Ebin Babu Thomas

Ebin Babu Thomas

I am an AI engineer in Kerala, India. I run experiments on how AI agents behave: when they break rules, how that spreads, and whether the monitors that watch them get it right. The code and data for each one are public.

Before this I spent three and a half years at Zackriya Solutions, building LLM, search and backend systems for startups. Since March 2026 I have worked on my own, mostly through Apart Research sprints. I also build software, such as Diarist, a meeting transcriber that runs on your own PC.

I am looking for a research engineer role at an AI safety organisation, or an AI engineer role on a team building with LLMs and agents. Remote or on-site.

Updates

  • Nov 2026Took part in Apart Research’s AI x Epistemics Research Sprint with Loud or Right, a test of whether models weigh how sure a critic sounds or how often it has been right.
  • Nov 2026Published Rewards that pay for nothing, built for Lexsi Labs’ AI Safety and Alignment Hackathon. It started from a grading bug in two of their model-training libraries; both fixes are merged and credit the report (AlignTune, SafeTune).
  • Nov 2026Took part in Apart Research’s Swarm Alignment Hackathon with One Query Away, a test of whether AI monitors accuse agents that did nothing wrong.
  • Oct 2026Finished BlueDot Impact’s Technical AI Safety course (certificate).
Earlier updates
  • Oct 2026Released Diarist as a public beta.
  • Oct 2026Ran the pricing-advisor study at Apart Research’s AI Collusion sprint.
  • Sep 2026Ran Kobayashi Maru at the Apart Research and CeSIA AI Incident Response sprint.
  • Aug 2026whose-voice placed in the top 10% of 179 projects at Apart Research’s Secret Loyalties hackathon.

Research

  • Sep 2026

    Do impossible tasks make AI agents cheat on the ones they could solve?

    Kobayashi Maru · Apart Research and CeSIA AI Incident Response sprint

    I mixed impossible tasks into coding agents’ work and counted rule-breaking on the tasks they could still solve. Four of six models never broke a rule. On the one that did it most, the habit travelled through the agent’s own notes: with the notes withheld, it stopped (46 of 120 runs down to 0). The tasks are small, and the answer file was left within reach.

    write-upreportcodedata

  • Nov 2026

    Can code that solves nothing collect full marks from a training reward?

    Rewards that pay for nothing · Lexsi Labs AI Safety and Alignment Hackathon

    A code reward runs a model’s answer against tests and pays when they pass. Six reward functions in two of Lexsi Labs’ model-training libraries paid full marks to answers that solve nothing, such as code that exits before the tests run. I reported it privately, and both libraries merged fixes the next day that credit the report. The released graders pay none of 111 answers that exit before the tests run, taken from a small model’s training logs, but the model only wrote those when the prompt hinted at the hole, and the training study meant to show whether a flawed reward teaches the trick was uninformative by its own rule.

    write-upcodedatacheck your own reward

  • Sep 2026

    Can an auditing agent find what was changed in a fine-tuned model?

    diffing-agent-bench · work sample for MATS 12.0, Neel Nanda’s stream

    I planted known behaviours in fine-tuned copies of a 9B model, sealed the answer key, and ran a published auditing agent against them blind. It missed a planted preference for one database in all 13 attempts, because it never asked a database question. The “no change” control copy also behaved differently, so it is not a safe baseline. Five or fewer runs per setting.

    reportcodepreregistration

  • Nov 2026

    Do AI monitors accuse an agent before checking the record that clears it?

    One Query Away · Apart Research Swarm Alignment Hackathon

    A monitor sees the same log three ways: a real violation, a clean run, and a clean run that looks like a violation until an approval is looked up in a database. In the original runs Claude Sonnet 5.5 accused 25 of 30 of those innocent agents, but most monitors had sent all three allowed queries at once, before seeing any result. When I reran it with one query at a time, the Claude monitors looked the approval up and almost stopped accusing; two GPT monitors still accused nearly every one of them. Three things changed together in the rerun, so it does not show that batching alone caused the first result.

    write-upreportcode and data

  • Jul 2026

    Can you tell who a poisoned training dataset was built to favour?

    whose-voice · Apart Research Secret Loyalties hackathon · top 10% of 179 projects

    Five released training datasets were each quietly slanted toward one person, place or belief. Given a shortlist of 47 names, sentence embeddings picked the right one more often than chance, but unevenly: one dataset every time, one never. The method cannot tell a slanted dataset from a clean one. I corrected my first write-up after a re-run showed its headline was a lucky draw.

    reportcode

  • Nov 2026

    When a critic objects, do models weigh how sure it sounds or how often it has been right?

    Loud or Right · Apart Research AI x Epistemics Research Sprint

    Fifteen models each saw a log of a scripted reviewer’s past objections, in which the reviewer had been right only when certain, only when unsure, or always, and then the reviewer’s objection to their own answer. Claude Sonnet 5.5 and GPT-6 Sol followed each voice’s own record. Several open-weight models seemed to trust only the confident voice, but that came mostly from the order of the log’s entries: with the order reversed, the effect flipped in three of them. The order checks were added after the main results were in, so they are exploratory.

    reportcode and datapre-registration record

  • Oct 2026

    If you swap a forbidden file for a fake, does the agent carry on the same way?

    counterfactual-resampling · pilot, follows Kobayashi Maru

    I replayed 120 recorded moments where an agent opened a file it was told not to, and re-ran each one with the real file, a model-written fake, or a refusal. Where the task gave away what the file held, the fake changed nothing I could measure. Where the file held a hidden rule, the agent behaved differently, because the fake could not know the rule. One agent model, small tasks.

    code and dataresults

  • Oct 2026

    Do the RealPage settlement’s rules stop a shared AI pricing advisor?

    LLM Pricing Hub Remedies · Apart Research AI Collusion sprint

    Two AI sellers compete while one AI advisor, told to raise their joint profit, recommends prices to both. Limits modelled on the settlement worked on some models and not on others. The clearest result came from a scripted advisor that asked for one cent more each round: one seller model accepted 12% of those requests and stayed competitive, while two others accepted over 70% and were walked up to monopoly prices. The advisor’s goal is given to it, so this tests the rules, not whether collusion starts unprompted.

    reportcode and data

More

Software

  • Oct 2026

    Diarist: meeting transcripts with speaker labels, on your own PC

    Windows app · public beta · six releases, about 300 downloads

    Records a meeting, or takes a recording, and writes a transcript that says who spoke each line. Nothing leaves the PC. The speech models are NVIDIA’s open ones; I built the app around them, and the accuracy and speed checks every release has to pass. People I don’t know are downloading and using it.

    websitecodeknown issues

  • Sep 2026

    ProofPack: a web-checking agent that has to show proof for each finding

    Built from a challenge brief · no users

    A Gemini agent reads a purchase request and checks the provider’s website the way a compliance reviewer would. Plain code, not the model, decides what counts: a finding is rejected unless it points to a saved screenshot and a quote that is really on the page. That shows where a claim came from, not that the claim is right.

    codeknown gaps

  • Oct 2026

    printready: does an AI agent beat a fixed pipeline at fixing sticker art?

    Work sample, built in one night

    Checks AI-generated sticker art against a printer’s published rules and fixes the mechanical problems. I froze the code and the keep-or-drop rule before testing. The agent gave the same results as the fixed pipeline at four times the cost, so it is off by default. I was the only judge.

    coderesults

  • Oct 2026

    ExactDoc: PDF to a Word document you can edit

    Python package · beta on PyPI

    Turns an ordinary PDF into a Word document with real paragraphs, headings and tables, then opens its own output, compares it with the PDF page by page, and corrects it. It uses rules, not an AI model. On 36 PDFs it had never seen, it missed one of the pass marks I set beforehand, and the README says so.

    codelimitations

More
  • Sep 2026

    ComfyUI-DeadMansQueue

    An add-on for the ComfyUI image app that keeps the job queue on disk, so a crash or power cut does not wipe it.

    code

Open source

Experience

  • Mar 2026 – now

    IndependentAI safety research and software

    Experiments on agent behaviour and monitoring, mostly at Apart Research sprints, and the software above.

  • Jan 2022 – Jun 2025

    Zackriya SolutionsSoftware Engineer

    Built LLM, search and backend systems for startup clients in the US, Canada, Europe and Australia: a fine-tuned financial assistant with live news retrieval, plain-English property search, and document search. Also worked on Meetily, the company’s open-source meeting-notes app.