Aayush Tiwari, home
Résumé
§2 Selected work

CONCLAVE

Multi-Agent LLM Decision System

Status
Research prototype
Role
Repository owner and primary implementer of the research pipeline, evaluation tooling, and dashboard.
Domain
Applied AI / Multi-Agent Systems
Source
Repository (opens in a new tab)
Figure 1Shows the full flow from public sources through normalisation, the feature engine, a four-way agent fan-out and fan-in, deterministic consensus, evaluation, and the dashboard. The counts it carries (20+ engineered metrics, six dashboard tabs) describe the system's shape. No measured scores, timings, or throughput appear.
§1Brief

The problem this had to solve.

Compare role-specialised consensus with simpler mutual-fund ranking baselines without presenting a research interface as financial advice.

CONCLAVE ingests public mutual-fund and economic data from AMFI, MFAPI, yfinance, and RBI sources, then turns it into a feature set covering more than 20 fund metrics.

Four role-specialised LLM agents can run through Gemini or local Ollama/Qwen2.5-7B providers. Their outputs feed deterministic consensus and fallback behavior before evaluation, backtesting, and six-tab Streamlit exploration.

Constraints

  1. The current 86-fund snapshot contains surviving Indian open-ended equity funds and is subject to survivorship bias.
  2. Available NAV history begins in 2013, so the persisted multi-vintage study covers a predominantly bullish market regime.
  3. Live ingestion depends on AMFI, MFAPI, RBI/manual inputs, and yfinance availability.
  4. LLM calls are optional and non-deterministic; heuristic and mock fallbacks are required for reproducible offline runs.
§2Architecture

How the system is put together.

Public data sources

AMFI, MFAPI, yfinance, and RBI data

source

Ingestion

Collect and normalize source material

process

Feature engine

Build 20+ fund metrics

process

Four specialised agents

Provider-agnostic Gemini or Ollama/Qwen2.5-7B inference

process

Consensus

Deterministic decision and fallback behavior

decision

Evaluation

Backtesting and result analysis

process

Dashboard

Six-tab Streamlit exploration

output
§3Decisions

What was chosen, and what it cost.

Separate evidence from model judgement

Choice
Engineer return, risk, cost, and consistency objectives before asking four role-specialised agents to rank candidates.
Rationale
Every agent receives the same bounded fund evidence while its role changes the priority applied to that evidence.
Cost
The role prompts are deliberately prescriptive, so the agents are closer to specialised rerankers than unconstrained analysts.

Keep a deterministic decision path

Choice
Aggregate proposals with explicit consensus and fallback logic instead of delegating the final result to one free-form response.
Rationale
Offline runs remain reproducible when a provider is unavailable and negotiation decisions can be replayed.
Cost
Fallback-heavy runs cannot be treated as evidence of LLM reasoning quality.

Evaluate against simple baselines

Choice
Backtest consensus alongside mean-of-objectives, five-year CAGR, random, and bottom-five controls.
Rationale
A complex agent workflow should be compared with inexpensive ranking rules before claiming an advantage.
§4Validation

How it was checked.

  • The repository includes automated tests, backtesting scripts, evaluation baselines, and persisted research reports.
  • The multi-vintage analysis covers 36 paired vintage/category cells; its reported consensus-versus-simple-baseline effect is directional but not significant at the two-sided 0.05 level.
  • The Streamlit application exposes six tabs for inspecting funds, sessions, comparisons, and investor-profile reranking.
CONCLAVE Streamlit dashboard showing the mutual-fund decision system interface.
Figure 2 · Repository artifactThe repository's Streamlit interface exposes fund evidence, agent sessions, comparisons, and investor-profile reranking across six tabs. Inspect the source (opens in a new tab)
§5Evidence

What was measured.

Each figure states the conditions it was measured under. None of them describe production traffic.

4

role-specialised agents

Independent roles contribute to the ranking decision.

20+

fund metrics

Features assembled from public market and economic sources.

6

dashboard tabs

A Streamlit interface supports interactive exploration.

§6Limits

Where this stops being true.

  • The repository explicitly excludes real-money execution, financial advice, hosted deployment, and guaranteed live-data availability.
  • Historical TER data is unavailable in the snapshot, category-specific benchmark history is incomplete, and external CRISIL/Value Research validation is not persisted.
  • The LLM study uses a single local model and mostly single-round negotiation; trained adapter inference was not evaluated end to end.
  • The dashboard does not yet surface the multi-vintage backtest summary.
§7Technology

What it is built with.

Runtime & interface

  • Python
  • Streamlit

Data & evaluation

  • Pandas
  • NumPy
  • SciPy

Infrastructure & providers

  • Ollama
  • Qwen2.5-7B
  • Gemini
Explore the repository (opens in a new tab)
Next documentTracePilot