diff --git a/README.md b/README.md
index b1ed7b6..8808338 100644
--- a/README.md
+++ b/README.md
@@ -1,4 +1,88 @@
 osmo-remsim - Osmocom remote SIM software suite
 ===============================================
 
-TBD
+This software suite is a work in progress.
+
+
+remsim-client
+-------------
+
+The client interfaces with GSM phones / modems via dedicated "Card
+Emulation" devices such as the Osmocom SIMtrace2 or sysmocom sysmoQMOD
+board + firmware.  This hardware implements the ISO7816-3 electrical
+interface and protocol handling and  passes any TPDU headers received
+from the phone/modem to remsim-client for further processing of the
+TPDUs associated to the given APDU transfer.
+
+remsim-client connects via a RSPRO control connection to remsim-server
+at startup and registers itself.  It will receive configuration data
+such as the remsim-bankd IP+Port and the ClientId from remsim-server.
+
+After receiving the configuration, remsim-client will establish a RSPRO
+data connection to the remsim-bankd IP:Port.
+
+As the USB interface for remote SIM in simtrace2.git uses one interface
+per slot, we can implement the client in blocking mode, i.e. use
+blocking I/O on the TCP/RSPRO side.  This simplifies the code compared
+to a more complex async implementation.
+
+
+remsim-bankd
+------------
+
+The remsim-bankd (SIM Bank Daemon) manages one given SIM bank.  The
+initial implementation supports a PC/SC driver to expose any PC/SC
+compatible card readers as SIM bank.
+
+remsim-bankd initially connects via a RSPRO control connection to
+remsim-server at startup, and will in turn receive a set of initial
+[client,slot]:[bankd,slot] mappings.  These mappings determine which
+slot on the client (corresponding to a modem) is mapped to which slot on
+the SIM bank.  Mappings can be updated by remsim-server at any given
+point in time.
+
+remsim-bankd implements a RSPRO server, where it listens to connections
+from remsim-clients.
+
+As PC/SC only offers a blocking API, there is one thread per PC/SC slot.
+This thread will perform blocking I/O on the socket towards the client,
+and blocking API calls on PC/SC.
+
+In terms of thread handling, we do:
+
+* accept() handling in [spare] worker threads
+** this means blocking I/O can be used, as each worker thread only has
+   one TCP connection
+** client identifies itself with client:slot
+** lookup mapping based on client:slot (using mutex for protection)
+** open the reader based on the lookup result
+
+The worker threads initially don't have any mapping to a specific
+reader, and that mapping is only established at a later point after the
+client has identified itself.  The advantage is that the entire bankd
+can live without any non-blocking I/O.
+
+The main thread handles the connection to remsim-server, where it can
+also use non-blocking I/O.  However, re-connection would be required, to
+avoid stalling all banks/cards in the event of a connection loss to the
+server.
+
+worker threads have the following states:
+* INIT (just started)
+* ACCEPTING (they're blocking in the accept() call on the server socket fd)
+* CONNECTED_WAIT_ID (TCP established, but peer not yet identified itself)
+* CONNECTED_CLIENT (TCP established, client has identified itself, no mapping)
+* CONNECTED_CLIENT_MAPPED (TCP established, client has identified itself, mapping exists)
+* CONNECTED_CLIENT_MAPPED_CARD (TCP established, client identified, mapping exists, card opened)
+* CONNECTED_SERVER (TCP established, server has identified itself)
+
+Once the client disconnects, or any other error occurs (such as card I/O
+errors), the worker thread either returns to INIT state (closing client
+socket and reader), or it terminates.  Termination would mean that the
+main thread would have to do non-blocking join to detect client
+termination and then re-spawn clients, so the "return to INIT state"
+approach seems to make more sense.
+
+Open topics:
+* detecting multiple connections from a server, logging this or even
+  avoiding that situation