entering the canopy…
Kartik Dubey
KATTI who?
scroll

— Identity

Kartik Dubey.

Data Scientist · AI Engineer · Systems Architect

who is Katti?

01 — About

I think in systems, not scripts.

I think in systems. Multi-agent architectures, RAG pipelines, knowledge graphs — the kind of designs senior engineers nod at. I spent my last year shipping an industrial RAG product at TCS after an ML internship at Reliance. Now I'm building my own stack.

Based in Navi Mumbai. Working with Claude as my engineering team. Architect first, engineer second — and getting faster at the second every week.

02 — Selected Work

Four bets. One workshop.

  1. 01

    katti-os

    Personal AI OS · multi-agent · in progress

    A persistent COO-style LLM orchestrating specialist agents over a 3-layer memory, with a council of local LLMs as an anti-sycophancy verdict layer.

    open dossier
  2. 02

    Lexara

    Multi-agent Legal RAG · spec'd

    A 4-agent legal-document Q&A pipeline (Parser → Librarian → Drafter → Critic) targeting 92–95% accuracy against a 65% single-LLM baseline.

    open dossier
  3. 03

    GRPO Trading

    Multi-agent backtest framework · spec'd

    Group-Relative-Policy-Optimization-inspired trading system. Specialist agents act on shared market data; a judge reweights their influence by recent profitability.

    open dossier
  4. 04

    TRIBE-Industrial

    Tri-modal foundation model · research

    Adapting Meta's TRIBE v2 trimodal architecture from neuroscience to industrial field diagnosis — predicting the manual section a senior engineer would consult.

    open dossier

03 — Where I've Shipped

I ship the real thing.

  1. 2025 — 2026

    TCS · Industrial RAG

    An industrial-document Q&A system over OEM service manuals — Qdrant, nomic-embed-text, custom chunking for engineering PDFs, deployed for field engineers.

    open dossier
  2. 2024 — 2025

    Reliance · Churn + EDA

    Customer-churn prediction end-to-end — feature engineering, gradient-boosted classifier, SHAP explainability — plus EDA pipelines turning raw data into decisions.

    open dossier
  3. ongoing

    Independent

    Designing and shipping the four projects above. ~200-commit target this quarter. Public portfolio, public learning.

    open dossier

04 — Contact

Let's build something that runs.

— end of dossier —