EXACT SEARCH
NP-complete instances solved exactly, faster — by racing independent engines against each other instead of trusting one.
REHAN THEORY LAB
I work on two hard problems at once: making exact computation faster — NP-complete search, kernel-level memory management — and making it safe for AI agents to operate real machines, through bridges with an explicit consent model.
NP-complete instances solved exactly, faster — by racing independent engines against each other instead of trusting one.
Allocators and page management inside a real RISC-V kernel — lock-free, arena, and slab, integrated into an existing subsystem.
Letting an AI agent operate the user's own machine, with consent as a first-class constraint rather than an afterthought.
Races 49 engines — published baselines and original work — and reports which one won. You get the answer, plus evidence for why that technique solved that instance.
baselines: Schroeppel–Shamir · MITM · BCJ · HGJ original: GDEP · adaptive partitioning · BigUint measured: n=60 → 524 s (i3-2100) = 1,649× vs BCJ projected: n=60 → 24.3 s (Ryzen 9) = 35,556× memory: ~12 GB floor at n=60
Gives an AI agent access to the user's own browser and desktop — the sessions they already have — instead of a sterile browser that hits a login wall everywhere.
engines: Chrome profile · CDP · patched driver interface: MCP server consent: per-session opt-in, visible, scoped never: credential export, silent operation
An agent system organised as a company: identities, managers, notebooks, scoped permissions, and prompts generated at runtime from each node's own context.
structure: CEO → manager → agent → sub-agent routing: domain → model, chosen from grades execution: mouse, keyboard, screen, browser status: research prototype — graders internal
A RISC-V kernel extended with lock-free, arena, and slab allocators integrated into an existing memory subsystem. A fork of caffeinix, credited as such in the first line.
upstream: caffeinix · Troy Mitchell · GPL-3.0 added: lock_free.c · arena.c · slab.c · palloc.c scope: ~390 lines into a ~7,500-line kernel
| Artifact | Evidence |
|---|---|
| Z++ — 49 engines, MIT licence | repository |
| n=60 → 524 s on i3-2100, 1,649× vs BCJ | measured, not projected |
| n=60 → 24.3 s on Ryzen 9 | projection — not yet measured |
| caffeinix fork: 3 allocators, +391 / −164 lines | commit 329116b |
| DOI archived | 10.5281/zenodo.20399806 |
| ORCID | 0009-0003-8748-6524 |
Measure, then claim.Every number carries the machine it ran on and whether it was measured or projected. A ratio computed from two estimates is labelled an estimate.
Credit upstream.A fork says what it is forked from in the first line. Attribution is a licence obligation under GPL, and a credibility requirement under everything else.
State the limit.If it isn't tested, the README says so before the reader finds out. Limitations sections are cheaper than retractions.