The problem worth solving

The pain point isn't "we don't have an early warning system." Every large PSB already has one
— usually a vendor rules engine bolted onto CBS that fires triggers on overdue days, cheque
returns, LC devolvement, stock statement delays. The real problem, and the one you'd have
seen up close:

The alerts are backward-looking, they fire in the hundreds, and nobody can explain them.
A credit officer at a zonal office gets a list of flagged accounts every month. Most are false
positives. Each one still needs a manual justification note before it moves to SMA-2 watchlist or
triggers a Stage 1 → Stage 2 migration under Ind AS 109. The officer opens the CBS ledger, the
last CMA data, the stock statement, the sanction letter, and hand-writes a paragraph. That
paragraph is what actually goes to the credit committee.

So there are two separate failures, and the second is the one nobody automates:

1.​ Detection is rule-based and lagging — it catches deterioration after it's visible in
repayment behaviour.
2.​ Justification is entirely manual — which is why SICR assessment across a large book
becomes a tick-box exercise rather than genuine judgement.

Traditional EWS frameworks rely on backward-looking data such as demographic and financial
information, which limits their predictive power. That's the gap, and it's documented enough that
your literature review writes itself. EY

The solution: SANKET

Signal-based Assessment for NPA-risk and Credit Early-warning Tracking. (Sanket = "signal" in
Hindi — plays well with a PSB audience.)

A two-layer system, and the two-layer design is the whole pitch:

Layer 1 — Predictive engine (ML). A gradient-boosted classifier trained on account-level

features (repayment behaviour, turnover-to-limit ratios, cheque return frequency, stock
statement lag, utilisation volatility, external signals like GST filing delays or MCA charge
modifications) that outputs a forward-looking 90-day probability of SMA/SICR migration. This
gives you hard, measurable accuracy numbers.

Layer 2 — Explanation and drafting engine (LLM + RAG). Takes the model output plus
SHAP attributions plus retrieved context from the bank's own credit policy, the RBI IRAC norms,
and the account's sanction terms — and drafts the SICR justification note with clause-level
citations. The officer edits and signs. Never auto-approves.

The USP is Layer 2. Anybody can train an XGBoost model. What no vendor product does well
is close the loop between a risk score and the auditable narrative a credit committee actually
needs. That's your differentiation, and it's also why your governance section becomes strong
rather than boilerplate.

The stack, and the "why"

Layer Choice Why over the alternative

Model XGBoost / LightGBM + Beats deep learning on tabular credit data, and
SHAP more importantly it's validatable. A neural net can't
survive RBI model-risk review; SHAP attributions
can.

LLM Llama 3.1 8B or Mistral 7B Not a cost decision — a data residency decision.
via Ollama, run locally Customer credit data cannot leave the bank's
perimeter. GPT-4 class models via API are
non-starters for this use case regardless of quality.

RAG LangChain + ChromaDB + All open-source, all local, no egress. Chroma over
BGE-small embeddings Pinecone for the same residency reason.

Interfac Streamlit Prototype speed. Note in your plan that production

e would be a React front-end on the bank's intranet.

Data Synthetic loan book + public Be explicit about this — examiners respect a stated
proxy (Home Credit / limitation more than a vague claim about "bank
Lending Club for model data."
validation)

The single sharpest line in your "why" section: every component runs on-premise with zero
data egress, at zero marginal cost per inference. That's a genuine architectural argument,
not a preference.

Answering the evaluation criteria

Performance. Report in the language a banker uses, not a data scientist: AUC-ROC, KS
statistic, and Gini — PSB risk teams live in Gini and KS. Realistic targets on a public proxy
dataset: AUC 0.82–0.87, KS 40–50, Gini 0.62–0.72. Add precision@top-decile, because the
actual business value is "of the 100 accounts we flag, how many genuinely deteriorate" — that's
your false-positive-reduction claim. For the LLM layer, measure separately: latency (tokens/sec
on your hardware), and faithfulness/groundedness via RAGAS on a held-out question set.

Analytical power. Frame it as three decisions unlocked: (a) which accounts to move to
watchlist before the 30-DPD trigger, (b) portfolio-level ECL sensitivity — if Stage 2 migration
rises 200bps, what happens to provisioning, (c) collection effort prioritisation by expected loss
rather than by overdue amount.
Cost. Prototype: ₹0, genuinely — everything above is free or open-source. Production,
indicative: a single GPU server capable of serving an 8B model to a few hundred concurrent
users lands around ₹12–20 lakh capex, or roughly ₹1.5–3 lakh/month on reserved cloud GPU.
Compare that against a per-token API alternative at portfolio scale and the on-prem case gets
stronger as volume rises — build the break-even chart, it's a good slide. Flag these as indicative
estimates; you'd want to sanity-check current India pricing before you present.

Trust and governance. This is where you win, because you can anchor to something live. The
RBI published its FREE-AI Committee Report on 13 August 2025, setting out seven guiding
principles and 26 recommendations across six pillars, applicable to all RBI-regulated entities
including scheduled commercial banks — Union Bank included. Key obligations include a
board-approved AI policy, structured governance across the AI lifecycle with independent
validation and periodic audits, and strict DPDP Act 2023 compliance. Map your controls directly
onto that: mandatory citation with refusal when retrieval returns nothing relevant, confidence
thresholds below which the system declines to draft, full audit logging of
prompt-response-officer-edit, SHAP attribution attached to every score, and human sign-off as a
hard architectural constraint rather than a policy suggestion. RegulationsRegulations

Prototype scope

Build these four screens and you have a demo that holds up under questioning:

1.​ Portfolio dashboard — accounts ranked by predicted migration probability

2.​ Account drill-down — SHAP waterfall showing which factors drove the score
3.​ Draft SICR note — LLM-generated with inline policy citations, editable
4.​ ECL impact simulator — reuse your existing PD/LGD/EAD logic to show provisioning
delta