1 of 5

CRISP

LUDDY HACKATHON / CASE 2 / 2026

CRISP.

Printed pages in. Compressed bytes out.

COMPRESSED RECOGNITION FOR IMAGE STORAGE PIPELINE

TEAM / SURYA · ANISH · LAHARI

2 of 5

STAGE 01

OCR Pipeline

IMAGE > TEXT

01

Denoise

CNN

02

Segment

CV

03

Recognize

CNN

90.6%

CHARACTER ACCURACY

111ms

MEAN STAGE 1 LATENCY

CAVEAT — DOMAIN GAP

Trained and benchmarked on EMNIST. Real-world smartphone scans differ in lighting, skew, font, and resolution — so live accuracy is lower than the benchmark.

3 of 5

STAGE 02

Adaptive Huffman

VITTER'S ALGORITHM V

2A

TEXT > BYTES

Compress.

Single pass, dynamic Huffman. Tree built as characters stream in. No dictionary transmitted.

1.88x

MEAN COMPRESSION RATIO

2B

BYTES > TEXT

Decompress.

Mirrors the encoder bit by bit. Rebuilds the same tree as it reads. 100% lossless round trip.

20/ 20

ROUND TRIPS VERIFIED

4 of 5

REAL-WORLD & VIABILITY

Why CRISP Ships

UNDER 200 MS / PAGE

01 / THE NEED

Scan-heavy archives.

Legal, medical, historical. Millions of printed pages stored as raw images.

Under 200 ms / page

02 / BUILT TO DEPLOY

Two containers, not a notebook.

Dockerized FastAPI services. Tested. Warmed. GPU preferred, CPU fallback.

20 / 20 lossless round trips

03 / ROOM TO GROW

Modular upgrade path.

U-Net denoiser. True-case recognizer. Arithmetic coder. Wire contract unchanged.

3 independent upgrade axes

5 of 5

CRISP / 2026

Q & A

Thank you.

Questions welcome.

SURYA ANISH LAHARI