Melove Gupta

Software Engineer•

00:00:00IST

I’m a CS undergrad student who spends most of his free time inside codebases that I didn’t write. Over the past year that’s meant a floating-point bug in LibreOffice, an instrumentation watcher for CNCF’s OpenTelemetry organization, a Bitcoin block parser built from raw bytes, and a citation-grounded RAG system for Indian Law.

What I am good at right now is making systems technically correct, parsers that handle edge cases, retrieval pipelines that don’t hallucinate citations, code that survives review from people who’ve maintained a project for a decade.

Experience

CNCF OpenTelemetry Ecosystem Explorerlink

Open Source ContributorMay 2026 - Present

Fixed a Faro SDK defect that unconditionally routed telemetry to a hardcoded production collector during local development, eliminating CORS failures across every application route.
Built a JavaScript instrumentation watcher from scratch, mirroring the existing Java and Collector watcher architectures to produce 46 registry entries, iterated through 16 rounds of maintainer review before merge.
Diagnosed a component-card overflow on mobile viewports via DevTools, traced it to an unconstrained grid layout, and shipped a targeted fix.
Authored contributor onboarding docs covering watcher architecture, CI wiring, merged 22+ PRs total while collaborating directly with CNCF maintainers at Grafana Labs.

Other Open Source Work

LibreOfficelink

Open Source Contributor

Fixed tdf#55960, a Calc bug open since 2012 where linear-sequence auto-fill accumulated IEEE 754 floating-point drift in nStartVal + nStepValue * nIndex (e.g., returning 0.6999999999999996 instead of 0.7) applied rtl::math::approxValue() to snap computed values to expected results and added a regression unit test, merged across two files after seven patch sets and later cherry-picked into a downstream release branch.
Diagnosed cross-component regressions across Writer, Calc, and Impress by tracing execution paths through the rendering engine, UI toolkit, and document model before scoping and submitting fixes.

In Between These Experiences

Projects

Chain Lens - Bitcoin Transaction & Block Visualizer link

Implemented a low-level Bitcoin parser for SegWit (BIP-141), Taproot (BIP-341), and legacy P2PKH/P2SH/P2WPKH/P2WSH script types by decoding raw blockchain binary data with no external APIs.
Validated block integrity via Merkle tree construction and root-hash verification against block headers, exposed the same parsing engine through a CLI, REST API, and web visualizer.

Pramaan - Citation-Grounded RAG System for Indian Law link

Built a hybrid retrieval pipeline (BM25 + dense vector search via BAAI/bge-large-en-v1.5) fused with Reciprocal Rank Fusion and cross-encoder reranking, grounding LLM responses in citation-verifiable chunks across a 20-document corpus of Indian statutes.
Designed a 122-question evaluation harness benchmarking 4 LLMs on answer rate, refusal precision, and citation accuracy, wired into a GitHub Actions pipeline that gates merges on quality regression.

Safe PSBT Transaction Builder link

Built a BIP-174 (PSBT)-compliant transaction builder with coin selection, RBF, and locktime support using bitcoinjs-lib, wrote 15+ unit tests covering input/output construction, signing workflows, and serialization correctness.

Card Drawer - Web3 dApp link

Architected a full-stack dApp in Next.js/TypeScript with a Solidity smart contract on the Flare Coston2 testnet - wallet connection, on-chain transaction dispatch, and reactive UI state over asynchronous confirmation flows.

Work in Progress

Parent Copilot for Autism

Professional autism support is expensive, scarce, and confined to sessions. But the situations parents actually need help with are things like a meltdown before school, a routine that suddenly stops working, or a new environment the child can’t process. The parent is often the most important person in the whole support system, and the one with the least continuous support.

Most tools in this space treat behavior tracking, visual schedules, and parent guidance as three separate apps. I think there should be 3 ways of interacting with one longitudinal record of a child.

I'm working on a loop where a behavior entry gives the Copilot context, the Copilot surfaces a pattern, and the parent adjusts a routine based on that.