I build agents that audit models, and models that survive their own backtest.
Final-year MSc–MTech Dual Degree at IIT Jodhpur (Data & Computational Science, Mathematics). I work on agentic systems and quantitative research infrastructure — the kind where the interesting result is usually the one that fails honestly instead of the one that looks good.
role: AI Builder — agentic systems, LLM tooling
focus: ReAct agents · evaluation & auditing · quant backtest infrastructure
principle: "A number you cannot defend is a number you do not have."
building: Honest Mistake · Predictive Engine (strategy series)| Repo | What it is | Stars | Last push | Latest commit |
|---|---|---|---|---|
honest-mistake |
multi-layer ML audit agent | ⭐ 3 | 3 days ago | Merge pull request #2 from… |
predictive-engine |
quant backtest thesis | ⭐ 1 | 3 days ago | Merge pull request #2 from… |
regret-zero |
decision-regret inventory optimizer | ⭐ 3 | 2 months ago | Expand findings with sensitivity and… |
Auto-refreshed by a GitHub Action · last run 03 Sep 2026, 03:13 UTC
① Honest Mistake — a multi-layer ML audit agent
Public credit-risk models report 0.90+ ROC-AUC. Most of that is post-loan leakage.
I stripped 41 leakage columns, rebuilt on a true temporal holdout, and landed at an
honest 0.7296 AUC / 0.4404 PR-AUC with a generalization gap of -0.002.
Then I built an agent to audit the model automatically:
| Layer | What it does | Status |
|---|---|---|
| L1 — Honest baseline | Leakage-free feature set, temporal split, SHAP audit → 3 findings | ✅ Shipped |
| L2 — ReAct audit agent | Raw ReAct loop (no framework), 8 read-only tools, 4 verdicts, JSONL tracing, 2 ablation switches, pgvector semantic retrieval over the data dictionary | ✅ Shipped |
| L3 — Runtime tool generation | Capability-gap detection → tool-spec synthesis → sandboxed exec → known-answer validation → HITL checkpoint | ⬜ Not started |
Four scored ablation runs: the planted column was caught in every canary configuration, and what varied between configurations was the false-positive count, not whether the leak was found. Numbers and limitations in LAYER2_EVAL.md.
Python · Anthropic API · XGBoost · SHAP · raw ReAct, deliberately no framework
② Predictive Engine — quant backtest thesis
A pre-registered systematic strategy ported to NautilusTrader, tested for the ways it could be wrong.
- Nifty100: CAGR 25.36%, Sharpe 1.88 (verified port)
- MidCap150: CAGR ~27–29% — and the edge collapses when one name is removed. That finding stays in the repo.
- Survivorship-bias module:
survivorship.py, 50/50 tests, hard validation gate - Pre-registered experiments EXP18 / EXP20 / EXP21 / EXP22 — rejected. Negative results are logged, not deleted.
NautilusTrader · pandas · pre-registration · survivorship correction
⓷ RegretZero — decision-regret inventory optimizer · live demo (free-tier app sleeps when idle; ~20s to wake)
LightGBM quantile forecasting (P33 → P90, zero leakage) feeding a newsvendor optimizer framed around decision regret rather than forecast error. Measured +12.3% cost improvement. Streamlit cockpit is live.
LightGBM · quantile regression · newsvendor optimization · Streamlit
Languages & Core
AI / Agents
Infra & Tooling
