Kuzushi Labs · Applied research and engineering

Engineering private financial AI

Small language models for a wealth-management workflow where household data stays on premises. From reviewed financial sources to quantized adaptation, multilingual tools, bounded execution and evidence that can support a deployment decision.

Arunank Sharan · August 2026 · 12 articles · 15,038 words

technical articles

12

target languages

3

multilingual test cases

24

A practical walkthrough of private financial AI, with source-linked datasets, a concrete QLoRA recipe, multilingual tool use, on-premises capacity calculations and an executable benchmark. Each chapter includes code, worked examples and a technical diagram.

The downloadable reference kit includes 48 passing offline tests, eight real FinQA sample annotations and a local model runner. Public examples use authored households; benchmark self-checks are kept separate from candidate model results.

1. PART 01 · 7 min read

   Private financial AI: from a question to a verified answer

   The complete local architecture: adviser workflows, permissioned evidence, deterministic calculations, model decisions and reviewable results.

2. PART 02 · 6 min read

   Financial documents to source-linked training data

   A reproducible data pipeline with document lineage, table semantics, family-level splits and eight real FinQA sample records.

3. PART 03 · 7 min read

   Generate financial QA with executable answer checks

   Question coverage, eight multilingual training-format examples, deterministic calculations and concise source-grounded explanations.

4. PART 04 · 7 min read

   Fine-tuning a local financial assistant with QLoRA

   A concrete Gemma 4 adaptation recipe: source-separated data, NF4, language-only LoRA targets, completion masks and reproducible checkpoints.

5. PART 05 · 7 min read

   Chat templates, loss masks and response contracts

   A token-level walkthrough of prompt rendering, completion supervision, thinking controls and parser-compatible serving.

6. PART 06 · 6 min read

   English, Hindi and Hinglish tool use

   Multilingual financial requests become the same typed arguments, permission checks and evidence-backed outcomes.

7. PART 07 · 7 min read

   A bounded runtime for financial agents

   An executable local agent loop with trusted tool traces, household authorization, numerical provenance, deadlines and explicit failure outcomes.

8. PART 08 · 7 min read

   On-premises serving: memory, latency and capacity

   Weight and KV-cache calculations, a practical local deployment topology, workload matrices and reproducible latency measurements.

9. PART 09 · 7 min read

   Typed scoring for financial answers and agent traces

   Numerical units, entity and date checks, evidence validation, missing-run denominators and executable scorer tests.

10. PART 10 · 7 min read

    Benchmark protocols for FinQA, FinanceBench and IndiaFinBench

    Pinned public datasets, candidate-only inputs, supplied-evidence versus retrieval protocols and comparable financial evaluation reports.

11. PART 11 · 7 min read

    WealthOps: an executable financial-agent benchmark

    Run 24 public English, Hindi and Hinglish scenarios through a local model, read-only tools and a strict outcome scorer.

12. PART 12 · 7 min read

    From benchmark runs to a reproducible release

    A practical release framework for base, adapted and quantized models: manifests, paired comparisons, regression gates, rollout and research references.

Continue with financial RAG, agent harnesses or the FIELD methodology.