Policies & licensing
Work with us
Interpretability tooling, SAE databases, audits, and research for teams of all sizes.
Reverse engineering intelligence with interpretability.
We reverse engineer how models are trained to bring interpretability into building better ML for frontier models, life sciences, industrial systems, scientific discovery, clinical research, and safety-critical deployments.
Backed by
awesome programs.
Building an interpretable VLM that maps text and vision towers on top of OpenCLIP.
Providing interpretability tooling for preference detection analysis in LLMs.
Using SAE tooling on a small LLM to sense conflict, ambiguity, and tone, and choose the right response.
Working with labs and research partners.
Training a ViT on wound recovery imaging, then interpreting it to distill a prediction system.
Find the features that drive recovery signals, then build and distill a leaner predictor you can actually inspect.
Mapping correlations between the human brain and LLM neural nets.
Building a testbed for drugs and cures aimed at Alzheimer's, dementia, and schizophrenia.
Complete interpretability tooling for frontier ML development.
Bring interpretability into how models are built and improved, not only into exploratory post-hoc analysis.
Deceptive feature finding for AI safety.
Working with individual researchers to locate and study deceptive features before they become failure modes.
Interpretability for ECG-related ML.
Opening cardiac models so predictions can be inspected, trusted, and improved in clinical settings.
Finding emotions in LLMs, and whether they refuse or agree internally.
Using SAE tooling to probe internal affect and decision posture, not just the surface answer.
Training SAEs and providing tools to frontier labs.
Dictionary learning and tooling so labs can inspect, steer, and develop models with mechanistic grip.
Putting interpretability into the development loop itself.
Reverse engineer intelligence to build with intention: inspect, attribute cause, intervene, and verify before you ship.
We solve hard problems.
1f/ \1 || 11 )/( j) |f)t(\(t
//xv (Ut jQ jf/\|1
\ 11/1 / ( \ 1 (
(|\1 ff j c j nvn\ /
1 1f un 11 x1j/ 1f\
| (tt v n x1 fx j 1) j1
\n)| (fr z fc C( n t )
/) (L 1j XJ zxY u /1 t/ 1/ j )r(
f z ft ( )U v zr xL uY) /| \x x f
/( (( \u jj nf vx C v (n 1 x
fx C |1 ((t C U v/t v0 r1 ( jfv)(
c v( r U jLr z vZ jn /j 1( j
) 1 \) v Y vO u\ z 1 j
/ X1 f c rC( Uu( c) u| t j x v
t( 1 z zu zc vU XC v/ XUxf/1 jU 1z u (
jx x 1 Y (Uf rO U vU xJ t r / f
rx nC \ vu /f )( JZ vf (U t r / j1 r
rj ) (fv nQ \cL /c nO v/ f x (( (( x
fU ( v ( f0 C /X fCY U / t x v1 1 j
) vt u )\x x0 rn vz Q cXu( \ J c \() f
x / )1 ut c0 1/ Q /v U1 x1 x( JU( 1 / )
c j1 v f0 fn tj X( z (L L fL\ | / ) 1
r x ux c U n C Y c z| / | rt f v
1U | x j xL X fz x \v v J \ f t zU n /n
U (z f ( x0 c| v/ vr z1 /X 1z f v( c / u
v 1( |x v \j \r zYzzcnxcQQ/1 fO rv X 1 /v z v 1 u
n (U // t 1c j/ |) vcYzzQOz)( ) uu t| \ u ) X t x
x ( 1 z0 zU (z \ fXt/ xU vc z (uY n )x t 1 u
Y f rx r J uX (UL00x v 1U )z /x( (C1/ rv / x
0 Y t t 1u 1n( nnt1 f0 f U U j1 U c u z
z/\( / f1 vX vf1 z |XU00Utx/ un (z /( t 1((C x
r n v 1v x())/vfjvCz( 1rJXvj\|(zc L /t u )x U
/u / \z u( )vnf/|\\\/nv) / v (zL0C /Q /x)1fvnrf ( Y r U
U 1Y v z) 1z( 1|uxtjY ) fJCvxz0 \c( ( | v( f x
U u )(x1 xLXf) )CZJ\ 1nf uC || j c n
|X \ / cn1 f1 x\ rx1 1LCCY ( )( xr /f /X u
v(z |z | /1 xn fxjfff( 1 )vXzf1( f0Cv( rv|1jf|\/| z f/ f| v
U j ( x\ f 1vz( vn()u/ CLLxfjfzf f/ c1 xjU
U f \ v 1()1jn/f/ttfr(1 (| xfnXv U0Yx/ \rr ) ( )|v\ 1Qt
z x / 1t fvrf(1 ( 1Xx| 1vf 1 10QCx1 ( vUxjx\r 1 (znvxzux( /Q
x(f u tX j| 1v//fxf 1vU CCnxx1xjfJCCcCr///(1tYCxt ( (Q
u)fc x| fz )) \ff( x f/1 vr YCQUUzxvzc|( (/Uxrjxzv1 ( xf
u X f)n U| cU1 |uv\ |fx( 1 z0LXcvUzj/f/fxXULj v(
cL v c j1 )f( 1 ) fvx/11 1)|jvX/ /x/( 1 \J JOQnj/ /xxcz/f() 1 1 \\f
Ct /r fr (\/))|//tnzUCLufj ) 1x 1z/(vU1 (u Czun/()/ffxj/( /ttft1(1 (x/1)
U| (tfr\ 1()///cXJYucuUCxrJxjYvrULvxJUcf/r( 1/|( \ vj u/ f/ 1(juvzn/)1
U( ( 1cffjtfzCLczXYf xf( 1 1 uz vvnQCLLv\ 1zj1 v C
f|/( \///nC0CvuYv( 1 z/ CUv0UXUCvj/|(|f/( \ u(|vv z
1tj/rxvUC0Cuf/ (\ x )v|zCz0QzCQzCLLLLr(1 (n ) (XU /|X
zZz/f ( / t x j|CYr0U00zcCnzvXXUf )(fr f U|
1Qr/) 1 )) ( x x1 z /n rz(vX 1 1/xCJYUf (\/ 1 x
f0x)\( ( |1 /x x|vvXxXxxttc|tYf ((rOCC( |Yt )| ///j
)Cfn/ (UfUCU0x (xxJU CxvU r| )f/1 (v(
vxff/1 1/ / r /vzzCUCLU vr(C 11 11(|/)11
Xr|f1/| \ v ( UfCOU Y (Y
z/tc t 1 v / t r(c0X| x\ u(
tj(\uf f| 1 )n f ( z1UZLc( ( 1z(
zf\tcv \(1 f( 1 tt f f XxxOz 1/cQYUQ(
zff|nzrnf)| )\\1v|)/ tY )1 ff/xfUZ\ C nZ
1Xvrj(nvcjrx/tt/(/fft\/ 1/1|j \/)YZ/ \ Uz
xYXx|t(fv| \1\ (j1 \n(fu fv|fJ0Z 1 nU
11/Xnrxj) n( (xxxnfxnnfnJ0JY/ vx
))fnv\(||(/zUvrUCLCY / U1
cZ// 1(x
uZ Y1
j0 (C
rL z/
nv rX
z )L
U 1 L
fc v|
v/ /C
U U
U1 xc
f ?||\xCYfzr||||ccfrr/tnXvjLnjcCnUuzuUj()-
~][}{(YYOqj i1tjuaU L&l I_tLCXzcrf1{]
_+zt(}-[}[t{|rxXr11}+Y[ tvj? |z]: ;!~~-]tx" [I <\\v/{
l[(uczr ]CJl ;/L[ {p> ]L; )ZwZz\|l~(Y)-:
]jff> !, +]~! >_x1> )xC), _mZ; +/- ;" >l ~}rx :}jOrLXjxJz)?
+cj_~[+>]{~uUUfl [w? |Z\ \OrI ,JO (L i, I:1rn+ i)/Lr{?
?_zZnc(l _/( <l +uJQf,: ]b- l>Cp?"?Mv?:I tZ_ +/: 1U] {n(ffi
]Xnf+uW_ ~ }\ -xL; l+k* jdOh ~mj :+-][1Xr\\fUX(: rY1<!>Lj?nQv(I
;[J\ }b- ~~CUjuj(no \w, fdQ! 0W" ?mhn/1)j~ }Yc1 Iz0 ?\0j,(J)
?Yu|||:>vu <{f| _zm|l ~! LL (ZC w0 l +Mi ?w) vZj}_ ;|r:_O1
+JO|-; f#l ;cr }vuU\ -#: ja : " ;qYU< _w%: :[ Qx+ ?)v) tLJqnv(+
}pt ,?jqu IQ~ : fa: LWQ(> >d> , , w@ LZ iYL 1/()|_I ?ru! [f\ (L{
I/n) +, <x( zn l? q? ; ,l: ih{ +>qB+ i-: mz ld| +f? i \r 1f /L
1Y, +| /| i+\kx; :CI p1 ; :?QMu ]omv{ pn ~mc; /U- [L< <0-~w1
;Uu <| x| ~(()1(fbI ;dm! >bp, qo<, ,kM ?wY\I !ka0JXx)[! +t! Laj! xd: )m_/ju-
lLO:l < /O _x] )~ }h[ <vaC( !+" QB: (oi qo - 1/O*l >fU; fQ|, X- |J,
+QO:[; IZ#Iz) wm i/YJ/\1l,nu >+\*$<,, qC- uo )Uj _Q[ !X1 Lr IQ]
!QL _| >(nfuCq) ;[ Z& ?></zpW ;Xp[ ZBli: qY ?J] Ii ]O vu {fvul 0[
Jo v[j\ q> 1v! !: C8+" (#; Xk<!]; ;C#hl /a- , YjU( 0~ Iqx :r/IL\
|nf qU m{ ;nzO: 0% iI~i, QZ : wbf /&) +pkmv\{~, {z> f* p} {zvn} ]tf!J}
~O{ X\ nx! mU p8 ,f|i xX< <~ j&; , XqI -or ; 0r 0v j*ou{< _r1 !Y>n\
<p ; >Lx 1, Ii; q1 I[J8? JZ ?UL }kul ~b< uW }W+ ++CWz, + >)c/+ {X n!x\
Q }! +0i j] ,> mO fp1 ?o_ -Z 1#- [h :p( _dcf)?}jt1 _\X~ }c+ QU J!wI
xa c >O !L ; : OW tk| lJd? O> zq ; fq I \* )z ]+ C? : U[ ;bo Z;YZ
-o+ _ub1 jC I-uf; I0%t )pc /qY)~-~!, ,qQ;, ;dx ,: ; jf ,- Yk xd i Jt_ ,?I 1qv_ pf -X |m_
-wj xv" cO [w{ ~\\[[zL{<< zQ; jdhbdpZOp&&Yr[+~lJB- >l } /~ ~] Qw +kf rI Xwj/fdj ; ~qf~ ix -m>
_Z/ j[ ! va YX-1Ut: Irp<, LX| cn Ifqphdb&Bbnu->; n\ ) {l mm Uc i z1 Im tnI ~k]\U{ li jO
0i u? x> ~+d%~ {do udl : <{ z ]jCkUXt0a: t! :qp ldI } i)umh ]Z[ nO >]! U+ /r |I mi
]h tJ Q0|- -Q{ [*< mkt, l ; , -, ;uaW880f_>,I +q< ro nb YOv? ~ ?i u#rY\ !-|Qq Y/ f+ 0|
(8:_h> lU( )U(+ xm! <rZu+, ;)+ Ii I ][ ;;tjZZUx-C%)l )CI -a] faf \Lx~, , foI <p_m, d+
<dXzvI X! I Jx wk; jwJx+~+<~<d; i : i ;cko8%aUOX]<>> "+mZ ud , -Yu( fU ruv#! )O
QlIZ {w l( :t i)xqt ::: !(OvnnYwJLq#du)>+,I,l ): :? , (i ~xQ*kwLzcubp: ,fW l?;>1! :[YU I m < n0 o
Xm 1X zb (/+_ >: mv ,, \nwZCXczzzXZwn]~:l:+Y , } ~w{udW8M{_~}X&[ l{YOnxCqZQCfju_ > ;{: >h_ Q? o
a[ t; ~rhl +fq! >bn >rdv_ I|rcm0ULh-; n >J*#q0d8_I;<?zpu )v\c|! +wu?I ?1 C>[O
o! / <} : +j 1m? lI<_)[(nuOxl [ ,/OWkCnt{>: ;: f>nM@*z/xZC{ ! ;{m#I ?cc :L :p>Z;
ck\z ?X <" jpZr({(Cx\( \0z)i : ! ,(tQ0x?]>,; ; ~ / rWM#h)> :!<)|u[ ,?nv| OQ I|ttl ,XJ Yk m
wvd cb >cj?-I ]Yr ]OZ l/J0LJCJv\{-1xi , ;nqkdJru}II ,1i l !:J%Mwv1><>-jQqcrLCczXc/< (d: ,~JY Cc|w
o<L ?uj :;I-(t0z J; : :1fxqbv_>I: + l} (jwZunmY, I , )MWW0CLCbC- -CY I-I\pr 0La
oiC }z! iw ~ i[ <{\runrLZYJYUUCQurtvcI [OCZkqj1i: [, <o%h0Y) ;I>])/fzQQ_!+{]]1n\v_tncqz: r&U
b?0_ <Y> xU l -CwQJvx)}1ufi , {? [rkOc[<!:l l[; -, )xwC(l-x- l r8&#Oxf\{1(vftwaOL0zQ]r/-jvbZwObmOv1 +-l:?X&i
IOuC! (m1 Uk (] ;]Lc< (t; - ", ,|xwYYC0Cf+<I :(xqal >_MMZOOxOLC*##p#QYXXurUh#OU[-il;il !;l)vfu&I
mnJp- ~0c|1|tJd| nn|(jzCJu l\- ]0! i(fCYr\],l<I> + IwQ} !, > \hM&aob0wbpcu}_??uYaOQL0bwr/1)11tvjtl?OJ
mjkl(JnZ! ~ac[I l !]?!<~~!ll~fpax|{]?+!; ;_; :(fcmqz_ ,! cC0u/(x] ? b8WkpqobLYCYJOkoWL)_:, :::>>l l!>fqu<
<pWi w- :, ]pt |Lr\)nJv)tjjtfrjnf[CwOXxrjrncLwkX{}i>iII: _XOXu/~)(|_ r >z*< _ f*B&ZLY[YOOpdYJvn[\x\\/(x11}}|jzzCj
/#U: t[ ,)l <_?/YQ~ !_tCQ->?uzXnncXXUZdoMWmCL{nj__<_~: ;1ff/ttr0j]i,ll?//}<" rbYuqoxl i um]-lI" < ~MbmZYunXJCOLXujjYUUJUxurjv0Yxn/
/ac~ l ~uUJQz)>!"l{/xvnXXYpk*hmpmoM0Q*0LhqQoWqO*opCXQv\//{], -\fxXcu\_ z +jwL); :, mY?>l 1 CX>/f\}: >txvLmqdZXnri
)av]{{v~I I;xpCJLUJbMWpbkhJ1OCv1!!r|?~\t<r~<md{wqZ&MWWwz-~\f\|t1]l!I : , rbLx1-[{! ]w I:+?>#
fCcYv\)tzXXXZM%Mqmhwu[{1)|){::[1t\i-x(I[:]1"ft>bY?#aw8ako#qLYcucCXu\/(\ztj1\(_>I +mvcqwf\)\-: i\> ttId1
|rULXQOqaM8MmCYt}jvzj}??++_~1f1]{}? -1t _f{i> 0f[nqcb#d8&d#&bMWWWWQvx[~~>_vZt?i ;i~ n< >vko1 + ,i,Yck{
b@bXJ|{]{{}I! ,, I ]}\{[|\:l}u- Y_>U[O1fLcMhQ8o88bp#ZdwkkaC{!/nuCQ};: < }C\[ : ?+~oc
r&QYn?{{r{>>]+[})[-<;}|{[-](]}}<>{nn?,u)l0+0x?b-YZ}Qbvqk?rt++_xXOM*hoCj><fvzX)_, i?}x~tO
J8Onzvi~]}+11! !_}u(? i++1_, <,I)cxjX0[0cqwk0k0OUUpcUhJil+ ; <vuQB#MvtchU->nc|(?l<~+?i~(+I>[?1/XYXL/
nMJZXf)-_i<-)\1i]l ;])[ <_[[}1)?,,<vaJaMo%O[[[[v00*afl , :}l(#0qo ,\Qct\+_)\t)jnJXr|j(fvqvj
wOJCXr/_\|) -)~ >) _t[! I??}xXf!iXjQ~Xwdb#a#Wa}_l, I!}qQvM !(tfjrrjrxucXnrr(
:kQcCrXc<?~jj< -){ +j!l[}+1/1!_}l ?z( w -fu1oCMBa{{!, ,;ljh]uh
~bYUp1[)1Ui!)xt<\\{<]] ?< ~[{I_|f) -w;+Y,U_Qvp8kc(_~,, ,:~<|Oz[mu
ULvzmC?_Jc)l)r})t!!t/:_{|}{l}<l(nZ> Cl/u/dxo@Wpv(__{(f|jutxbv
(bJzUpw]tzvx(](]Jv<{}]|i1x11||UU,1J>tJ[kO0Bb)+::(xXp&ha&v,
tbJJcZbQZCncj]nzzrwcnY\?||Uh<lnxiJJXOJo@z!i i +#]Z$/
(xkwQLuZwpLQOYUUYuYCCUzY?~rXrcL]zYnh@Y:: i ,z>/ad
fOhk0cUuJwc}zrz}vLr]zZvJm[JwcJ*8$I ! x!|Za
{rxXkZQ0LnjZvtu0OOZJ0ZZJZ*8*hX (<)q0
\nnJZqzvccuXdowQaMW#h_Y : \]jor
p$XY>}] fxu0
m$_\>( ; }jhr
L%>?), ~u#f
QW!t} ; )bY
Zq>j : ~Qk
,b+}{ I!nW
to_r : /Wi
Cp-? : +qc
wX> l;XM
[a~< I ]o_
}or; ~Op
-//}]_\CjNeurology for machine learning.
Mapping correlations between the human brain and LLM neural nets to build testbeds for drugs and cures in Alzheimer's, dementia, and schizophrenia, while covering the wider med-bio stack from wound imaging to ECG.
Learn moreIntroducing Aquin tooling
build models with same precision as writing code.
locate, debug, simulate and improve.
build transformers & llms
layered attention
simulate lora
low-rank weight adaptation
The science: Mechanistic interpretability
Mechanistic interpretability reverse-engineers how networks compute, not just what they output. Aquin uses sparse autoencoders, logit lens, activation patching, and causal tracing to find which features, layers, and circuits drive each token so you can patch failures at the source.
geography / capitals → Paris
Work with us
Interpretability tooling, custom SAE databases, mechanistic audits, circuit reports, and hands-on research, experiments, and studies for teams of all sizes. Reach us at aquin@aquin.app
Not sure if Aquin is right for you?
Policies & licensing
Aquin