02 · Projects
Things I've built.
Each card says what the project does, what it achieved, and where the code is. Numbers come from the repositories' own evaluation runs.
AI agent · PaymentsRazorpay AI Buildathon · 2026
Wapas
A revenue-recovery agent for Razorpay merchants. It detects revenue at risk, diagnoses why a payment failed, and runs a bounded recovery workflow with deterministic policy checks before every action. Results are measured against a control group the agent never touches.
- 90.5% diagnosis accuracy on 5,000 simulated episodes
- 0 policy violations reaching an actuator · 0 of 33 red-team escapes
- In development · Razorpay test mode
- Python
- PostgreSQL
- Redis
- NVIDIA NIM
- Docker
AI agent · RAGHackerRank Orchestrate · May 2026
Sentinel
A support-triage agent for three product domains. A rule engine routes and escalates first, hybrid retrieval (FAISS plus BM25 with reciprocal rank fusion) finds the policy, Gemini drafts the reply, and an auditor strips any claim it cannot find in the corpus.
- Four-stage pipeline: triage, retrieve, generate, audit
- Prompt-injection and fraud triggers escalate in under 5 ms
- Python
- Gemini
- FAISS
- BM25
- Pydantic
AI evaluation · LinuxOpenEnv · Apr 2026
SRE Sandbox
An OpenEnv-compliant environment that breaks its own Linux container (nginx config, DNS, firewall, disk, permissions, cascading failures) and scores an AI agent on how it diagnoses and repairs the damage with bash.
- 8 scenarios from easy to cascade · dense regex-based rewards
- Command sanitiser with 17 blocked patterns · 90+ unit tests
- Python
- FastAPI
- Docker
- Bash
- Hugging Face
Machine learning · FintechKLU ML-02 Hackathon · Nov 2026
Payment Fraud Decision System
Scores every card or UPI transaction and decides approve, step-up OTP, or block, with cut-offs chosen by rupee cost rather than a default 0.5 threshold. The same pipeline finds strong signal in real card data and correctly finds none in a simulated UPI set.
- 0.796 PR-AUC on card data, 610× random · ROC-AUC 0.982
- 78% of fraud caught at the cost cut-off · 81% lower expected cost than doing nothing
- Python
- XGBoost
- Random Forest
- SMOTE
- Isolation Forest
Real-time web · OperationsJul 2026
Token Queue Management System
A visitor queue for offices and hospitals: reception issues thermal-printed tokens, cabin operators call and process them across configurable levels, and a public TV display announces tokens by voice. Admins manage users, cabins and sessions with live analytics.
- Socket.IO pushes every change to all screens instantly
- Thermal printing over raw TCP, USB or the browser via QZ Tray
- Next.js 16
- React 19
- PostgreSQL
- Prisma
- Socket.IO
- NextAuth
Machine learning · GeospatialFlipkart Round 2 · Jun 2026
Traffic Impact Forecaster Pro
Predicts the impact of traffic incidents in Bengaluru with an XGBoost classifier, enriched with live TomTom congestion and OpenWeatherMap visibility. An LLM incident-commander chat suggests resource deployment, and one click produces a PDF incident report.
- ~94% classification accuracy on incident impact
- Automatic model fallback between NVIDIA NIM endpoints on rate limits
- Python
- XGBoost
- Streamlit
- Folium
- NVIDIA NIM
Security · MLJan 2026
Mail & Link Authenticator
Phishing detection for URLs and email text. Structural URL features feed a Random Forest, TF-IDF email features feed Naive Bayes, and heuristic overrides make sure raw-IP links and "urgent password reset" patterns are never missed. Supports bulk and CSV scanning.
- Hybrid engine: ML verdicts with rule-based overrides
- Deployed frontend on Vercel, stateless FastAPI backend
- React
- TypeScript
- FastAPI
- scikit-learn
AI agent · Multi-agentAgentathon 2025 · GDG Hyderabad
Autonomous Marketing Agent for MSMEs
Finance, inventory, competitor and audit agents analyse a small business's data and recommend whether to launch a campaign, clear dead stock or hold cash. A crisis-simulation mode forces the strategy to pivot from growth to survival.
- Built and deployed to Google Cloud Run in 24 hours
- Append-only audit log of every AI decision
- Python
- Streamlit
- Gemini
- GCP Cloud Run
AI agent · ERP integrationMay 2026
ERPNext Telegram Bot
Lets warehouse staff say "moved 50 metres of red cotton from main store to the shop floor" in Telegram. Gemini function-calling maps the words to item codes and warehouses through FastMCP tools, checks stock, and creates the ERPNext entry only after a human confirms.
- Built for a textile business running ERPNext in the UAE
- Human-in-the-loop before every write
- Python
- Gemini
- FastMCP
- ERPNext REST
No projects match this filter.