LedgerMind: a hackathon AI agent with persistent memory that learns recurring vendor GST errors across cycles

Vast-Moose639 · reddit · 2026-11-11

LedgerMind, built for the HackwithBangalore 4.0 hackathon, tackles Indian GST reconciliation: existing tools treat each month as a blank slate, so finance teams re-discover the same vendor errors every filing cycle.

The autonomous agent reads purchase registers and GSTR-2B files to match invoices and learns vendor behavior over time via Hindsight by Vectorize: month 1, it spots a GSTIN mismatch and stores the resolution; month 3, it recalls the pattern and drafts a specific follow-up using the communication channel that worked; month 6, it proactively warns the finance team before a chronically late vendor's deadline—shifting from reactive matching to predictive intelligence.

Stack: Hindsight SDK for memory, Groq inference (GPT-OSS-120B / Qwen3-32B) with heavy function calling, Python + FastAPI + Pandas for deterministic CSV/Excel matching (keeping the LLM away from raw math), HTML/CSS/JS frontend. The team is soliciting feedback on invoice-matching edge cases and cross-session agent memory.

Original post →

More from coding & agent

coding & agent channel →