Investor diligence surface · 2026-05-28

Diligence on Sipsa Labs, in 5 minutes.

Public artifacts are linked below — every PPL ratio, every architecture, every honest negative we have catalogued. The NDA-gated set (full USPTO claim text, revenue funnel internals, customer pipeline state, AFWERX Phase I submission text) is available on request. This page exists so you can do 80% of diligence without a call.

PPL-verified architectures (registry)

1.0066×

Hermes-3-405B on 32 GB GPU

30:23

honest negatives : verified wins

paid acquisition spend

1. Public artifacts (start here)

Public auditor registry

github.com/sipsalabs/ultracompress/blob/main/docs/benchmarks.json

Every PPL ratio, every architecture, every reproducibility metadata field. 22 verified records spanning 0.6B–405B parameters, dense + MoE + 1 SSM (with comparator-note caveat). Each row reproduces via pip install ultracompress && uc verify <pack>.

→ benchmarks.json

Honest negative-results ledger

github.com/sipsalabs/ultracompress/blob/main/docs/concepts/catastrophic-failures.md

30+ catalogued refuted approaches grouped by family. Why we publish them: the 30:23 ratio IS the moat. Any team claiming their 5-bit codec beats AWQ on every architecture either is lying or has not tested enough architectures.

→ failure ledger

Customer-side audit primitive

pip install ultracompress · uc audit <pack>

Emits a structured JSON receipt (unsigned by default; Ed25519-signed when you supply a key) that compliance teams attach directly to FDA SaMD pre-submissions, SR 11-7 model risk validation packages, DoD ATO accreditation packets. Schema: audit-receipt-schema.md. The regulated-buyer credibility tile nobody else in the 5-bit band has.

Public verification + cost surfaces

sipsalabs.com/{leaderboard, tools/verify, calculator, badge, request}

Side-by-side leaderboard (UC vs AWQ vs GPTQ vs EXL3), paste-any-HF-pack-URL verifier, GPU-savings calculator, embed-badge generator, request-a-pack demand-signal form. Full buyer funnel: discover → verify → prove → justify → request.

→ leaderboard → verify-any-pack → savings calculator

Hermes-3-Llama-3.1-405B pack on HuggingFace

huggingface.co/SipsaLabs/hermes-3-llama-3.1-405b-uc-v3-bpw5

Largest dense near-lossless 5-bit transformer published. 405B parameters reconstructing reproducibly on a single 32 GB consumer GPU at 1.0066× PPL ratio against bf16. Per-tensor SHA-256 manifest. Gated:manual access by Sipsa Labs Anchor Design Partner policy.

→ HF pack (gated)

Cross-architecture portability

docs/hardware/apple_silicon.md · docs/hardware/amd_rocm.md

Working converter from UC v3 pack to standard HuggingFace safetensors, verified on Qwen3-0.6B with 0.9966 forward logit cosine into mlx-lm (Apple Silicon). AMD ROCm path documented (zero code changes via PyTorch's CUDA-to-HIP shim). One converter, three hardware targets.

2. Traction snapshot (live, organic)

Distribution (organic, $0 paid)

last 30 days · pypistats canonical + HF API + GH traffic

PyPI: 3,537+ installs (mirrors excluded). HuggingFace: 1,769+ pack downloads across 41 SipsaLabs repos. GitHub: 1,731 clones from 469 unique cloners (last 14 days). Hacker News is the #1 referrer. ~5,300+ touch events in 30 days, zero paid acquisition.

Customer pipeline (outbound)

Resend log + cold-pitch state machine

Round 3 (7 named): Palantir CTO Shyam Sankar, HF transformers eng Marc Sun, C3.ai CTO Nikhil Krishnan, DataRobot VP Lisa Aguilar, 3 press (The Information, Platformer, Ahead of AI). Round 4 (14 named, healthcare + frontier-labs + audit-compliance) staged for 2026-06-02. Round 5 (10 academic-first per data-driven playbook) staged for 2026-06-09. Reply window 24-72h.

Grants + government (non-dilutive)

Federal non-dilutive grant trajectory in progress

AFWERX SBIR Phase I submitted; currently in review. NSF SBIR Pitch in. NIH PA-27-100 SBIR Phase I draft refreshed (deadline 2026-09-05). Standards-body engagement plan staged across FDA / NIST AI RMF / ISO/IEC 42001 / SR 11-7 / DoD ATO / EU AI Act.

3. The moat (4 layers, all defensible)

1. Cryptographic guarantee. SHA-256 reproducible reconstruction is qualitatively different from "approximately equal." Every other 5-bit codec is the latter. The regulated-AI buyer (FDA SaMD, SR 11-7, DoD ATO, EU AI Act) needs the former.

2. Patents pending. Two USPTO provisional applications filed April 2026; non-provisional window April 2027. Additional provisional filings in preparation for deployment-architecture extensions. Full claim text under NDA.

3. Federal non-dilutive grant trajectory in progress. Defense, scientific, and health-research grant routes all engaged. Details under NDA.

4. 30:23 honest-negatives-to-verified-wins ratio published publicly. Auditors and serious engineering buyers respect this; competitors with "everything works" pitches do not have it. It is also a powerful filter on hiring — the candidates who self-select against it are not the candidates we want.

NDA-gated set (request via the form below)

Full USPTO provisional application claim text (both filings)
AFWERX SBIR Phase I submission text (public-redacted competitive sections)
Revenue funnel state (Stripe + Resend + sipsa-inference DB internals)
Customer pipeline state machine (current targets, hooks, signal-watch column)
Cofounder hire pipeline (15 named candidates, current outreach state)
Engineering-agent session reports (technical depth for technical DD)
Internal compression-method documentation (codec architecture under NDA)
30-month financial model + use-of-funds projection
Cap table + 409A status

Request the NDA set →

4. The ask

Round

Pre-seed

Target

$750K – $1M

Vehicle

post-money SAFE

Cap (target range)

$4M – $6M

Runway

18 months

Lead profile

defense / regulated-AI specialist with 5 warm-intro slots

Founder

Missipssa Ounnar (solo + cofounder hire pipelined)

YC status

S26 — in review (decision ~6/05)

Use of funds (60 / 15 / 15 / 10 split):

$450K — Hire 1 cofounder + 1 senior production engineer. Closes the vLLM kernel integration gap that is the binding constraint on enterprise rollout.
$120K — Compliance + sales. SOC 2 Type I attestation, first enterprise sales-engineering hire, MSA template legal.
$60K — GPU + cloud infra. Multi-arch testing matrix, customer-side audit infrastructure.
$170K — Cash buffer. 6-month runway extension cushion if PMF takes longer than 12 months.

Conservative milestone: two paying enterprise customers landed by month 6 (under realistic Phase 0 → Phase 1 conversion at $5K / 5-business-day cadence).

5. How to engage

If you have read this page and want to move forward:

Email: founder@sipsalabs.com — same-day reply during business hours (US Mountain).

Include: fund / angel name, stage focus, thesis match, expected check size, next-available slot for a 30-min call.

If you want the technical version (deep-dive on the codec method, the failure ledger, the architectural-limit research arc): request the NDA-gated set above. We turn around the executed NDA the same day and the data-room within 24 hours.