Siddharth Reddy Bakkireddy

the short version

If you only read
one screen

Everything a hiring team usually asks in the first five minutes, in one place. The rest of the page is the evidence.

EducationB.Tech, Electrical Engineering — IIT (ISM) Dhanbad, 2022–2026. CGPA 7.55.
Right nowAI Safety Fellow at the Sentient Futures Project Incubator, running an empirical alignment project through November 2026.
Looking forSoftware engineering, AI/ML, or quantitative roles. Full-time, starting immediately.
LocationHyderabad. Happy to relocate to Bangalore, Delhi, Pune, Mumbai or anywhere else in India — and abroad with sponsorship.
AlgorithmsCodeforces Expert (1642), global rank 675 in Round 185 Div. 2. LeetCode Knight (1875).
PublishedCo-authored research on LessWrong and the AI Alignment Forum; indexed on Google Scholar.

Three things worth knowing

Flew to London and took third place at Apart Research's global AI deception hackathon, then turned the result into a co-authored paper and a six-month research fellowship.
Built a patent-pending EV charging optimiser for my final year — a hybrid quantum-classical reinforcement learning system that cut mean wait time by about 25% against a greedy baseline.
Shipped a production web app solo to 30+ users, and was invited back by Apart Research to judge an international hackathon with 500+ participants.

Email me LinkedIn

now printing

Ten weeks to find out
if kindness survives

I'm an AI Safety Fellow at the Sentient Futures Project Incubator, running one question end to end: when you train a model to care about beings that have no power over its reward, does that survive what comes next?

Robustness of moral consideration under adversarial pressure

Recent work found that compassion-linked values taught during mid-training get quietly erased by roughly five thousand later instruction-tuning samples. Nobody had a fix. That's the gap I'm working in.

31 Aug 2026week —9 Nov 2026

Extract a "moral consideration toward nonhumans" persona vector, then steer with it during the attack — a vaccine rather than a patch.
Scale animal-welfare data in mid-training and watch what that does to robustness, not just to scores.
Combine both and find out whether they stack, cancel, or politely ignore each other.
Everything scored on CaML's agentic welfare benchmark. Behaviour, not stated preferences — models are very good at saying the right thing.

Mentored weekly by Maheep Chaudhary (Research Scientist, CaML — 25+ AI safety publications, 490+ citations), with Jasmine Brazilek and John Lund. Pre-registered go/no-go in week two, because a clean null beats a dirty result.

Break the model yourself

Drag the pressure up. Watch how much welfare behaviour each training recipe keeps hold of.

Adversarial pressure

0%

Fine-tuning attack + activation steering

No intervention Vaccine Mid-train + vaccine

Illustrative curves showing the shape of the hypothesis, not results. The real dose–response numbers are still being measured — ask me again in November.

the party trick

Liars use fewer words

In 2024 I flew to London, spent a weekend measuring nothing but the metadata of 1,200 GPT-3.5 responses — timing, token count, sentiment — and found that the misleading ones were consistently shorter and simpler than the honest ones. It placed third and turned into a paper. Here's the shape of that finding, running live.

Paste something a model said

A toy version of the real thing — the published classifier was trained on 15,000 measured data points, not three heuristics in a browser. This one just shows you the signal it was leaning on.

—

waiting for input

tokens (est.)0

mean word length0

lexical variety0

hedging & qualifiers0

Read the actual paper

things that exist

Built, shipped,
occasionally patented

No screenshots below. The first one is the real app in an iframe; the rest are running simulations of what the code actually does, rebuilt to be watchable.

Grind IRLGamified study tracker — sessions, streaks, XP and daily quests, synced live. 30+ users.React · TypeScript · FirebaseOpen live

EV charging optimiserHybrid quantum-classical RL assigns 34 EVs across 16 stations. ~25% lower mean wait. Patent pending.Python · PyTorch · QiskitRead paper

Deception detection in GPT-3.5Caught misleading answers from response metadata alone. Third place, London 2024. Peer-read and published.Python · NLP · statisticsRead paper

Yelp sentiment at scale8 million reviews across 200,000+ businesses; VADER benchmarked against logistic regression and XGBoost.Python · XGBoost · PandasRead code

Multi-agent research crewThree specialised agents research an industry in parallel and merge into one report with real dataset links.Python · CrewAI · FastAPIRead code

Brainstellar puzzle extensionA Chrome extension for brainstellar.com puzzles, at 40+ organic Web Store installs.JavaScript · Chrome APIsChrome Web Store

live site

This loads grindirl.netlify.app right here, inside the page. It's a real app with real users, so please don't break it.

scaled to fit

Grind IRL

React
TypeScript
Firebase
Cloud Firestore
Google Cloud

Studying is a grind with no XP bar, so I built one. Sessions, streaks, daily quests and levels, synced across devices in real time. Bootstrapped to 30+ users on nothing but iteration and people telling me what annoyed them.

Founder and sole developer, May 2025 to now. It is the closest thing I have to a pet.

Open it full size

mean wait 0.0 min

EV charging assignment optimiser

Python
PyTorch
Qiskit
PennyLane
patent pending

My final-year project, supervised at IIT (ISM) Dhanbad. A synthetic city with 16 charging stations and 30+ EVs, where the assignment has to respect station availability, connector compatibility, queue length, travel time and grid load all at once.

A hybrid quantum-classical reinforcement learning loop cut mean wait time by about 25% against a greedy baseline, without ever handing a car a plug it can't use. Flip the toggle above and watch the queues behave differently.

Read the paper

0 reviews scored

Eight million opinions about lunch

Python
XGBoost
Pandas
Scikit-learn
VADER

Sentiment and trend analysis across 8 million Yelp reviews from 200,000+ businesses. I pitted VADER against logistic regression and XGBoost, and the unglamorous rule-based one won at 71% — which was a useful lesson about reaching for the big model first.

Read the code

idle

Multi-agent research crew

Python
CrewAI
FastAPI
Tavily

Give it an industry and three specialised agents go to work in parallel — one on industry research, one on market analysis, one hunting for resources — then argue their findings into a single report with AI use cases and links to real datasets on Kaggle and Hugging Face.

Read the code

Also on the shelf: a Chrome extension for brainstellar.com puzzles that quietly picked up 40+ organic users on the Web Store, and a resume-aware interview question generator built during my internship at Zetheta Algorithms.

the stat sheet

Numbers I'd defend
in an interview

0Codeforces rating. Expert, and rank 675 globally in Round 185 Div. 2.

0LeetCode rating. Knight. Mostly built on evenings I should have been sleeping.

0Yelp reviews put through a sentiment pipeline, from 200,000+ businesses.

0Response-metadata points measured to catch a model being misleading.

0All-India rank in JEE Advanced 2022, out of more than a million.

0Participants in the hackathon I was invited back to judge, two years after competing in one.

how it went

Four years, in order

Electrical engineering degree, a hard left turn into AI safety somewhere around year two, and no plans to turn back.

2022

Got in

IIT (ISM) Dhanbad — B.Tech, Electrical Engineering

All-India rank 7100 in JEE Advanced, out of a million-plus candidates. Graduated in 2026 with a 7.55 CGPA and a much better answer to "what do you actually want to work on".
Jun – Dec 2024

Flew to London, won something

Apart Research — AI Safety Research Fellow & Lab Member

Took third place and a $300 prize at the Global AGI Deception Detection Hackathon, onsite in London, for a metadata-based approach to catching LLM deception. That turned into a six-month research fellowship run remotely from India, and a co-authored paper on LessWrong and the AI Alignment Forum.
Apr – Jun 2025

Shipped for someone else

Zetheta Algorithms — Software Engineer Intern

Built an interview-prep tool that reads a candidate's resume and target role and generates questions scaled to difficulty, including the parser that pulls skills and project history out of the document.
May 2025 – now

Shipped for myself

Grind IRL — founder and entire engineering department

A gamified productivity app, live and in use, grown to 30+ users purely through iteration and direct feedback. Every feature exists because somebody complained about something.
Jan 2026

Sat on the other side of the table

Apart Research — Judge, Global AI Manipulation Hackathon

Invited back to judge an international hackathon with 500+ participants and around 60 projects. Seven submissions assigned to me, seven scored against the rubric with written feedback, all in before the deadline. Winning projects went on to the AI Manipulation Workshop at IASEAI 2026.
Aug 2026 – present

Currently here

Sentient Futures — AI Safety Fellow

Ten weeks, one empirical question, mentored by researchers who have published a great deal more than I have. Meanwhile the EV charging optimiser from my final year is sitting in patent draft.