Graphify — Daxos teardown

The one question that mattered

A hundred thousand stars in seventeen weeks is among the fastest ever. We tried to prove it was bought. It isn't.

GitHub disabled stargazer enumeration platform-wide, so the star curve can't be pulled directly by anyone right now. Three independent proxies all say organic: the fork curve shows textbook spike-then-decay (no step functions), 58% of fork owners are pre-2022 accounts, and PyPI installs run 70-80k on weekdays and drop to 37k at weekends. Paid traffic doesn't take Saturdays off.

Corroborating: 185 contributors, 2,407 issues in four months with real reproductions (version numbers, OS, internal function names), and an unprompted third-party ecosystem — a .NET port, a Rootly importer, HKUST academic use, a Rust reimplementation. Star-farmed repos are hollow. This one isn't.

Residual doubt: downloads-per-star is 17, against 136 for llama-index and 2,036 for uv. Confounded by age and by this being a once-per-developer install rather than a CI dependency, but it's the one metric that doesn't corroborate a 101k audience. The honest read is a large shallow audience over a real, smaller, active user base.

Where it came from

Not Hacker News — every Graphify post there got single digits. A rival founder's own Ask HN records the actual origin: "Graphify launched off the back of a Karpathy tweet — 12k stars in 48 hours." One tweet from the right person, then compounding.

The company

Raised	$500,000 pre-seed, 28 May 2026. Y Combinator is the only investor. That's the standard YC deal, not a priced round. No Form D on file, no outside capital.
Valuation	Not disclosed. Harmonic's $2M is its own model, not a mark.
Team	YC says 2, LinkedIn 3, Harmonic 4, GitHub org 0 public members. One engineer. Safi writes 944 of ~1,100 commits; the next contributor has 27. The other staff read as operations and business development.
Founder	Safi Shamsi. MSc Data Science, Birmingham, finished Oct 2025. One engineering job before this — AI Engineer at Valent, five months. Built the repo nights and weekends while employed there. No elite org, no prior exit, largest previous repo 32 stars.
Revenue	None. No pricing page, no cloud tier. Enterprise is a waitlist marked "early access." YC lists Rootly, Geotab, Tweddle Group and Superagent as customers — unverified, though Rootly did ship an importer.

For and against

Distribution money can't buy. Sourcegraph raised $232M and doesn't have this developer mindshare. 4.1M installs since April.
Raised almost nothing. $500K for a top-0.01% distribution event is the cheapest entry we've seen this year.
Real engineering. 36 tree-sitter grammars, 200 releases in four months, near-daily shipping.
Honest benchmarking. Shared model, competitors run as adapters, blind-validated judge. Rare, and it says something about the founder.
Cross-vendor by design. Works across Claude Code, Cursor, Codex, Gemini, Copilot, Ollama. A platform owner won't build that.

Nothing to buy. No product, no price, no paying customer. Every commercial fact is unproven.
One person. Solo founder, five months of professional engineering before founding, 86% of commits. No co-founder.
Apache-2.0, no CLA. Cursor ($3.4B, ships its own indexing) or Anthropic can fork it. 185 contributors make a later relicense politically hard.
The precedents are graves. CodeSee, code graphs, exited at $0. Bloop, LLM code search, dead. Continue was absorbed by Cursor.
Category drift. The OSS product competes with Greptile ($29.6M). The enterprise pitch drifts into Glean's market ($765M raised, $7.25B).
Wins on cost, not quality. Their own numbers tie dense RAG and lose to supermemory on QA accuracy. The moat is $0 build credits, and cost arguments erode.

Claims that don't hold

The 100k announcement carried two boasts. "5th YC-backed OSS project to hit 100k" — it's the 7th, behind build-your-own-x, opencode, ollama, firecrawl, browser-use and supabase. "First Indian developer to ship a 100k-star repo" — false: TheAlgorithms/Python (223k, Anup Kumar Panwar), TensorFlow (Rajat Monga), PyTorch (Soumith Chintala) all predate it by a decade. The founder is also London-based, Indian by origin.

Separately, the README's headline "71.5x fewer tokens" comes from a hand-picked 52-file corpus; their own honest number is 5.4x. Smaller housekeeping: the Discord invite in the README is expired, pyproject declares MIT while the repo declares Apache-2.0, and after 200 releases it has never cut 1.0.

Verdict

5.75

The traction is real and we could not break it — that's the finding, and it's worth something. But Daxos would be underwriting a solo founder with five months of professional experience, no revenue, no moat under a permissive licence, in a category where the two closest analogues went to zero, against a platform owner one quarter of engineering away from copying it.

The practical problem is price. At $500K post-YC this would be interesting. After 100k stars and an S26 demo day, the next round clears well above where a $100–500K check buys meaningful ownership. Track it, take the meeting, don't chase the round. The thing to watch is whether a paid SKU ships before the star curve flattens.