Engineering AI, from research to production.

Founding AI engineer and published ML researcher. I build LLM-powered retrieval and multi-agent systems alongside high-performance machine learning pipelines, backed by peer-reviewed research spanning astronomy and accessible robotics.

fig. 1 — lorenz attractor · σ=10, ρ=28, β=8/3 · ● live · why?

● OPEN TO INTERESTING PROBLEMS

LAST COMMIT TODAY

CURRENTLY BUILDING AMPHOREUS

§ 01

Selected Work

ALL PROJECTS →

Amphoreus D&D — Multi-Agent LLM Simulation

A fully automated multi-agent simulation of a D&D campaign: a Dungeon Master, an Orchestrator/referee, and player-characters, each backed by an LLM, coordinated by a structured engine loop. World state changes only through typed tool calls — never parsed from prose — so every change is logged, replayable, and testable. The engineering focus is on reliable ways to promote and evaluate long-horizon planning in agent behavior, and also observe inter-agent coordination. Ongoing project.

Multi-agent orchestration. Three distinct cognitive roles coordinated by a 4-stroke turn loop; world state mutates only through typed, logged tool calls, never parsed prose.

Search-earned, typed memory. Agents pull top-K relevant context from owned, vector-indexed stores (ChromaDB); provider-independent, so any role can run a different model.

Evaluation of non-deterministic agents. A variance-aware harness (N≥3 per condition, distributions not point values, an independent LLM-judge, pre-registered thresholds) to help decide what to build by measuring first.

WATCH THE DEMO ▸ READ THE WRITE-UP

fig. 2 — the turn cycle · every intent passes through the orchestrator · no orbit repeats

2026 ONGOING AI ENGINEERING

Accessible Wheelchair Routing

Built two datasets of vibration data from manual and powered wheelchairs across fifteen surface types. Trained a surface-classification model using adaptive activation functions and extended it to powered wheelchairs via transfer learning, reaching 98.8% accuracy — outperforming SOTA by 3% with 40% less train time.

PAPER · SN COMPUTER SCIENCE

fig. 3 — wheelchair users on mixed paving — the terrain the router learns to read

2024 PUBLISHED ML RESEARCH

Foreground Emissions Modeling for Astronomy

Parsed and analyzed deep-sky observation data from the GALEX satellite's spacecraft state files. Derived an empirical model of UV foreground emissions using robust regression, and generalized it to deep and medium-sky observations.

PAPER · ADVANCES IN SPACE RESEARCH

fig. 4 — the GALEX ultraviolet telescope, artist's concept · NASA/JPL-Caltech

2023 PUBLISHED ASTRONOMY / ML

§ 02

Writing

ALL POSTS →

RAG, Revisited — field notes from 2023–24, re-read against the 2026 canon

6 PARTS · JUN 2026

№ 00 RAG, Revisited JUN 30

№ 01 Atomicity Is the Target, Not the Algorithm: Chunking in 2026 JUN 11

№ 02 You Can't Retrieve What You Didn't Index: Coverage-Bounded Recall JUN 15

№ 03 When Your Embedder Thinks Everything Is a Finance Question JUN 19

№ 04 Routing RAG by Question Complexity: A Bloom-Taxonomy Lens JUN 23

№ 05 Good Idea, Good Result: The RAG Evaluation Gap JUN 27

LATEST

✳ Amphoreus D&D: orchestrating believable, long-horizon agent behavior JUL 02

§ 03

Publications

AdaGen: Adaptive Generalized Knowledge Transfer Framework for Sensor-Based Surface Classification for Wheelchair Routing

SN Computer Science, 2024 · doi:10.1007/s42979-024-03181-w

Validation of Foreground Emission Model using GALEX Medium Observations

Research in Astronomy and Astrophysics, 2023 · doi:10.1088/1674-4527/acd09f

A model of foreground emission in UV using GALEX deep observations

Advances in Space Research, 2023 · doi:10.1016/j.asr.2022.07.086

§ 04

About

I've been drawn to the same question since I first saw a Lorenz attractor: how do simple rules produce behavior you couldn't have guessed? These days I ask it of LLM agents instead of differential equations.

Earlier work spanned vibration data from wheelchairs, ultraviolet light from a space telescope, and retrieval over financial text — plus a focused two years studying India's public policy and governance, which sharpened how I analyse complex, real-world problems.

NOW

Building — Amphoreus: long-horizon planning & coordination in agents
Writing — what comes after RAG, Revisited
Elsewhere — GitHub · LinkedIn · [email protected]