Hi,
A problem with IR is that you can't directly "see" what's happening. But many digital cameras are sensitive to IR, which may help. Ceratainly it would be much easier if you have access to the inside of the tank, but to expand on my suggestion in #4. :
Light will be also reflected from the front face of the glass so you will need a barrier betweeen the LED and photosensor. But simple plastic tubes (e.g. 3mm black sleeving) touching the front face may be all that's needed. Optical systems often use transparent "glue", but that may modify the beam width of the LED/sensor, adding to the unpredictability.
Fresnel's Equations indicate that about 4% of the light will be reflected at a glass-air (or air-glass) interface when there is "Normal incidence" (at right angles to the surface - ignore all the maths before that) and less than 1% at a glass-water interface, which could be easy to discriminate. BUT we need to consider what happens to the other 99% of the light. If the inside glass surface or the water is "dirty" then light will be scattered backwards, which might be similar to (or more than) that reflected back by the glass-air interface.
However, you have said that the water is clean, so where does the light go? If the inside of the tank is "bright" brass, then light shining directly in through the window may be refelected straight back, so you probably need to angle the LED and sensor slightly. That's where the beam widths could become important and a fair amount of experimentation may be required. Good luck.
Cheers, Alan.