Sandeep Alluru — I tune models, make them fast.

INFERENCE_ENGINEERING

54.0 tok/s on one DGX Spark @ Q8.

This site is a brain, not a brochure. A 35B model served on one desk — every number measured, receipts included.

FOCUS: INFERENCE ENGINEERING
DECODE: 54.0 TOK/S @ Q8
RECEIPTS: 981

MODEL_FINE_TUNING

A model is only as good as its data.

140,545 training rows, decontaminated in-house — the gate caught 604 leaked held-out rows before they could reach training.

FOCUS: MODEL FINE TUNING
ROWS_TRAINED: 140,545
LEAKED_ROWS_CAUGHT: 604

[ MODEL_FINE_TUNING ]

MODEL_ACCELERATION

Same model. 1.54× faster.

MiniMax H3: 312.7 s → 202.5 s per clip on one DGX Spark. Nothing about the model changed — only the engineering.

FOCUS: MODEL ACCELERATION
CLIP_TIME: 312.7S → 202.5S
PER_STEP: 14.40S → 8.39S

[ MODEL_ACCELERATION ]

MOTION_GRAPHICS_ENGINEERING

Motion, engineered like software.

31 devices, 34 tracks, 91 of 106 samples pass their gates — failures published in their own files.

FOCUS: MOTION GRAPHICS
DEVICES: 31
SWEEP: 91/106 ALL PASS

[ MOTION_GRAPHICS ]

THE_MODELS

Small models, working models.

A 4B triage model under tuning. Qwen3.6-35B-A3B served locally by Tessen. MiniMax H3 behind the film work.

FOCUS: THE MODELS
TRIAGE: 4B — TUNING
SERVED: QWEN3.6-35B-A3B
FILM: MINIMAX H3 33B

THE_METHOD

Every number here was measured.

No benchmarketing. A figure goes on this site when there is a run behind it — a manifest, a workload, a raw output file. Misses stay in the table.

FOCUS: THE METHOD
CLAIMS: MEASURED ONLY
FAILURES: PUBLISHED

THE_WHOLE

So ask it yourself.

Tap a question and hear the answer in my voice — or skip the theatre and go straight to the work. Connect has the door.

FOCUS: ASK
SESSION: READY
ANSWERS: IN MY VOICE

[ ASK_THE_BRAIN ]

A model is only as good as its data.

140,545 training rows, decontaminated in-house — the gate caught 604 leaked held-out rows before they could reach training.

FOCUS

MODEL FINE TUNING

ROWS_TRAINED

140,545

LEAKED_ROWS_CAUGHT

604

[ MODEL_FINE_TUNING ]

Same model. 1.54× faster.

MiniMax H3: 312.7 s → 202.5 s per clip on one DGX Spark. Nothing about the model changed — only the engineering.

FOCUS

MODEL ACCELERATION

CLIP_TIME

312.7S → 202.5S

PER_STEP

14.40S → 8.39S

[ MODEL_ACCELERATION ]

Every number here was measured.

No benchmarketing. A figure goes on this site when there is a run behind it — a manifest, a workload, a raw output file. Misses stay in the table.

FOCUS

THE METHOD

CLAIMS

MEASURED ONLY

FAILURES

PUBLISHED

Sandeep Alluru — I tune models, make them fast.

54.0 tok/s on one DGX Spark @ Q8.54.0tok/sononeDGXSpark@Q8.

A model is only as good as its data.Amodelisonlyasgoodasitsdata.

Same model. 1.54× faster.Samemodel.1.54×faster.

Motion, engineered like software.Motion,engineeredlikesoftware.

Small models, working models.Smallmodels,workingmodels.

Every number here was measured.Everynumberherewasmeasured.

So ask it yourself.Soaskityourself.

Sandeep Alluru — I tune models, make them fast.

54.0 tok/s on one DGX Spark @ Q8.54.0tok/sononeDGXSpark@Q8.

A model is only as good as its data.Amodelisonlyasgoodasitsdata.

Same model. 1.54× faster.Samemodel.1.54×faster.

Motion, engineered like software.Motion,engineeredlikesoftware.

Small models, working models.Smallmodels,workingmodels.

Every number here was measured.Everynumberherewasmeasured.

So ask it yourself.Soaskityourself.

54.0 tok/s on one DGX Spark @ Q8.

A model is only as good as its data.

Same model. 1.54× faster.

Motion, engineered like software.

Small models, working models.

Every number here was measured.

So ask it yourself.

54.0 tok/s on one DGX Spark @ Q8.

A model is only as good as its data.

Same model. 1.54× faster.

Motion, engineered like software.

Small models, working models.

Every number here was measured.

So ask it yourself.