Full Stack Software Engineer at Adelev8 in Bengaluru, India.

I build diagnostic tooling and CLIs for people shipping with AI agents: token cost meters, agent benchmarks, fresh-clone replay, instruction-file audits.

That's a real npx ctxvitals run on my own machine: 886 skills and 14 MCP servers spend 138,940 tokens before the model reads a word. Grade: F. That number is why these tools exist.

• ctxvitals (npx ctxvitals) - a Lighthouse for your AI agent stack: measures the real token cost of MCP servers, skills, and rules; flags security issues and duplicates; grades the setup A-F.
• compactsim (npx compactsim) - replays local Claude Code and Codex transcripts to measure Context Rent (payload_tokens * turns_carried), catch post-compaction amnesia re-reads, and simulate autoCompactWindow thresholds before you edit settings.json.
• agent-trials (npx agent-trials) - run the same JSON task suite through every coding agent CLI you have installed (Claude Code, Codex, Gemini) in sandboxed repo copies; scoreboard of who passed, how long they took, what it cost.
• ruleherd (npx ruleherd) - audits CLAUDE.md, AGENTS.md, .cursor/rules and friends for contradictions, dead file references, and copy-pasted rules that drifted apart. Also in Rust: ruleherd-rs.
• mcp-tax (npx mcp-tax-cli) - how many tokens your MCP servers burn before you type a word. Also in Rust: mcp-tax-rs.
• mcp-bouncer (npx mcp-bouncer) - policy-enforcement proxy for stdio MCP servers: tool allowlists, method gating, path traversal and SSRF blocking, rate limits, JSONL audit logs.
• agentblast (npx agentblast) - two-plane blast-radius auditor for Claude Code and Codex CLI: checks ~/.claude/settings.json and ~/.codex/config.toml for plaintext API keys and missing deny rules, replays session transcripts to catch unrestricted ssh root@ calls, non-/tmp rm -rf, and .env reads, and emits least-privilege permission policies.
• egressleak (npx egressleak) - outbound HTTP egress, User-Agent email leak (#107757), and context PII auditor for Claude Code, Codex CLI, and MCP configs: separates synthetic noreply@anthropic.com trailer emails and 127.0.0.1 loopback calls from real external domain egress and ?authuser= query leaks, and emits a deterministic PreToolUse guardrail hook.
• repotrap (npx repotrap) - prompt-injection, invisible Unicode steganography (U+E0000 tags, Bidi overrides, standalone U+200B), spoofed <system-reminder> delimiter, and .claude/.mcp.json/.vscode repo-open execution trap scanner for skills, cloned repos, session transcripts, and untrusted GitHub issues (--gh-issue owner/repo#123).
• loopbrake (npx loopbrake) - coding-agent edit-thrash, shell-write bypass (39.2% of source writes across 153 transcripts via Bash heredoc/sed/.write_text()), deleted-sandbox CWD escape, and consecutive tool-error doom-loop auditor + PreToolUse/PostToolUse circuit-breaker hook generator.
• clonesieve (npx clonesieve) - codebase structural function clone (Type-2 normalized + 4-gram Jaccard shingle), cross-file duplicate helper name (370 duplicate names & 117 clone groups across 3,300 files), and zero-caller dead-export auditor + Claude Code and Codex blind-write transcript scanner (39.0% of new functions written before any Grep/rg search).
• canimcp (npx canimcp) - preflight checker for MCP servers: which clients silently drop or mangle your tools, checked against a versioned quirks corpus (Claude Code, Cursor, Codex, Windsurf, Zed, Continue).

• coldstart (npx coldstart-cli) - replays your README's install instructions in a clean clone and reports the exact step where a new user hits a wall.
• readmeroast (npx readmeroast you/yourrepo) - scores any GitHub README out of 100 and roasts what it's missing.
• slopsieve (npx slopsieve) - score a diff or PR for 15 AI-slop review signals: test weakening, secret leaks, phantom imports, debug leftovers, dead exports. Also in Rust: slopsieve-rs.
• uislop (npx uislop) - static linter for React, Next.js, Vue, Svelte, and Tailwind that scores frontend files for 16 "vibecoded" AI UI tells (indigo/violet palette dominance, hardcoded Unsplash stock avatars, fake "10,000+ developers" social proof, Lucide icon-in-tinted-box cards, and dead href="#" links).
• dep-census - who actually maintains the code inside your node_modules.

• MaintainerOS - command center for open-source maintainers: repo health, stale backlog, security readiness, issue triage, PR review, Actions reports. Live at maintaineros.prince.sh.
• Quantiti - live trading journal with real Zerodha Kite order routing.
• Day job: built and still operate the full stack for Bigg Bounce, a 30,000 sq ft entertainment park in Bengaluru - ticketing site, self-service kiosks, POS, QR turnstiles, Razorpay UPI reconciliation, rain-day transfers. 10,238 transactions recorded across 3,526 paid bookings and 7,116 POS orders, peaking at 994 tickets and 937 gate check-ins, backed by 46 Postgres migrations and a five-node Linux fleet.

36 PRs merged across query engines, compilers, drivers, and SDKs:

• tokio-rs/bytes (Rust): sign_extend bounds assertion #862 and minimal-versions CI check + serde floor #863 merged.
• seanmonstar/reqwest (Rust): start read_timeout timer when polling the next body frame #3118 merged.
• agno-agi/agno (Python): preserve all tool calls across streamed HuggingFace chunks #11526 merged; cumulative streaming usage fix for OpenAILike in #11549.
• Vercel AI SDK (TypeScript): MCP tool-output serialization fix carried into vercel/ai#23935 (merged with co-author credit); Anthropic compaction iteration cache token usage fix in #24764; root-cause reproductions in #23819, #23830, #23834.
• NousResearch/hermes-agent (Python): sibling-site authority profile regression tests carried into NousResearch/hermes-agent#150992 (with co-author credit); active PRs #146938, #146940.
• apache/datafusion (Rust): correlated-subquery decorrelation #26551 merged.
• apache/arrow-rs (Rust): record-batch interleaving #11845 merged.
• lancedb (Node/Python): SDK fixes #4584, #4586 merged.
• oxc (Rust): linter AST traversal and WHATWG autofill rules, #28008 merged, #28109, #28110.
• pmd (Java): Logstash structured-argument guard in GuardLogStatement #7225 merged.
• JSqlParser (Java): 6 merged SQL grammar/AST fixes, #2838 through #2843.
• Pebble (Java): 5 template-engine fixes merged, #842, #843, #846, #847, #845 - #843 closed a bug open since 2021. Also Handlebars.java#1200, #1201 merged.
• connectrpc/connect-es (TypeScript): HTTP/2 session state handling #1849 merged.
• rakazo (TypeScript): 8 merged fixes in #2125-#2138. Also skardela#12, #13 merged.
• sqlalchemy (Python): preserve recursive CTEs through union().select() #13710 merged via Gerrit.
• redis/lettuce (Java): connection state handling #3966 merged.
• In review: Cohere v2 cross-provider history sanitization in litellm#50674, MCP session ID propagation onto OpenTelemetry v2 spans in litellm#50662, Bedrock Invoke stream abort response closure in litellm#50657, Arize Phoenix proxy parent span cost dedup in litellm#50655, and prefixed-model catalog lookup in litellm#50645; run-scoped experiment scores inside steps in inngest-js#1880; VBAR CombinedLinear FP8/NVFP4 & unprefetched encode guards in ComfyUI#17617; SpanQuery projection ordering in Arize-ai/phoenix#17746; Mistral error classification in pydantic-ai#12864; git worktree resolution in trigger.dev#5105 and run-metadata JSON parsing in trigger.dev#5052; menu bar status item targeting in vercel-labs/agent-desktop#624 and wait-option validation in vercel-labs/agent-desktop#610; MaybeReversedEdgeReference accessors in petgraph#1066; GPU-budget accounting in ray#67229; C-extension mapping conversion in msgspec#1302; explicit columnVisibility on parent columns in TanStack/table#6639.

Portfolio (w.prince.sh) · Engineering Dossier · MaintainerOS · Quantiti · LinkedIn · X · mail@prince.sh