Accountable operations

Verifier, maker-checker, idempotency, budgets, audit replay, cohort metrics, kill switch and rollback.