Instruments of resolution
Noisegoes in.Signalcomes out.
A repository. A Postgres backup. A bank export. A lab PDF. A git history. Nine systems that take input nobody vetted — and refuse to guess about the parts they can't resolve.
01 / the ratchet
carabiner
A security baseline that only tightens.
- Problem
- Repository posture rots silently between audits.
- Hard
- A ratchet that never loosens — not by accident, not by a well-meaning PR.
- Cost
- 60 repos calibrated · 83 tests · 18 drills against live GitHub settings · SARIF · a pre-commit hook · PyPI, GHCR, Marketplace
- Found
- Run on 40 popular repos, its build-failing findings went from 5 to 2 — both real.
carabiner fixpins every action to its commit: 226 in grafana alone.
02 / the vessel
firedrill
A backup is not proven until something restores it.
- Problem
- A
pg_dumpthat exits 0 can still be a truncated file nobody has ever tried to read back — andpsqlrestores a truncated plain dump with exit 0, rows silently missing. - Hard
- Matching the archive's own database version and restoring inside a disposable container with no path to production, ever — including point-in-time recovery from pgBackRest and WAL-G, to the second.
- Cost
- 6 Postgres majors, both directions on the corpus · MySQL and MariaDB · SQLite with no container at all · Slack and Discord alerts · PyPI, GHCR, GitHub Action
- Found
- pagila's published dump says Postgres 12 and needs 18: restored as labelled, 140 statements fail and
psqlstill exits 0. 64 bytes zeroed in one pgBackRest file: caught.
03 / the rotor
recur
Which charges actually recur.
- Problem
- A bank export is a wall of rows. The subscriptions are in there somewhere.
- Hard
- Postgres row-level security, so a query that forgets its tenant filter returns zero rows instead of someone else's money.
- Cost
- FastAPI · SQLAlchemy · argon2id · OAuth 2.1 + PKCE · remote MCP server · pgvector matching, measured and cut · 0 bank credentials stored, ever
04 / the spectrometer
LabLedger
A lab PDF, resolved to codes.
- Problem
- Every lab formats results differently. The numbers stop meaning anything.
- Hard
- Resolving each test to a LOINC code, converting units, and picking the right reference interval — an LLM only for the ambiguous ones.
- Cost
- FastAPI · MongoDB · role-based access, 2FA, audit log · FHIR R4 export · one-command demo · anything uncertain goes to a human
- Found
- Checking 88 real report rows by hand found a positive COVID result read as negative. Now every row is a test: 0 values wrong, 0 tests matched wrong, 7 left for a person.
05 / the tumbler
QuizNest
Knowledge either aligns, or it doesn't.
- Problem
- Reading something is not knowing it, and a score out of ten doesn't tell you which part you missed.
- Hard
- Generating questions that actually discriminate, then turning a stream of answers into analytics a learner can act on.
- Cost
- MERN · Gemini-generated questions and graded written answers · Google sign-in · per-topic analytics
06 / the manifold
AI-Recipe-Maker
Five measured inputs, one plan.
- Problem
- "What can I cook with this?" is easy to answer badly and hard to answer with real nutrition behind it.
- Hard
- Dosing the constraints — what's in the fridge, what it costs nutritionally, what a week of it looks like — into one coherent plan.
- Cost
- TypeScript · Claude-powered generation · nutrition analysis · meal plans
07 / the prize wheel
Snazzlebop
A game show where your friends are the contestants.
- Problem
- Party games online mean an install, an account each, or one person sharing a screen while everyone else watches.
- Hard
- Server-authoritative engines: each phone only ever receives what its player may see. Your dice, the bomb's fuse, who is secretly the House — none of it reaches another device before the reveal, and a deploy doesn't end a game in progress.
- Cost
- 37 games · FastAPI + WebSockets · React + TypeScript · show nights, an audience of up to 30, a Discord Activity · 1,046 backend tests
08 / the selector
AniNest
Every page answers, even with a source down.
- Problem
- An anime and manga hub built on free public APIs — which rate-limit, and go down.
- Hard
- AniList first, Jikan when it fails, and the last good copy from its own database when both are down. One dead upstream greys out one panel, never the page.
- Cost
- Vanilla JS · Express · SQLite on Turso · anime, manga and novels · official free episodes · games, XP and alerts · CSRF, 2FA, strict CSP · Playwright plays every game
09 / the city
afterglow
A repository's history, grown into a night city.
- Problem
- The parts of a codebase that matter — hotspots, the bus factor, code nobody has touched in years — are invisible in a file tree.
- Hard
- Reading any public repository's history safely: a sandboxed worker, a bare blobless clone that never checks out a file, network egress to github.com only, and no email address ever kept.
- Cost
- FastAPI · a Postgres job queue · a WebGL2 city that also works as a plain table · CodeQL, Semgrep, Trivy and gitleaks on every PR · GHCR images with build provenance
10 / the core sample
Wizcommerce
Backend in production, on a multi-tenant platform.
Backend DeveloperGurgaon · Apr 2025 — Sep 2026
- 30+ REST APIs on FastAPI — 30% faster data retrieval
- A multi-tenant backend: each client's data private and isolated, and the system still easy to scale
- Search on Elasticsearch — response time down ~40%
- Long-running work moved onto Celery, so the API stays fast and responsive for users
- PostgreSQL queries and SQLAlchemy code reworked — 25% less database load
- Centralized logging and Sentry alerts, so issues are caught and fixed faster
11 / the feeler gauge
What I reach for
Seven leaves, each ground to one thickness.
- Language
- Python · SQL · TypeScript · JavaScript (React, Node.js) · Java
- Backend
- FastAPI · REST · JWT · OAuth2
- GenAI
- RAG · embeddings (pgvector) · LLM evaluation · prompt engineering · OpenAI, Gemini & Claude APIs · MCP · Claude Code hooks, Cursor, Copilot
- Data
- PostgreSQL · MongoDB · SQLAlchemy
- Search
- Elasticsearch · Redis
- Ship
- Celery · Docker · Kubernetes · Git · CI/CD · Sentry
- Judgement
- Data structures & algorithms · system design · API design · performance tuning · debugging
B.Tech Computer Science Engineering, Chitkara University, 2022–2026 — CGPA 9.02. Certificates in deep learning, data science and product management.
12 / the rack
More, on GitHub
Nine above are one system each, worth their own mechanism. These share a rack instead — real, tested, published, each with its own CI, read at a glance rather than one at a time.
- receipt — verified receipts for what a shell command, a browser click or an AI coding agent actually did — one hook for Claude Code, Codex and Cursor — hash-chained and signed with an SSH key like a git commit
- invariant — declare what must stay true, check it in CI with its own GitHub Action, keep a proof bundle of every run
- playhead — replay an AI coding session like a video timeline, imported straight from its transcript
- satchel — a local-first reading archive with a one-click browser extension; imports Pocket and Omnivore, tags included, and rescues dead links from the Wayback Machine
- sourced — checks each claim of LLM output against its source, measured on 20,261 real LLM sentences from RAGTruth; an optional Claude judge must quote its evidence verbatim
- lockstep — checks what is really installed — on a machine, in a container image or an SBOM — against the lockfile, uv.lock and npm's package-lock.json included, and catches code edited after install
- escrow — finds scheduled jobs that quietly stopped, by cron schedule or across a whole GitHub org or Kubernetes cluster with no setup, with a self-hosted status page; one pydantic workflow had failed all 118 scheduled runs on record
- drift — checks a device's real traffic, HTTPS included, against its maker's own RFC 8520 MUD file, from a capture or live on the network
- witness — a browser extension that holds a site to the claims it publishes about itself
Let's buildsomethingthat holds up.
Fremont, California · Green Card holder, authorized to work in the U.S.
How this page works
One object, never cut. Every boundary is the shape before it turning into the shape after. Pick one to watch it happen again.