The brief was continuous breathing monitoring for patients outside intensive care, with nothing on the patient that needs a battery. Our team of five built a passive RFID tag that sticks to clothing. I led two parts of it, the sensor (a capaciflector) and the antenna, from the first simulations to the finished tags.

The first puzzle was the chip. It had a single sensing pin, and its datasheet only covered simple two-plate capacitors, but a capaciflector has three electrodes. Working from first principles, I routed the ground plane to the sensing pin so the chip drove both the reflector and the sensor electrode, and that went into the final design. I simulated the sensing principle in Multisim, and designed and simulated the antenna in ANSYS HFSS, including the effect of the wearer's body.

Then fabrication failed, twice. Our screen-printed sensors shorted because of air bubbles in the dielectric, in week 7 and again in week 8. On my supervisor's advice I switched to Kapton–copper clad and made it work reliably. I designed ten flexible-PCB layouts in KiCad, chose the four most promising to manufacture, fixed their manufacturing-rule errors and assembled the tags by hand. Testing raised two more problems. Long wires made the data noisy, and a more compact version lost read range because the sensor coupled to the antenna. A smaller sensor geometry fixed both.

In the end it read breathing clearly from a metre away. With breathing paced at 6 per minute by a metronome, we tested from 80 to 160 cm. The pattern was clear up to 100 cm, still recognisable at 120 cm and weaker beyond. 100 cm was the best balance between patient comfort and signal, with an average peak-to-peak swing of 75 raw counts. The reader picked up several tags at once, each with its own ID, so one reader could in principle watch a whole ward. We only had one working breathing sensor, though, so readings from several patients at once are still untested.

If I did it again, I'd test the fabrication route in week 1, not week 7.

An engine controller needed a clean vibration signal, scaled to 100 ± 5 mV/g, from a piezoresistive accelerometer, all on a single 5 V supply. I designed the whole chain, with the sensor modelled as a Wheatstone bridge, followed by an amplifier and cascaded second-order Butterworth filters passing 20–170 Hz, the vibration range of an engine running at 1,200–10,200 RPM.

My first amplifier quietly threw most of the signal away. A simple differential amplifier loaded the sensor bridge. I predicted 78% signal loss, and simulation showed 60%. It also couldn't swing negative on a single 5 V supply, so it clipped at +10 g. Raising its resistors tenfold still lost 26% on paper and 13% in simulation, more than the 10% a precision measurement can tolerate, so instead of retuning further I looked for the cause. The bridge's output impedance was too high for the amplifier's input impedance. A three-op-amp instrumentation amplifier, with its much higher input impedance, cut the loss to 0.35%, and biasing its output at 2.5 V kept the full ±10 g range inside the supply.

The amplifier delivered 102.5 mV/g, inside the 100 ± 5 mV/g target. With the filters added, sensitivity was 96 mV/g at 100 Hz and peaked at 110 mV/g at 50 Hz, 10% over target. Gain varied by 1.2 dB from 50 to 100 Hz, within the 1.5 dB flatness target. A final change cut the current draw by 7.1%, to 2.87 mA. I calculated every value by hand first, then checked it in Multisim. It has only been simulated so far, so building it on a protoboard is the obvious next step, to see real op-amp offsets.

The goal was to turn a bare piezoelectric disc into a working accelerometer. I designed a 3D-printed housing and a two-stage charge amplifier with a gain of 20, and first checked the disc itself on an LCR meter, which measured 18.44 nF at 1 kHz, within the datasheet tolerance.

The biggest peak was in the wrong place. On a shaker, from 0.5 Hz to 5 kHz with no added mass, 0.7 g and 1.4 g, the biggest peak sat near 850 Hz, far from the datasheet's 4.2 kHz. Rather than blame the sensor, I attributed it to the shaker's own resonance, since lab shakers typically resonate at a few hundred hertz. With mass added, a second peak appeared at 2.5–2.7 kHz and moved down as the mass grew, most likely the disc's own resonance.

Simulation and measurement didn't match exactly, and I could explain why. The protoboard added parasitics, a chain of series resistors stood in for one 10 MΩ part, and the probe loaded the circuit. The lab had no 30 MΩ resistor, so the low-frequency cutoff came out at 1.6 Hz instead of the specified 0.5 Hz. A reference accelerometer on the shaker is the first thing I'd add, to separate its response from the sensor's.

Three messages, three unknown ciphers, no keys. I wrote the attacks in Python, and every answer had to pass the same test. Re-encrypting the plaintext had to reproduce the original ciphertext exactly.

The substitution cipher beat pure statistics. Its letter frequencies still followed the shape of English, but mapping the most common letters straight onto e, t, a and so on produced nonsense. With only 269 letters the rankings were unreliable, and two letters were tied at 4.8%. Combining the frequencies with word patterns cracked it, and the finished key had a structure that explained why my first guess, a simple Caesar shift, had failed.

The hex file gave itself away by halves. A single-byte key, worked out from a hint about the first letter, turned every other character into English and left the rest garbled, and a brute force over all 256 single-byte keys confirmed none worked. The even and odd bytes also sat in different value ranges, which pointed to a 2-byte key, so I split the file into two streams and brute-forced each one separately.

The transposition took six attempts. The text's length, 84 letters, limited the possible grid sizes. Simple columnar, rail-fence, route, skip and row-rearrangement ciphers all failed, and I ruled out every rail-fence setting with a quadgram fitness score, a measure of how English-like text is. Then I brute-forced every column order of a keyed columnar transposition, scored each the same way, and one produced English.