Skip to content
View f20220970-debug's full-sized avatar

Block or report f20220970-debug

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
f20220970-debug/README.md

Hi, I'm Pavitra 👋

Final-year dual-degree student at BITS Pilani (B.E. Mechanical + M.Sc. Chemistry, 2027), interning in Client Success & Program Delivery at Goodera, Bengaluru.

I measure things that run in production — processes, models and growth systems — with guardrails, evaluation and the failure modes written down. Most of what follows carries a finding that reversed my own first conclusion.

📫 f20220970@pilani.bits-pilani.ac.in · LinkedIn


Operations, risk and controls

Project Links What it found
process-transformation-lab Code · Results Process mining on a real 1,434-case workflow. One transition held 31% of all queue time — the obvious fix. But the clock decomposes to intake 38%, logged workflow 9%, closure 52%, so that bottleneck was worth 1.8 days of 59. Recommended instrumenting the unmeasured windows first.
card-portfolio-risk Code · Results SQL risk profiling of 30,000 real card accounts (180,000 account-months): roll rates, concentration, lead time, 6-sheet Excel model. A status field coded inconsistently across months made the book look like free fall; the comparable definition read as recovery. Both were the coding.
bread-credit-mlops Code · Results Model validation with promotion gates. Blocked a candidate that ranked fine and was wrong by 12.5 points on calibration. My own serving monitor missed a stuck upstream feature until it watched inputs as well as outputs: 3/4 → 4/4 faults caught, zero false alerts.
sc-transaction-monitoring Code · Results Fraud detection scored by alert budget, not accuracy (99.83% by flagging nothing) on 284,807 real card transactions. Predicted risk was 24× overstated before recalibration.
ozi-kids-qcommerce-analytics Code · Results 11 SQL analyses on 99K real orders. At equal delivery time, a missed promise lifts 1–2★ reviews from 8.8% to 25.3% — promise-kept beats speed as the KPI. Caught a NULL bug reporting 72% repeat rate against a true 1.9%.
finance-recon-controls Code Reconciliation controls over finance operations.
credit-risk-model-monitoring Code Scorecard performance, population stability and drift tracked over time, with the thresholds that should trigger a model review.
evaratus-fde-lab Code Field-level extraction evaluation on 2,000 noisy invoices: exposed a 19.4% document pass rate hiding behind a healthy-looking accuracy number, tracked to 99.8%.

Agents and experimentation

Project Links Highlight
chaos-liveops-agent Code · Daily eval Tunes a simulated puzzle game's difficulty: A/B tests each change, ships or rolls back inside retention and revenue guardrails, logs every decision. Re-evaluated daily on GitHub Actions.
llm-triage-agent Code · Daily eval Claims guardrail → Gemini → local fallback → confidence router. On unseen wording, 20% → 72% accuracy vs the local model. Refuses to publish when the LLM fails.
game-ua-budget-agent Code Thompson sampling on predicted D30 player value: +10.6% revenue vs even split; caught a bidding incident in one day.
retention-uplift-engine Code Uplift targeting 113 vs 69 extra orders per 1,000 against churn-risk targeting; Kaplan–Meier; propensity-matched loyalty ROI.
marketplace-pricing-experiments Code Switchback carryover bias (37%), elasticity with clustered SEs, CUPED (−87% variance), constrained price optimiser.
dynamic-pricing-bandits Code Shared-model Thompson sampling: 98.4% of oracle, 60% less regret than an A/B test.
dispatch-matching-sim Code Hungarian batching plus forecast-led repositioning lifted fulfilment 63.6% → 83.5% on the same 80 drivers (3-seed mean). Separately, a fleet-sizing test reached an 80% service level with 80 drivers against 160 under greedy dispatch (single seed), at 114 empty km per driver per day.

Engineering

Project Links Highlight
xr-3d-automation-pipeline Code · Results Mesh cleanup, unit detection and ML-driven LOD generation over real CAD assets: full-engine processing 179s → 52s; 31% fewer triangles than a fixed ratio. Found two validation defects in my own pipeline, including a glTF export mis-scaling an asset 1000×.
distributed-rl-rollouts Code Shared-memory actor–learner system: 1.9× throughput on 2 cores, worker recovery under 0.1s, and a measurement of the 30% of rollouts a strict on-policy learner would discard.
container-terminal-rl Code PPO written from scratch on a terminal-dispatch environment: matched or beat the best of four rules in all six scenarios, +7% on an unseen six-crane layout.
call-insights-api Code Queue-backed FastAPI service — Redis, MongoDB, retries, dead-letter, schema-validated JSON: 94% intent accuracy vs a 62% keyword baseline on 50 synthetic calls labelled before testing.
trip-data-pipeline Code ETL over 5.97M trips.

Business, growth and finance

Project Links Highlight
india-agri-carbon-market Code · Results Sized India's agri carbon-removal opportunity on a 44.16 Mha rice base. A cross-check against published farmer earnings broke my own model: a flat 50/50 split implied a developer share of −105% to 41% for rice methane — the credit is worth less than the payment it must fund. Screening six real programmes then broke two more assumptions, in opposite directions.
kaho-genz-growth Code · Results Positioning and acquisition arithmetic for a consumer app. The cheapest channel costs 3.2× less per install and 57% more per paying user, and breakeven sits 13% above the disclosed conversion rate — so the first move is sign-up drop-off (17.6% of CAC, no media spend), not more installs.
hyderabad-housing-analytics Code Price prediction on 2,464 real listings across 238 localities: gradient boosting at 6% median error (R² 0.91) against 10% for regression, which I kept as the interpretable benchmark. Clustered 41 localities into 5 segments and named the thin-supply premium pockets.
ev-market-sizing Code India E2W aftermarket sized bottom-up from Vahan data: ₹870 Cr FY26 → ₹5,600 Cr FY31 base case. Spend per vehicle moves FY31 by ₹2,200–3,000 Cr against ₹800–1,600 Cr for sales growth — so the lever is share of wallet, not market growth. Excel model CI-tested against a Python re-implementation.
soma-growth-console Code · App Founder planning tool: move growth levers and watch ARR, cash runway and Monte Carlo risk update.
arr-growth-model Code 24-month ARR and cash model in live Excel, matched to a Python implementation; Monte Carlo.
revenue-pipeline-forecaster Code Deal-level Monte Carlo within 3.8% against 83% error for stage-weighted CRM forecasts.
soma-brokerage-ops-lab Code · Memo Insurance brokerage operations: speed-to-lead, carrier routing, renewals, AI triage evaluation.
porter-category-growth-lab Code Pricing × fulfilment, partner acquisition and M1 retention, driver-support redesign.
qcommerce-retention-sql Code 9 SQL analyses on 3.35M Instacart orders, growth experiments, teardown.
noon-funnel-growth Code Funnel analysis and growth experiments.
eurekaforbes-service-insights Code Service analytics case study.
nobroker-strategy-case Code A 34× spread in modelled value per lead by journey — and ranking the same journeys by conversion rate inverts the order, so the metric a growth team reaches for first points at the lowest-value one. Rewrote my own FY24 recommendation after the company's FY26 revenue mix outdated it.
lng-configuration-screener Code Techno-economic screen of four LNG liquefaction configurations against four customer types: levelized cost of liquefaction, scenarios and sensitivities in live Excel.

Every repo states what its data is. Projects on real public datasets say which one; projects on simulated data say so and show method rather than company facts. Where a result rests on a single run or a small sample, the README says that too.

Tools: Python · SQL (CTEs, window functions) · pandas/NumPy/scikit-learn · Excel · FastAPI · Redis · Docker · GitHub Actions CI · pytest · Streamlit · Gemini API · Claude Code

Pinned Loading

  1. ai-benchmark-diligence ai-benchmark-diligence Public

    How random train/test splits overstate AI accuracy on chemistry data: R² falls from 0.70 to 0.09 on new molecular scaffolds. Includes a 5-point checklist for checking any AI startup’s accuracy claims.

    Python

  2. chaos-liveops-agent chaos-liveops-agent Public

    A LiveOps agent that finds high-churn levels daily, A/B tests easier versions and ships or rolls back on evidence, with spend caps, a retention guardrail and a full decision log. Re-evaluates itsel…

    Python

  3. india-ai-deal-tracker india-ai-deal-tracker Public

    Sourced database of Indian AI funding rounds, a weekly GitHub radar for India-based AI builders, SDK release-pace tracking, a market map, an investment memo and a technical diligence note on Composio.

    Python

  4. llm-triage-agent llm-triage-agent Public

    An LLM triage agent that re-evaluates itself daily on GitHub Actions against labelled messages, with rate-limit retries, a local fallback model and a rule that it refuses to publish when the LLM fa…

    Python

  5. retention-uplift-engine retention-uplift-engine Public

    Kaplan-Meier churn, T-learner uplift targeting, budgeted channel allocation and propensity-matched loyalty ROI on simulated customers.

    Python

  6. soma-growth-console soma-growth-console Public

    Streamlit planner for a brokerage founder's office: move drivers and operating levers and watch ARR, cash runway, unit economics and Monte Carlo plan risk update live. Illustrative inputs, not comp…

    Python