About Me
I'm Kundan Singh. I lead engineering at DCRefers, an initiative of Open Law Library. I build backends, APIs, write algorithms, manage infrastructure, and ship LLM systems. Mainly Python, sometimes NodeJS, full-stack if needed.
🌴 In my free time: LLM/evals experiments, growing vegetables, or reading sociology
💼 I take on work in applied AI, LLM systems, and Python backends — US hours
🤝 Questions, ideas, or work — email me: hi@kundansingh.me
👨💻 Things I've Built
DCRefers
DCRefers is a legal-tech platform built using Python/Django, connecting modest-means citizens in Washington DC, USA with qualified attorneys and mediators. I've been the sole developer for four years and counting — architecture, backend, infrastructure, and frontend. I built the recommendation & matching engine, onboarding flows, state machine architecture for notification pipelines, and an LLM based categorizer with PII redaction ahead of inference. I also designed a configurable BPMN workflow engine (celery/joeflow) that extends as workflow complexity grows, with a pluggable architecture exposing apps through an AppConfig registry. Integrated a11y to ensure WCAG compliance and i18n/l10n support, and currently building UIs for data and analytics.
Lowbono
Lowbono is an open-source PyPI library — the Python/Django based referral engine extracted from DCRefers, generalized for any pro-bono or low-fee professional-referral portal. Referrals, legal practice-area taxonomy, income-eligibility screening, admin-configurable notification engine on a Celery/Joeflow workflow base. Its optional LLM categorizer runs PII redaction using Presidio, spaCy ahead of inference, classifies queries against a fine-tuned US legal taxonomy, and instruments failure modes for evaluation. Explicit contract over convention — easily plug own professional types via AppConfig contract, and everything is overridable without forking.
Parallax
Parallax is an applied LLM system that reads a social phenomenon through several theoretical paradigms simultaneously and surfaces where their concepts collide. LLMs generate quality content around sociological issues, but fails when comparing multiple paradigms or contention occurs on a phenomenon. I built an architecture that bifurcates the system: constrained generation producing schema-validated property graph per paradigm, each under its own ontology, then collision detection as a separate deterministic stage — so nothing generated reaches the render path unvalidated. Try cancel culture, ozempic use, global governance.
SapiensInk
SapiensInk is a curated list of the best human writing on the internet — essays, books, threads, papers. Here is why I made this, and why it matters. Suggest a reading, if you've found something worth sharing. Also: no signups, no tracking, Pagefind search, and a single JSON file as a public API to build on.
Mindmaps
Mindmaps — an offline-first, privacy-focused, mindmaps app. ReactJS & TypeScript, six production dependencies (two are fonts), no network requests after load — enforced by a build check, not a README promise. Open-source, hand-crafted tree layout engine, ~600 tests, IndexedDB behind an injected adapter, mobile responsive, full keyboard control, easy backups (JSON or Markdown).
🛠️ Experiments & Hacks
EasyPrompts Extension
A browser extension for ChatGPT, Claude, Gemini, etc., that manages your frequently-used prompts and auto-pastes them. It also has an Airtable integration that syncs your prompts across all devices.
What Can I Do?
What Can I Do? — a decision-first weather app. A custom hourly, daily and weekly scoring engine for 8 activities like running, stargazing, photography, etc. and best-time available to do them. Pure static, no backend, no keys, privacy-focused, powered by Open-Meteo.
WhatsApp Web Mutation
archived
Before WhatsApp allowed APIs officially, I wrote custom APIs to automate several business use-cases using the MutationObserver interface. No longer public, but makes for an interesting hacking story.