AGI Labs · Kumbakonam, India
Founder, AGI Labs/AI Engineer, Neo Wealth
I build AI systems that answer from real sources — and I don't accept a result I can't explain mechanistically.
That instinct runs in two directions. One is production: knowledge graphs, grounded retrieval and agent orchestration for enterprises that need the answer to be right. The other is mechanistic interpretability, where the question stops being whether the model works and becomes what it is actually doing.
Now
One job, one lab, one research direction. They pull on the same thread: making a system's answer traceable back to something real — a document, a graph edge, a circuit inside the model.
Jan 2026 — present
Building the AI backend for a conversational trading platform: agentic workflows that turn a plain-language request into portfolio insight and trade actions, with the orchestration, tool integration and conversational state that real-time finance demands.
Aug 2026 — present
Building Tark, a rehearsal courtroom for Indian advocates grounded in Supreme Court judgments — plus assistants on other curated corpora where every answer is fetched from a source before it is spoken.
Ongoing
Replicating results, running small experiments, and writing them up. The bottleneck on this path isn't ability or compute — it's legible public output. So I'm producing it.
AGI Labs
The flagship, Tark, lives on the AGI Labs site →
Most conversational systems answer from a model's memory of the internet. Mine retrieve from an actual corpus — the 701 verses, the 1,556 pages, the clinic's own intake rules — and can show you the line they used. The interesting engineering is not the chat; it's the data layer underneath it.
All 701 verses with audio, word-by-word meanings and classical commentaries, served directly from the corpus. Ask a question in plain language and get the verse it came from, not a paraphrase of one.
1,556 pages of classical astrological texts indexed, with the source layer kept strictly separate from corrections and from the rules derived on top of them — so a claim can always be traced back to the page it came from.
A WhatsApp system for an Ayurvedic practice, live with real patients. It greeted each one in their own language, collected exactly what the doctor needed, and handed the conversation to a human at the boundary where a machine shouldn't be answering.
CV analysis and structured self-appraisal, with a graph as memory: what the agent learns about you persists as nodes and edges it can query later, instead of being flattened into a context window.
Record
The numbers below are the ones I'd defend in an interview: what was built, how big it actually was, and what it cost to run.
Jan 2026 — present
Neo Wealth Asset Management · India
Jul 2024 — Jan 2026
Karini AI · Pune
Aug 2023 — Jun 2024
Mercedes-Benz Research & Development India · Pune
Jun — Nov 2022
HSBC WCR Analytics · Bengaluru
Research
I'm an applied mathematician by practice, not a trained pure one, and I don't take a working black box on trust. What pulls me toward interpretability is that it's the one part of the field where the honest answer to "why did it do that?" is still being constructed. I'm working toward research at a frontier lab or a funded PhD, and I treat public, reproducible output as the thing that gets me there rather than a byproduct.
Interpretability research on natural language autoencoders — the pair of models that turn a residual-stream activation into an English explanation and back again. Experiments are running now; a technical write-up will follow, and I'll link it here when it does.
Supervised and preference-style fine-tuning of a Qwen 3 model on Thinking Machines' Tinker API, teaching it agentic behaviour over a graph: when to reach for a tool, how to write a Cypher query, and how to read what comes back. Running on research credits since November 2025 — and the more interesting question underneath it is what actually changes inside a model when you teach it a query language.
A faithful implementation of hierarchical navigable small world graphs straight from Malkov & Yashunin (2016) — algorithms, probability distributions and complexity analysis translated directly from the paper, so the index is something I understand rather than something I import.
An agent framework backed by Neo4j: state is stored as a graph rather than abstracted away from one, so an execution can be watched live in the browser and the whole history of every run can be interrogated with Cypher.
Hands-on work on ARC-AGI tasks — using coding agents to propose minimal ONNX networks per puzzle — and, more usefully, a written account of where that approach succeeds, where it quietly fails, and what the gap says about program synthesis under an agent.
Writing
Technical explainers built the long way round: start at the smallest case you can hold in your head, count it upward with a pencil, and only then check whether the machine agrees. Each one has interactive widgets and a measured implementation, so you can see the claim rather than take it.
Why attention runs out of memory before it runs out of compute — deriving the online softmax from a small matrix by hand, then building the tiled kernel that falls out of it.
Read ↗ Explainer28 Jul 2026PyTorch's free-normalization work, approached from a 2×2 matrix and worked upward — memory hierarchy first, so the optimisation reads as inevitable rather than clever.
Read ↗Written in my own words after working the derivation through, with an AI tutor for the parts I hadn't understood yet and a proof-reading pass at the end. The understanding is the point; the article is the artefact that proves it.
Shorter pieces on research as it lands: architecture papers, lab strategy, and the parts of the tooling most people skim past.
Credentials
Indian Institute of Information Technology, Prayagraj
Sri Sairam Engineering College, Anna University, Chennai
University of Alberta · full four-course sequence
Competing since 2020 — 44 competitions entered, best finish top 500 in a playground series.
Contact
I read everything. If you're a lab or a professor looking at interpretability work, a team hiring research engineers, or someone with a corpus that deserves a proper assistant — write to me.
giridharanadithya@gmail.com