Full-stack engineer working on GenAI applications. B.Tech in Computer Science and Artificial Intelligence at Newton School of Technology, Rishihood University (2024–2028).
I care about the unglamorous half of AI engineering — making a model's output checkable. Most of what I build ends up being a verification layer wrapped around an LLM, plus the benchmark that proves it works.
GenAI Trainer — Octa Learning · Aug 2025 – present Teaching applied generative AI to 20+ students: LLMs, retrieval-augmented generation and prompt engineering. Designed the hands-on projects where learners build and integrate real AI-powered applications.
Full-Stack Developer, freelance · Jan 2025 – present Delivered production web applications for 10+ clients, most of them live with real users. Backend systems and REST APIs, deployed end to end — owned from requirements through to keeping them running.
Software Engineer Intern — Klariti Learning Innovations · May 2025 – Jul 2025 Built full-stack features on a React and Node.js ed-tech platform used by students and instructors. Shipped REST APIs for user accounts and course workflows serving thousands of requests a day, and cut frontend load time through component-level optimisation.
verbatim — meeting intelligence for Hinglish calls
Otter, Fireflies and Fathom all fall apart on Hindi–English code-switching, and independent testing has found hallucinated content in a meaningful share of AI-generated meeting summaries. verbatim takes the opposite stance: cite-or-drop. Every minute, decision and action item carries a timestamped citation back to the transcript — and any claim it cannot ground gets deleted rather than guessed.
precision 100% ← zero invented claims on the adversarial set
recall 91–93%
grounding 100% ← every surviving claim carries a citation
cost ~$0.08 / meeting-hour, self-hosted
Groq Whisper v3-turbo for transcription · gpt-oss-120b for extraction · sherpa-onnx for local speaker diarization · Qdrant + fastembed for cross-meeting RAG · RapidFuzz to validate every citation · Playwright bot that auto-joins Google Meet. Runs on your own machine, so the audio never leaves it.
Python FastAPI Pydantic Qdrant Playwright · MIT
Gauntlet — does a prompt-injection defense survive attacks it has never seen?
Most detectors report 99% on the distribution they trained on. Gauntlet holds whole attack families out of training, evolves new attacks with a 21-operator mutation search gated by a victim oracle, and then checks the result against 662 real human-written injections. Five defenders benchmarked at 3 seeds each; plug your own detector in over HTTP with bench. Runs on a laptop CPU, no API keys. The finding: canonicalisation plus sentence windowing took real-data AUC from 0.56 to 0.81, and swapping TF-IDF for a fine-tuned transformer did not.
Python scikit-learn PyTorch (optional) · 63 tests, CI on 3.11 and 3.13
Hindsight — financial news intelligence that grades itself
Ingests 14 live free sources (RSS, GDELT, regulators), dedupes them into events, scores market impact with grounded LLM analysis, and seals each directional call before the outcome exists, so accuracy is measured against real price moves rather than asserted. Every claim traces back to a source URL. Read-only FastAPI API, a six-view Next.js frontend, and scheduled recording through GitHub Actions.
Python FastAPI Next.js uv · 287 tests
Nova — a visual website builder for the browser
Drag blocks onto a canvas, style them in an inspector, and export a real, dependency-free site as HTML or React. Built as a full editor rather than a demo: 46 blocks, 24 starter templates, responsive breakpoints, reusable components, undo/redo, a command palette, accounts, one-click publishing, and form submissions collected in a real backend.
React TypeScript Vite · MIT
Witness — bug report in, verified reproduction out
Takes a vague, human-written bug report and drives a real browser to reproduce it, claiming CONFIRMED only after deterministic checks trip on three clean re-runs. Precision is gated at 100% in the eval harness; recall is reported honestly. Ships with Kirana Cart, a storefront seeded with 22 bugs and 3 decoys, and a lint that fails the build if a model import appears anywhere near the oracles.
TypeScript Playwright Vite
ClipStudio — a video editor that runs entirely in the browser
Multi-track timeline, eased keyframes on any transform, chroma key with spill control, colour grading, transitions, an offline audio mixdown and a real MP4 encoded on your machine through WebCodecs. Captions are transcribed locally with Whisper via transformers.js, so the media never leaves the tab. Accounts, saved projects, and shareable watch pages with comments.
Next.js React TypeScript WebCodecs Prisma
ArchForge — plain English → production system architecture
archforge.vercel.app · live
Type "a ride-sharing platform", get a verified 15–20 component architecture with capacity math (QPS, storage, bandwidth), per-component reasoning, and a Well-Architected review across all six AWS pillars. Three deterministic verification layers sit between the model and the canvas, so it cannot emit a structurally invalid diagram. Benchmarked at 98% recall against the real published architectures of Uber, Instagram and Netflix. 75 tests. Exports to PNG, Mermaid or Markdown.
React Vite React Flow Node Express elkjs · Groq / Cerebras / Gemini with automatic failover
easycron — cron that survives free-tier hosting
npm i -g easycron-cli · v3.0.2 published
Render, Railway and Fly.io put free instances to sleep and kill background jobs, which quietly breaks every in-process scheduler. easycron takes a schedule in plain English, compiles it to cron, and wires up an external trigger — GitHub Actions, UptimeRobot or cron-job.org — so the job actually fires. Plugin system with regex-safety validation, configurable retry backoff, execution logs, and endpoint scaffolding for Express and Fastify. The README documents the real limits too: the 60-day Actions sleep and the 2,000-minute monthly budget.
Node.js npm GitHub Actions
Email SDK — a delivery pipeline built to fail well
Multi-provider email delivery across AWS SES, SMTP and SendGrid, written to make the resilience patterns explicit rather than hidden: per-provider circuit breakers, token-bucket rate limiting, exponential-backoff retries, a priority queue holding 10,000 jobs drained by five concurrent workers, a dead-letter queue for what still fails, and event-based observability throughout. Documented with sequence, activity, class and use-case diagrams.
TypeScript Vitest nodemailer Handlebars
Solar forecasting + agentic grid optimisation — ML forecast of solar DC output, feeding a retrieval agent that returns structured grid recommendations · CryptoVision — crypto analytics dashboard with a live market ticker · FairMarket India — systems-thinking study of small sellers on Indian e-commerce, with causal loop and stock-flow models · geo-panorama-viewer — 360° panoramas on a map, exportable as GeoJSON · HomeScope360 — real-estate platform with 360° virtual property tours
Active contributor across the Apache Foundation and other projects — merged and in-review, all verifiable via GitHub.
| Project | Contributions |
|---|---|
| Apache DataFusion — Rust query engine | 11 merged, 3 in review · Substrait query-plan serialization (output types on aggregate and window functions, LIKE encoding, CASE equality conditions, intersection nullability), decorrelation of grouping sets with correlated filters, DISTINCT aggregate arity, date_trunc on scalar timestamps, overflow and underflow in vector functions, and CI wiring for the Substrait check job |
| Apache Airflow — workflow orchestration | 4 merged, 6 in review · stopped the FAB app being rebuilt on every users/roles API call, fixed SFTPHookAsync letting a connection's known_hosts override an explicit path, deferred EKS-cluster-deletion logging, Google OperationHelper test coverage; in review: failing a task whose execute is interrupted by SIGTERM, scheduler session handling for executor events, rendered template fields for start-from-trigger tasks |
| RUXAILAB | 6 merged, 1 in review · DOMPurify sanitisation of rich-text v-html sinks, an RBAC read check on analytics subcollections, auth redirect for signed-out visitors, locale-aware i18n date formatting, a cross-client email-rendering fix, a dead-route fix in heuristic analytics |
| Kolibri — Learning Equality | 3 merged, 2 in review · fixed a crash on resources with more than one learning activity, surfaced every activity on coach lesson cards, accepted Content-Range responses with an unknown total length |
| AsyncAPI | 2 merged · made the Case Studies page table mobile-responsive, cleaned up every stale link in CONTRIBUTING.md |
| Tin — open-source marketing system | 1 merged · an outreach speaking-shortlist workflow package |
27 merged pull requests across 6 organizations, 12 more in review. Full history →
Languages JavaScript · TypeScript · Python · Rust (DataFusion contributions)
AI / ML LLMs · RAG · prompt engineering · vector databases · PyTorch · eval harnesses and grounding checks
Frontend React · Next.js · Vite · Tailwind · React Flow
Backend Node.js · Express · FastAPI · Pydantic · REST APIs · microservices · queues, retries, circuit breakers, rate limiting
Data PostgreSQL · MongoDB · Qdrant · NumPy · Pandas
Concepts Distributed systems · API design · caching · async processing
Tooling Git · GitHub Actions · Docker · Vercel · Postman · Playwright · Vitest
Open to internships and to collaborating on open source.
namanjainpy@gmail.com · Portfolio · Résumé · LinkedIn



