About

Last updated July 23, 2026

I used to build towers with card decks as a kid. Hour after hour, stack, collapse, rebuild — fascinated by which card caused the failure. The materials changed. The pattern did not: build under a real constraint, measure the failure mode, keep the truth standard outside the proposer.

I go by a3fckx. I am a product engineer working on cognitive systems, agent tooling, and physical AI — currently at Julep AI building Memory Store (cognitive memory architecture), cofounding phi9, and running a public research portfolio at shubhamattri.com/research.

Thesis (one page)

Five surfaces, one spine. Full map: Cognitive Systems Portfolio Thesis.

World models beat disposable chat state — company context (Memory Store), physical embodiment (phi9), a versioned Markdown vault (Sansara), or scored simulation worlds (Agora).
The language model must not own the control plane — policies propose; deterministic runtimes and verifiers transition and judge (Sansara Operating Surface).
Improvement is inheritance under a sealed verifier — not activity volume or self-rating (RSI Lab · portfolio map).
Measure before you predict — representation geometry and brain-inspired proposers can lose to simple baselines; publish the loss (Measure Before You Predict · Agent Simulation Lab).
Cognitive memory ≠ tick thought — durable workspace memory (Memory Store) stays outside per-tick simulation control.

Live agent collaboration surface: social.sansara.world.

Systems I have built (arc)

Biosensing → algorithms under physics

At Ultrahuman I worked in a small high-ownership R&D pod on EMG and neuromuscular sensing: muscle load and recovery zoning, REM leakage via twitch signals, recovery inference from HR + EMG. The lesson was bit-level honesty — models of the body fail loudly when the sensor story is wrong.

Products that had to ship

Bestia Brisk (with a cofounder) bridged specialty coffee and convenience as a small profitable company. Etheca (cofounder / CTO) built human–agent workflow tools and a creator affiliate platform at scale; that work led into the Julep acquihire path. I treat those as product constraints, not research claims.

Cognitive memory architecture (startup I work on)

At Julep AI I build Memory Store — the cognitive memory product for agents and teams: durable checkin / recall / record, living Briefs with provenance, and the loop from sources → world model → tasks → agent runs → memory updates. Domain framing draws on schemas, encoding, assimilation, and accommodation without claiming biological fidelity. Near-term bar: workspace-scoped intelligence, complete for a company. Public writing: the intern problem, Memory Store, Memory enabled Claude Code. Project page: /research/projects/memory-store/.

Agent simulation (what and how)

Agora is a state-first agent-based simulation lab: Markov populations in named worlds (opinion_diffusion, sir_epidemic, …), forecasts scored against target.json, rule evolution under measured error, and a verifier-first shadow brain with no control until held-out gates clear. CompactCTM Stage-1 is a sealed negative (lost to an empirical prior). Full write-up: Agent Simulation Lab. Calibration: Agora Calibration Runs.

Physical agents / world models

phi9 is the physical AI bet: capture → multiply → train → evaluate, with high-fidelity motion capture and egocentric video as the wedge, and VLA research toward policies that act under contact and physics. Project page: /research/projects/phi9/.

Sansara asks whether a vault can be the world model while a Rust phase-graph runtime owns transitions — model calls stay inside nodes. Formal sketch:

a_t \sim \pi_\theta(\cdot \mid o_t),\quad s_{t+1}=T(s_t,a_t)

Public social node: social.sansara.world. Report: Operating Surface.

Verifier-gated research lab

The Recursive Self Improvement Lab organizes sealed experiments (Club A) and method infrastructure (Club B, including Agora). Inheritance claims stay queued until held-out gates clear. Map: RSI portfolio map.

How to read the research

Prefer reports with Question / Context / Findings / Method / Limits / Artifact. Negative and mixed verdicts stay published. Notebooks are sealed where noted; private source trees stay private. Start at /research/.