September 2026
- Three certificates, and what each was for
CS50x, Anthropic's Claude architect exam and two Neo4j certifications. None of them is research. Each covered a specific layer of what I build.
- Where the global layer goes decides when a hybrid learns to retrieve
Our ICLR 2027 submission: 49 small models, 15 designs, one mechanism changed at a time. Attention-sink size did not predict recall. Layer order and key binding did.
- Four validated upgrades that don't add up
LAF-YOLOv10 combined four published small-object detection techniques in one detector. Together they cost 7.8 points of mAP, and the loss traces to one interaction.
- LaTeX to Word without losing the equations
Nine months of work, now on the VS Code Marketplace: editable equations, live cross-references, and failed builds explained in plain words.
- Instructions hidden in log fields
Guardian-AI v3.0 sits between untrusted security telemetry and the LLM that reads it. A regex list catches the obvious attacks. The rest need layers.
- Flood and fire damage per property, from two Sentinel-2 passes
AlphaEarth goes from cloud-masked composites to damage tiers for each parcel, and scores itself against independent damage records.
August 2026
- A data node where the text never leaves the device
Collecting Telugu, Tamil, Hindi and Bengali text for training without shipping anyone's health records to a server. The rule is one line; enforcing it took four separate guards.
- I audited my own paper
HAGD's search is fast. The circuit it found in GPT-2 Small kept no more behaviour than random features, and the preprint's larger results were not in the code. What the audit found, and what I changed.
July 2026
- What our AgentCompress notebook shows, and what it doesn't
The paper's benchmark is not public, so the released notebook runs on synthetic workflows and its numbers are lower. Both columns belong in the README.
- Four proposals, one question each
What the proposals we sent to Google, Amazon and GARA this year each set out to answer, including the one that names its own likely negative result.
- Vak joins the Sarvam AI startup program
An AI literacy and interpretability platform built natively in Telugu and Hindi, not translated from an English tool.
June 2026
- Done with the degree
Three years of diploma, three years of engineering, and the two habits they gave me.
May 2026
- Does the explanation predict the model?
In May I began working with Prof. Hamdi Al-Jamimi at KFUPM on a narrow question about sepsis decision support: when a model explains itself, is the explanation true?
- Flood claims from radar in under five minutes
At the 418 Hackathon at Mahindra University we built OrbitClaim Sentinel: Sentinel-1 radar in, a flood-claim decision out, in a JSON small enough to downlink.
April 2026
- Looking for two people who work below the framework
We were preparing proposals for the Google TPU Research Awards and the Amazon Research Awards, so I asked for collaborators who care about memory hierarchies and kernels.
- A cost check for code an agent wrote
At the OpenAI Codex Community Hackathon in Hyderabad we built Gravity, a FinOps check that runs before AI-written code is merged.
March 2026
- Overdue is not a status
A small tool for compliance analysts, and the one modelling decision in it that I would defend in any review.
- Four pull requests to pyGAM, starting with the p-values
Generalized additive models report p-values for their smooth terms. Getting those right needs Wood's rank-r eigendecomposition.
February 2026
- Eleven pull requests to pgmpy in one week
Reading a causal-inference library closely enough to fix it, from d-separators to NaN treatment effects.
- Blocking ads on every device from a Raspberry Pi Zero 2 W
Pi-hole filters ads at the DNS level at home, and Tailscale carries the same filter to my phone when I am out. About two hours of setup.
- An AI chief of staff in 36 hours
At the Hack-Nation Global AI Hackathon with the MIT Sloan AI Club we mapped how information moves inside a company. The RAG pipeline died at hour 22.
January 2026
- Searching a hierarchy instead of the graph
HAGD, our circuit-extraction preprint: cluster the attribution graph, search it from the top, and report a circuit only after it passes causal tests.
- Not every agent step needs the whole model
AgentCompress picks a precision for each step of an LLM agent's workflow, from the first 32 tokens of the request.
- Pruning for the disease, not for ImageNet
Our first preprint of the year keeps the channels that separate one leaf disease from another, so a few-shot model fits on a Raspberry Pi 4.
December 2025
- A PDF in, a training series out
At the ElevenLabs Worldwide Hackathon in Hyderabad we turned long training documents into short narrated video episodes.
- 2,300 developers, one room, and a fraud-detection agent
Agentathon 2025 set a Guinness World Record for the largest agentic AI hackathon. What I built there, and the one design idea I kept.