Agentic AI Engineer at Zetheta Algorithms, Bengaluru. I build AI products for finance, compliance and hiring, and I hold them to one rule: every claim in a README is a number CI can reproduce. Each project below ships with a benchmark, a held-out test it was never tuned on, and a quality gate that fails the build if scores drop.
| Project | Problem it solves | Measured result |
|---|---|---|
| InvoiceIQ | GST invoice checks and GSTR-2B reconciliation before claiming input tax credit | 100% extraction and error detection on 160 invoices · unseen layout: 0 wrong postings, 60/60 sent to LLM fallback |
| FairShortlist | Blind, explainable resume screening for recruiters | Ranking 100% vs 42% TF-IDF and 30% keyword ATS · swapping name and gender changes scores by 0 |
| ChurnShield | Which bank customers will leave, why, and who to call within budget | ROC-AUC 0.864 on held-out customers · €406,848 retained value vs €111,726 random targeting |
| Clausewise | HR and policy helpdesk that quotes the exact clause, says "not covered" instead of guessing | 0 HR-only leaks vs 16 for a typical RAG · held-out 82% vs 39% · version-aware |
| ClearMIS | Monthly MIS from a raw Shopify, Tally or Excel sales export, every number verified | Exact revenue where the raw export is up to 14% too high · right driver 12/12 months · no AI-written code runs |
| WorkBridge Compliance AI | Anti-money-laundering monitoring for Indian NBFCs, banks and payment companies | 128 alerts vs 35,143 from threshold rules · every planted case caught for the right reason · 46 of top 50 cases real |
| BalanceBook | Tax-aware portfolio rebalancing for Indian advisers, with adviser approval | Same risk control as band rebalancing · 50% less tax (78% on a held-out market) · 10-year backtest |
| TiltGuard | Risk desk for crypto prop firms and copy-trading: tilt, size spikes, overtrading, limit breaches | 92% of injected episodes caught vs 50% for fixed limits · 1/7 of the alerts · real Hyperliquid trades |
| ShelfWatch | Competitor price and stock monitoring for online sellers, every crawl checked before it is trusted | 3 false alerts vs 7,264 for the original ETL · 99% of real changes caught · Airflow + PostgreSQL |
| TickTrust | Crypto reference prices a payment gateway can safely quote on, with feed-integrity alerts | 0 bad quotes vs 2,870 from a single API on held-out data · 58/58 feed faults caught, 0 false alerts |
| ProofPost | Real customer reviews to compliant social posts every week (n8n + FastAPI) | 0 problem posts vs 606 with naive reposting · verbatim quotes · DPDP-aware redaction |
| RegLens | Cited answers to SEBI regulation questions | LangGraph planner, analyst and compliance checker · recall@4 100% · eval-gated releases |
| ArchitectAI | Plain-English requirements to a working FastAPI backend | 8/8 generated apps boot · 112/112 CRUD checks pass in CI |
| SupportPilot | Production ticket triage API | 1M records ingested in 9.4 s · PII redaction · API keys and rate limiting · on-call runbook |
| Practice | Why it matters |
|---|---|
| Benchmarks with a held-out set | A perfect score on data you tuned against proves nothing |
| CI quality gates | Regressions fail the build instead of reaching users |
| Fail closed | Uncertain outputs are blocked or sent for review, never silently accepted |
| Fairness tested, not asserted | Counterfactual swaps and segment metrics, reported in the README |
| Honest limitations | Every project says what it cannot do yet |
Open to AI Engineer, Agentic AI and GenAI Engineer roles · Bengaluru or remote · available now
LinkedIn · Email · Portfolio



