I'm a final-year Electronics and Communication Engineering student at Delhi Technological University (Class of 2027). I build software: AI agents and agentic workflows, and full-stack web and mobile apps in React, TypeScript, React Native and Expo.
I can't sit idle, so I build things to learn them. That's why my projects are varied: an agent that verifies its own citations, a finance app in Expo, a research workbench in React, a hybrid RAG system. Each one is a way to try something new, and each one gets tests, measurements and an honest write-up.
| Area | Details |
|---|---|
| Target roles | Software Engineer, AI Engineer |
| Education | B.Tech in Electronics and Communication Engineering, DTU, 2023 to 2027 |
| Experience | Software Engineer Intern at Globe Capacitors Pvt. Ltd. (Jun to Jul 2025) |
| Building now | AI agents, RAG systems, and full-stack web and mobile apps |
| Reach me | yadavkapish22@gmail.com |
Experience in detail: Software Engineer Intern, Globe Capacitors Pvt. Ltd.
Built backend automation and cloud integrations connecting the company's enterprise ERP platform, relational databases and Slack for automated reporting and alerts on AWS.
- Cut manual administrative effort by 40% across 3 operations teams through automated tracking and notifications
- Diagnosed and resolved 30+ production defects on GCP infrastructure, reducing average load times by 25% for 100+ internal users
- Automated approval routing and governance workflows with Python backend APIs, raising approval completion rates by 15%
Click a project to expand the details.
OpsPod · agentic dispute resolution: the LLM proposes, a policy engine decides
- About 90% of cases are decided by deterministic rules with no LLM call, and the model never supplies a monetary amount
- PII is tokenized before any LLM call, and 0 prompt-injection cases caused an action in the evaluation gate
- Idempotent ledger with reversible actions, replayable audit records, and an approval queue for human review
- 400+ tests including chaos and crash-resume suites, with OpenTelemetry, Prometheus and Grafana
Finio · personal finance app in React Native (Expo) with a FastAPI backend
- Mobile app built with Expo Router and strict TypeScript, with Zustand and AsyncStorage for a local-first experience that needs no account
- Home, transactions, insights, and goals screens, including spending charts and a no-spend streak tracker
- Monochrome design system with light and dark variants, specified in Figma
- Monorepo with a FastAPI and MongoDB service, and three GitHub Actions workflows for lint, type-check, tests and Docker image publishing
Scout · durable, citation-verified multi-agent research assistant
- LangGraph workflow (planner, researcher, supervisor, writer, citation validator) that checkpoints to Redis and resumes after a crash
- Per-run and daily budgets on time, model calls, tokens, searches and page fetches
- Every citation is checked against the fetched page text before the report is published
- Live progress over Server-Sent Events, with a Rich CLI and a REST API, running on free-tier models
TabNet Studio · PyTorch research workbench with a React and TypeScript frontend
- TabNet implemented from the paper's equations in PyTorch, including Sparsemax attention, Ghost Batch Normalization and GLU feature transformers, with no library wrapper
- Interactive workbench for hyperparameter tuning, live training curves and Sparsemax attention heatmaps
- FastAPI backend, React and TypeScript frontend, packaged with Docker
GraphMind · hybrid knowledge-graph and vector RAG for SEC 10-K filings
- A router chooses graph, vector or hybrid retrieval, and hybrid results are fused with Reciprocal Rank Fusion
- Request-scoped evidence registry: every citation must resolve to evidence retrieved for that request
- On a 100-question benchmark, hybrid gained +15 points over a vector-only baseline on 2-hop and 3-hop questions (the overall +4 is not statistically significant, and the README says so)
- FastAPI, React, a terminal CLI, and 151 automated tests
SeatLock · ticket booking backend that can't oversell during a flash sale
- Atomic Redis holds, Postgres row locks and a unique-index backstop so a seat can't be double-booked
- Idempotent checkout, HMAC-signed payment webhooks, and an SQS notifier with retries and dead-lettering
- Three FastAPI services behind an Nginx gateway, with Docker Compose, Kubernetes manifests and 101 automated tests
Arbor-DB · relational database engine from scratch in C++17
- Slotted-page storage, LRU-K buffer pool, B+Tree indexes, ARIES-style recovery tested with a real kill -9, and strict 2PL or MVCC
- Hand-written SQL parser, optimizer and Volcano executor, verified under ASan, UBSan and ThreadSanitizer
Expand: six ML and quant projects, each starting from a question
| Project | Question | Result |
|---|---|---|
| Stock-Return-Predictor | Can ML predict next-day US equity returns without leakage? | No tradable edge after costs |
| Earnings-call-sentiment | Does call tone explain abnormal returns? | Strong same-day link (t = 5.1, N = 2,050); FinBERT adds nothing beyond LM |
| Neural-Vol-Surface | Can neural methods calibrate a real SPY vol surface? | Neural pricer calibrates Heston 5.6x faster |
| FraudLens | Can fraud be caught with cost-aware thresholds? | PR-AUC 0.749, 36 ms p95 API latency |
| ChurnGuard | Can churn be predicted, explained and acted on? | PR-AUC 0.673, ROC-AUC 0.842 |
| ReviewLens | Can reviews become product decisions? | DistilBERT beats TF-IDF, macro-F1 0.695 vs 0.658 |
| Habit | What it looks like |
|---|---|
| Learn by building | A new idea becomes a working project, not a bookmark |
| Tested | Property, concurrency, chaos and crash-recovery tests, not just happy paths |
| Measured | Evals with confidence intervals, held-out sets and pre-declared primary metrics |
| Honest | Every project documents its limitations, including results that didn't hold up |



