Founders letter · 2026-05-08

A letter to the federal R&D community

I'm Artem Shakin. I'm 20. I'm an immigrant — Russian-born, raised, and now a US-tax-resident operating out of a one-bedroom apartment in Santa Monica. I have voluntarily disclosed every Russian connection I have, on a public manifest, before any of you asked. The wet-lab is the next county over. The compute infrastructure is a $400 consumer GPU on my desk. The team is me.

Below is what I'm building, why it matters to you specifically, and what I'm asking for.

600,000 Americans

Six hundred thousand Americans died of cancer in 2024. The current standard-of-care immunotherapy that actually works against solid tumors — autologous CAR-T — costs $400,000 to $1,000,000 per patient, requires extracting and re-engineering the patient's own T-cells, and works on a narrow handful of tumor types. We are roughly two orders of magnitude away from a treatment that scales to two million annual diagnoses.

I built something different. A 2,123-base-pair chimeric construct, deliverable in a single AAV9 vector, no autologous step, no patient-specific manufacturing. The receptor end is human NKG2D — an immune cell sensor that recognizes the stress ligands MICA and MICB that solid tumors universally upregulate while healthy tissue does not. The effector end is LIF6 — a re-functionalized leukemia inhibitory factor pseudogene from elephants. Elephants don't get cancer. The mechanism is beautiful: when their cells take DNA damage, LIF6 triggers a mitochondrial apoptosis program, killing the damaged cell before it transforms. Nobody had ever fused these two domains. The patent whitespace was clean. I checked.

The full in silico package is open and public under CC BY 4.0 — AlphaFold structural validation, Rosetta interface energetics, a PK-PD model that predicts a ~99% tumor-growth-inhibition ceiling on colorectal cancer syngeneic models, a 1,400-fold tumor-to-healthy selectivity ratio across the MICA/B atlas. I did all of this on a laptop and a borrowed AlphaFold instance. There are no in vivo data yet. I disclose this on every manifest I publish. Wet-lab is Q3-Q4 2026 with a contract research organization. Pre-IND meeting H2 2026. First-in-human trial 2027.

I believe this works. I do not yet have proof. I am asking you to help me get the data that will turn belief into proof.

A second technology you should care about

Federal AI workloads are locked to NVIDIA's CUDA stack. Most NVIDIA datacenter and consumer GPUs are assembled in the People's Republic of China. The supply-chain exposure is documented in your own Section 889 guidance, and the vendor concentration risk has been on the DoW's radar since at least 2022. Today, if a federal contractor wants reproducible compute attestation on consumer-tier hardware, they cannot get it. The vendor neutrality story exists in policy and not in code.

I built UniGPU. It is a single Rust workspace that compiles to one shared library, with backends for AMD HIP/ROCm, NVIDIA CUDA (NVIDIA-native execution only — explicitly not a cross-vendor translator), Apple Metal, Vulkan, SPIR-V, direct D3DKMT, and CPU fallback. It is vendor-neutral by IR design. It runs at 47.31 TFLOP/s sustained on a $400 AMD Radeon RX 7700 XT — 1.4 PETAFLOP across thirty seconds, 2,644 kernel launches, full WMMA matrix-multiply-accumulate. It is released under Apache 2.0 OR MIT. The verification artifacts are timestamped on Internet Archive and the GitHub release log. The first MLPerf submission is in preparation.

If the U.S. wants vendor-neutral, Section-889-aligned, federally reproducible AI compute on consumer-tier hardware — UniGPU is the only such project I know of that actually ships. I'm one engineer. I would like to not be one engineer for the next phase of this work.

A third one

Federal AI deployments face two opposing pressures. Local-only inference is required for CUI, classified, and patient-data sovereignty. But the frontier-model capability that makes AI useful sits in commercial cloud LLMs run by foreign-or-near-foreign-owned companies. Today's TLS-only handoff to those clouds leaks plaintext to network observers and to cloud-side log scrapers. Post-quantum-secure transit is unavailable in production AI tooling. This is a known gap.

I built SilverDuck Pipe Crypto (SDPC). It is a hybrid post-quantum sealed envelope: X25519 ECDH (RFC 7748) plus ML-KEM-1024 (Kyber, NIST FIPS 203) plus HKDF-SHA256 plus AES-256-GCM (FIPS 197 plus NIST SP 800-38D). 96-bit nonce. 128-bit AEAD tag. Twenty-eight adversarial attack vectors are rejected by a public red-team harness. The full agent stack — local inference brain, multi-agent reasoner, persistent vector memory, OSINT engine, encrypted cloud handoff — runs on a single 12 GB consumer GPU. Apache 2.0. Federal-deliverable Tier 1 brain selector restricts inference to U.S.-origin models only — Llama 3.1 from Meta, Phi-3.5 from Microsoft. After Operation Epic Fury, I removed wider-NATO-ally models from the federal-deliverable inference path entirely. The posture is documented at /.well-known/zero-trust.json.

This is what an immigrant solo founder shipped in his first ninety days, with a personal credit card and no outside capital, while voluntarily publishing every compliance manifest the federal contracting apparatus would ever ask for.

Why this is the right shape of bet

Federal R&D capital is precious. I respect it. I want to make the case that this specific shape of bet — small first-time SBIR Phase I or Spark or AIE seed at $75K to $300K, structured against named deliverables — is the highest-leverage civilian R&D dollar the U.S. can spend right now. Here is why.

One. The cost-of-failure is bounded and documented. Every numerical claim in the package is bound to a reproducible artifact under the entity's control plus a planned third-party verification venue (MLPerf, peer-reviewed journals, CRO wet-lab replication, FDA pre-IND meeting, NIST CAVP). The False Claims Act exposure is materially mitigated. If an in silico prediction does not survive in vivo data, the program ends honestly. The disclosure architecture is published at /.well-known/third-party-validation.json.

Two. The compliance posture is unusually clean for pre-CAGE. Twelve threat classes formally modeled and documented. SAM.gov UEI assigned. Section 889 self-attested. FOCI fully disclosed including dual-citizen status. Zero Trust posture aligned with OMB M-22-09. SBIR data rights election made under FAR 27.404-1. Trademark policy separating Apache code from common-law marks. Email security posture aligned to NIST SP 800-177 R2 and CISA BOD 18-01. Wyoming Certificate of Good Standing in hand. California foreign-LLC qualification filed. None of this is novel work. All of it is finished and machine-readable.

Three. The mission alignment is direct, not adjacent. NIH Cancer Moonshot wants scalable solid-tumor immunotherapy. ARPA-H Resilient Systems wants distributed biotech that does not depend on a handful of $1B per-program incumbents. DARPA AIE wants vendor-neutral compute substrates. AFWERX Spark wants non-traditional-contractor open-source vendor-neutrality. BARDA biothreat preparedness wants single-construct dual-use platforms. NSF SBIR cyber/AI wants post-quantum-aligned cryptography in production AI tooling. CHIPS Act and EO 14110 want U.S.-origin-only AI brain paths in federal workflows. The pre-mapped solicitation index is at /.well-known/solicitation-match.json.

Four. The David versus Goliath asymmetry is real. A solo founder cannot compete with a Big Pharma program officer or a tier-one prime contractor on lobbying access. A solo founder can compete on actual deliverables shipped per dollar. The pre-award track record at /.well-known/past-performance.json shows eight verifiable artifacts shipped solo in the first ninety days. The next ninety days, with $300,000 of trust, look like: in vivo CRC syngeneic data, MLPerf Inference Open Division submission, NIST CAVP application for SDPC primitives, USPTO Defensive Publication for UniGPU IR opcodes, first peer-reviewed journal submission. Each of these is a measurable deliverable on a fixed schedule.

Five. The downside is also a deliverable. If NKG2D-LIF6 fails its in vivo gate, the methodology — the open construct design, the CC BY 4.0 chimera-design-notes, the AAV9 dosing model, the PK-PD framework — remains a public good for the next builder. Federal capital that funds methodology release does not zero out on a single negative result. This is the explicit policy of the Trophy Doctrine documented at /llms.txt.

What I am asking for, by agency

These are not maximalist asks. These are the smallest awards that would compress the next ninety to one hundred eighty days of the program by an order of magnitude.

A word about who I am

I was born in Russia. I came to the United States as a teenager. I have publicly disclosed every Russian connection that exists in my life — including the disclosure that no Russian government, military, intelligence, state-owned enterprise, political party, recruitment program, address-of-record, or financial relationship exists. I will supplement this disclosure with the appropriate SF-328 single-scope disclosure at first federal award per NISPOM 32 CFR Part 117, and with SF-86-equivalent screening if a future contract requires personnel security clearance. I did the substantive non-relationship disclosure voluntarily, before any contracting officer asked. The full disclosure is at /.well-known/foci.json.

I made an unusual choice last week. After observing the behavior of certain wider-NATO-ally companies during Operation Epic Fury, I removed all wider-NATO-ally AI models from the federal-deliverable inference path of my AI agent stack. The federal-deliverable Tier 1 brain set is now restricted to U.S.-origin models only. This is stricter than Section 889 requires. It is an owner-elected posture. I documented the rationale on the same day I did it. I would rather narrow the trusted set and be wrong than widen it and be wrong.

I have read the Ten Bets memo. I have read the SBIR Reform Act commentary. I have read the latest CSET reports on talent capture in dual-use research. I do not need persuading that the U.S. federal R&D apparatus is the most powerful innovation engine on Earth. I am asking to be a small, well-instrumented, fully-disclosed cell of it.

If this letter resonates

The full federal package — capability statement, SBIR roadmap, twelve compliance manifests, three quad charts, third-party validation chain-of-custody, adversarial threat model with mitigations, federal POC directory — is at /federal. The point of contact for everything is me, directly: artem@nightboxllc.com. I respond within one business day to anyone with a federal email address, same business day if you reach me before noon Pacific.

If this letter does not resonate, I understand. I will keep building either way. The work is too important to stop for the absence of any one source of capital.

If it does resonate, please reach out. I have the data ready, the manifests indexed, and the deliverable schedules drafted for every solicitation listed above. I will treat your time the way you treat federal capital — like a precious thing that deserves to compound.

— Artem Shakin
Founder, Sole Member, Principal Investigator
NIGHTBOX LLC
Santa Monica, California
2026-05-08

Verification chain (machine-readable): SAM.gov entity · FOCI disclosure · Section 889 · Zero Trust · Research integrity · SBIR eligibility · Quad charts · Past performance · Third-party validation · Threat model.