Mihir BrahmaniyaGenAI Intern · AI / Data Science
LinkedIn
All projects
Intelligent ClassificationIndustry Project · NoBrokerHood

Ledger Flash

Ledger classification that learns from reviewers, reusing known decisions and calling the LLM only for what’s new.

My roleDesigned and implemented the system end-to-end, including the AI logic, backend, frontend, data processing and automation workflows.

Explore Try Demo

Problem

Transactions often land in the wrong ledger head. Reviewing every posting by hand is slow, and sending every transaction to an LLM is costly and repeats decisions reviewers have already made.

System pipeline

  1. Transactions
  2. Memory match
  3. Classify
  4. Confidence
  5. Human review
  6. Learning

Approach

Validation is part of the system, not an afterthought.

Flags transactions posted to the wrong ledger, recommends the right one with confidence and reasoning, and learns from every review. The workflow keeps inputs, decisions and exceptions visible so that a reviewer can understand how the output was reached.

My contribution

  • Similarity-first retrieval and resilient batch classification
  • Approve, reject and correct review workflow
  • Reviewer-confirmed learning memory and audit history
  • Import, reporting and tenant-isolated data handling

Try the demo

Ledger Flash: interactive demo

Narration → ledger: normalisation, similarity search, local classifier, confidence, approve / change, learning memory.

  • Computed
  • Simulated visual
  • Local model
Transaction

Read id, date, narration and amount.

Session memory: 16 approved examples
◇

Run the sample or upload CSV. Recommendations come from token similarity and transparent local keyword features—not an external API.

This interactive demonstration is a simplified, synthetic representation inspired by an enterprise project. Production systems, company data and proprietary implementation details are not publicly exposed.

Result

A result that can be inspected.

The portfolio demo classifies narrations with approved-memory similarity and local features, exposes confidence and reasoning, and learns corrections for the current session.

Technical architecture

Stage-by-stage system design

  1. 01
    Transactions · Tenant-isolated ingestion

    CSV / XLSX / PDF → Transactions with current ledger

  2. 02
    Memory match · Similarity retrieval

    Narration + current ledger → Reused decision or miss

  3. 03
    Classify · Resilient batch inference

    Unmatched transactions → Suggested ledger + reason

  4. 04
    Confidence · Confidence scoring

    Suggestion → Scored decision

  5. 05
    Human review · Human-in-the-loop

    Scored decisions → Confirmed ledger

  6. 06
    Learning · Active learning

    Confirmed ledger → Updated memory

Tech stack

Tools used in the production system

  • FastAPI
  • Gemini 2.5 Flash
  • Similarity matching
  • Google Sheets API

The portfolio demo itself is static TypeScript running locally in the browser with synthetic data and no API key.