I build AI systems, then I try to break them.
Final-year MSc–MTech in Data and Computational Science at IIT Jodhpur. I work on agents that audit models, models that admit when they're unsure, and a trading strategy I keep testing for the ways it could be fooling me.
The number each one actually earned.
A big number is easy to get. These are the ones that were left after I went looking for what was wrong with them.
Claude Sonnet 5 scored 0.6148 on the same 3,000 clauses, zero-shot, at 496× the cost. A fine-tuned DeBERTa on a CPU did the job for $0.00088 per thousand.
Most public models on this data report above 0.90. They read columns written after the loan closed. I removed 41 of them, then built an agent to find that kind of leak on its own.
A stock-ranking strategy run from Nifty 50 to Smallcap 250 after brokerage, slippage and capital-gains tax. On Nifty 100 and MidCap150, every rebalance is reconciled against a NautilusTrader port.
£54,631 saved over 12 held-out weeks by ordering for the cost of being wrong, not for forecast accuracy. Same forecasts, different decision.
Three habits I keep across projects.
Rules before results
The accept rule for an experiment gets written down before it runs. Reasonable Doubt's pre-registration runs past 6,000 lines, and when a rule turns out wrong, the fix goes in as a dated amendment, not an edit.
Failures stay in the record
Two of Reasonable Doubt's three planned tiers were dropped because they lost on measurement. Four pieces of Honest Mistake's third layer were cut and the reasons are in the repo. Predictive Engine keeps every rejected idea next to the one that passed.
Other people can check it
Each project is public. Three are live, and the fourth publishes its code, experiment record and results, though a rerun needs 396 MB of vendor price data that isn't in the repo. Honest Mistake lets you run the audit yourself with your own API key, and RegretZero's dashboard imports the same module that produced its README numbers.