Currently building document and voice AI for enterprise teams

Praveen Kumar

Principal AI Engineer & Solution Architect

Six years of it. Mostly speech AI

I take AI systems from the first rough prototype to the version that survives real traffic, real documents and real users. Everything I learn on the way ends up here, written down properly.

Currently building ThinkLess, a Python library for agent decisions, and CleanFrame, for cleaning data through saved recipes.

See the work About Resume GitHub

Selected work

Three I'd show you first

The project directory has the full collection. These three are the ones that taught me the most about shipping things other people depend on.

LLM & RAG Python pandas PyPI

CleanFrame

A data-cleaning engine you can actually put in a pipeline. An LLM writes the plan once; the plan compiles to a YAML recipe that replays over next month's file in pure pandas: no API calls, same answer every time. When the schema drifts it stops and tells you, instead of quietly corrupting the column.

On PyPI as cleanframe-engine Apache-2.0 Runs fully offline Read the write-up

MoM AI meeting minutes application homepage

LLM & RAG Flask Celery Redis Whisper Docker

MoM AI: Minutes of Meeting

Drop in a recording, get back minutes. FFmpeg normalizes the audio, a Celery worker routes it to one of six speech-to-text backends, including Whisper on your own machine, and an LLM returns the summary, the decisions and the action items. Both layers swap at runtime, from a dropdown.

6 STT providers, one runs locally Open source Read the write-up

Learning Resource Vanilla JS HTML CSS

Py Internals

An interactive course that teaches Python by drawing its memory. Code on the left, stack and heap on the right, one idea per animation step. No backend, no build step, no account: hand-authored snapshots rendered by a few hundred lines of vanilla JS.

9 sessions with interactive labs Open source Read the write-up

Browse the project directory

Writing

Notes from the workbench

Long-form, measured, reproducible. When a question keeps coming up at work I sit down with real hardware and answer it properly, then write down what the numbers said.

Which model should we use? Eight decisions in order, from writing the contract to sizing the bill, with the model chosen at step six

LLM & RAG Model selection Encoders vs decoders Evals Serving cost

Which model should we use? A production framework for choosing architectures, models, evals and training, from the use case to the bill

The order I make model decisions in, written so it works on a use case I haven't seen: the task shape, when an encoder beats a decoder and when it didn't, searching Hugging Face without being fooled, the eval set that actually decides, how much training a gap needs, languages and domains with no good model, and the memory and cost arithmetic. Six use cases worked end to end, and a one-page sheet.

18 October 2026 Long read Grounded in models trained, served and shipped Read the post

Value-level F1 on the same 4,577-document held-out PII test set: Kavach PII 270M v3 92.0, Kavach v2 79.0, GLiNER-PII 59.3, regex and checksum rules 30.3, Presidio 29.5

LLM & RAG PII detection Gemma 3 270M LoRA llama.cpp

Introducing Kavach PII 270M: a 270M model that finds personal data better than GLiNER, Presidio and spaCy, in 29 languages, trained on one laptop GPU

Updated for version 3: a quarter-billion-parameter model that scores 92.0 value-F1 where GLiNER gets 59.3 and Presidio 29.5, on the same held-out set with identical gold, and 90.4 on documents no language model touched. Why taking character offsets out of the training targets was the decision that made it work, the re-annotation that found the old test set missing one value in ten, how 43 labels across 29 languages were built and checked, the one number that got worse, and the three things that failed.

23 September 2026, updated 29 September Long read Every number from the same held-out set Read the post

Two minutes of a held-out meeting: the labelled speakers above, the container's diarization below, one colour per person

Speech & Audio Speaker diarization Kaldi PLDA Docker

Who spoke when, on your own data: training, tuning and shipping a Kaldi speaker diarizer that runs on a CPU

From a folder of labelled recordings to a container that writes an RTTM in seconds on a CPU: the x-vector pipeline stage by stage, the back-end trained in under a minute, the threshold cliff a trained PLDA creates and the Bayesian HMM that fixes it, every knob tuned on dev and measured on held-out meetings, and the honest comparison with pyannote and NVIDIA's Sortformer.

20 September 2026 Long read Every number measured on held-out meetings Read the post

Browse all posts

The short version

Models are the easy part

The hard part is everything around them: the queue that retries, the tenant boundary that holds, the GPU bill that doesn't spiral, the transcript a human can actually correct. That's the work I've spent six years getting good at: speech intelligence, document processing and LLM automation, mostly for enterprise teams who need it to hold up on a Monday morning.

Years building AI systems

10+

Production AI platforms

50+

Models deployed

Projects in the directory

The longer version

Say hello

Got something hard to build?

I like problems where the obvious approach doesn't work yet. Email is fastest, but any of these reach me.

Email LinkedIn GitHub Twitter Resume