


Reconciliationbreaksthemomentrealmoneymoves.
TruncatedRTGSnarrations.AsymmetricTDSdeductionsunder§194C&§194J.Missinginvoicenumbersacrosssplitbatchsettlements.
MilaaneliminatesgreedycollisionswithprovableHungarianoptimization.










Probabilistic Fellegi-Sunter scoring + LLM adjudication over noisy synthetic bank transaction narrations
Modeling asymmetric TDS §194C/J deductions, GST net invoices, and UTR clearing
Kuhn-Munkres O(n³) Hungarian assignment resolving greedy allocation collisions
Gemini 2.5 Flash-Lite → Groq Llama 3.3 70B failover explains every ambiguous match in plain English with a bounded ±20 confidence delta
Verifiable confidence scores across 60–80 synthetic records — AI enhances explanation, deterministic code enforces the final decision



TruncatedRTGSnarrations.AsymmetricTDSdeductionsunder§194C&§194J.Missinginvoicenumbersacrosssplitbatchsettlements.
MilaaneliminatesgreedycollisionswithprovableHungarianoptimization.
Evaluated on synthetic test batches modeled after Indian banking narrations and statutory TDS deduction structures.
Evaluated on simulated Indian bank feeds and ERP rows containing intentional noise, truncated RTGS narrations, and split payments.
Calculates minimal-cost bipartite optimal assignment matrix to eliminate greedy allocation collisions across concurrent rows.
Formal benchmark against live production enterprise datasets will be conducted post-buildathon evaluation.
Three algorithmic pillars converge into one deterministic confidence score.

PAN-UTR exact pairing, statutory TDS §194C/J threshold brackets, and bank IFSC routing tables.

Fellegi-Sunter log-likelihood vectors, sentence-transformer embeddings (all-MiniLM-L6-v2), phonetic fuzzy resolution — ambiguous cases escalated to LLM adjudication.

Kuhn-Munkres O(n³) Hungarian assignment and subset-sum solvers resolving multi-invoice bulk payouts.
Pass 5 invokes an LLM only on genuinely ambiguous records. AI enhances explanation quality. Deterministic code enforces the final decision.
Invoked only on medium-confidence ambiguous records. Proposes a plain-English explanation + a bounded confidence delta. Never makes the final accept/reject call alone.
Seamless failover when Gemini rate-limits mid-batch. Exponential backoff, circuit breaker, and graceful degradation ensure the batch completes even if both LLM providers are exhausted.
Every LLM response is validated with Pydantic, stored as raw_llm_response in the audit log, and capped at a bounded confidence delta. Hallucinations cannot silently corrupt the ledger.
Design principle: AI proposes. Deterministic code decides. LLM confidence adjustments are strictly bounded. The final accept / review / reject threshold is enforced by auditable, inspectable code. Hallucinations cannot silently pass a bad match.