Srijan Upadhyay

(About) who am I?

I build the systems large models stand on, and the companies that put them to work.

0PRs landed in PyTorch, TensorFlow, NVIDIA and HuggingFace

0merged PRs across public open source

0cricket deliveries in Midwicket

0GPA, B.Tech Data Science at SPPU

I'm a B.Tech Data Science student at Savitribai Phule Pune University, graduating 2028. My craft is the infrastructure beneath machine learning: process groups that deadlock, losses that diverge in bf16, models that fracture when sharded across GPUs. I trace each failure to its root, write the smallest fix that holds, and send it upstream.

That work now runs inside Megatron-LM, PyTorch, TensorFlow and HuggingFace Transformers, and TensorFlow maintainers invite me to review their security and correctness fixes. I co-founded LakkeSide, a GPU-backed AI agency running 20 H100s, and I also run ScoreOne, an AI-backed JEE and NEET prep platform with 2,500+ students taking real exams on it.

(Experience) where I build for clients

2022 to now, Pune

Co-Founder & Technical Lead,
LakkeSide

A GPU-backed AI agency run by five founders on a fleet of 20 H100s. I architect its ML deployment stack, from agentic pipelines that write, test and ship code through CI/CD to the technical roadmap for LLM inference, multi-GPU orchestration and backend reliability.

0H100 GPUs in the fleet

0co-founders

0concurrent client deployments

<100msretrieval over 8,000+ embeddings

Enterprise financial services

Payment processing pipeline

Led the ML engagement for a confidential client and designed a high-throughput transaction pipeline with lower processing latency.

Government research institution

Sinhala and Nepali to English

Led the NLP engagement: a multilingual transliteration pipeline covering script normalisation, tokenisation and sequence mapping over large text corpora.

Production ML

Vision, NLP, inference

Delivered computer vision, NLP and distributed inference systems, including vector retrieval at sub-100ms query latency.

(Founder) my startup

Apr 2026 to now, scoreone.in

Founder & Lead Engineer,
ScoreOne

An AI-backed JEE and NEET prep platform. Students sit full mock exams that score exactly like the real NTA paper, and coaching institutes run their whole batch on it. I lead it end to end: product, architecture, security, and every reliability call at 3am.

Visit scoreone.in ↗

0students on the platform

0merged PRs

0questions in the JEE + NEET banks

0OMR accuracy on answers it commits to

Exam engine

Scores like the real paper

NTA-parity marking, including negative marking on integer questions, and checkpointed attempts that survive a dropped connection or a closed tab.

Proctoring

Cheating the client can't fake

Tab-switch, fullscreen-exit and split-screen detection, with every violation recorded server-side where a tampered browser cannot reach it.

Institutes

A whole coaching centre on one login

Organisation, branch, batch and student levels with strict tenant isolation, a 500-student bulk roster import, test assignments, and per-student behaviour flags from per-question timing.

(Founder) the stars, read by hand

Nov 2025 to now, thekarmacompass.co.in

Founder & Palm Reader,
The Karma Compass

5000+ seekers guided, across the globe.

I founded The Karma Compass with my brother, who joined as co-founder and reads the charts while I read palms. Seekers from across the globe come to us for Vedic and Western birth charts, Rashi charts, palm readings, children's readings and kundli matching, and every one leaves with a personalised report. I lead the practice and the business, and I read every palm myself.

Visit The Karma Compass ↗

0seekers guided worldwide

1000spersonalised reports delivered

0kinds of reading

2founders, brothers

Vedic chart
Rashi chart
Western chart
Palm reading
Children's reading
Kundli matching

For seekers

Book in minutes, from anywhere

Choose a reading, share your birth details and pay securely online, from any country, without a single phone call.

After the booking

Speak to your reader

Every seeker gets a private line to their reader: live chat, palm photos and charts shared securely, and voice or video calls.

Trust

Your chart stays yours

Birth details, palm photos and conversations are private by design and never shared. Trust is the whole practice.

(Founder) Banarasi, woven for the world

Live, srijan-sarees.vercel.app

Founder & CEO,
Srijan Sarees

10 lakh+ pieces sold. Crores in revenue.

One of the largest Banarasi saree manufacturers, dressing women across India and around the world. I founded Srijan Sarees and lead it as CEO, from the looms to the last delivery: sarees, suits, dupattas and purses, each sold at one honest price, with every order handled personally.

Visit Srijan Sarees ↗

10 lakh+products sold

Croresin revenue

Globalacross India and abroad

1 tapto order on WhatsApp

Banarasi sarees
Suits
Dupattas
Purses
Every Indian weave

Manufacturer

From our looms to you

We make what we sell. Buying straight from the manufacturer means real Banarasi craft at an honest price, with no middlemen.

Reach

India and the world

Shipped pan-India and abroad. Customers order on WhatsApp and speak to us directly, the way a family business should feel.

Ahead of the market

A modern saree house

AI product photography, an online catalogue with one-tap colour swatches, and a guide to every Indian weave. Tradition, run like a modern brand.

(Co-founder) helping people get hired

Mar 2025 to now, JobCoachX

Co-Founder,
JobCoachX

Interviews, placed fast.

Students, career switchers and working engineers come to us when applying alone has stalled. We get on a one-to-one call, learn their strengths and gaps, rebuild the resume and LinkedIn, then take them straight to hiring managers with a referral. Most land interviews within one to two months. My brother coaches the conversation, including salary negotiation; I build the AI tools: an interview helper and an auto applier.

Visit JobCoachX ↗

0people consulted

1-2months to interviews

24/7on call, WhatsApp, LinkedIn

0video playbook for every member

01
Know the person
One-to-one call to map strengths, gaps and the job they actually want.
02
Fix the profile
Resume fine-tuned and LinkedIn rebuilt so recruiters stop scrolling past.
03
Open the door
A warm introduction to the hiring manager, backed by a referral.
04
Win the offer
Interview prep and salary negotiation, so the offer matches what they hoped for.

Who we help

Tech, sales, marketing, finance

Mostly technical roles, and anyone seeking a new job or a fresh career. Success is measured on payday, when the first salary lands.

Our promise

No result? We take over

Members get a six-video playbook and our tools to run the search themselves. If a month passes with nothing, we take it into our own hands and line up interviews directly.

What I built

Two AI tools, free with every package

AI interview helper: my own Cluely-style assistant for interview prep. Auto applier: applies to matching jobs automatically. Plus the JobCoachX site itself.

(What I do) where I'm useful

Where I go deep.

(1)

Distributed Training

Multi-node pretraining and fine-tuning that launches cleanly, stays in lockstep and converges. I work across the parallelism stack: data, tensor, pipeline, sequence and expert parallel, and the collectives that hold them together. Rendezvous and process-group init, world-size and rank derivation, MoE router numerics, mixed-precision drift, checkpoint resharding across topologies.

What it buys you: A 64-GPU run that gets past init on the first try, survives a node loss, and keeps a finite loss at step 40,000.

FSDP2 / ZeRO-3
Megatron-Core TP/PP/CP
Expert parallelism
NCCL collectives
FP8 delayed scaling
Distributed checkpointing

(2)

LLM Inference & GPU Kernels

Serving paths that are fast and economical, from the scheduler down to the kernel. Paged KV-cache, continuous batching, speculative decoding, weight-only and activation quantisation, and kernels written and profiled until the roofline says stop.

What it buys you: Same model, higher tokens per second per dollar, with a benchmark that proves every claimed gain.

PagedAttention
Speculative decoding
AWQ / GPTQ / FP8
Triton kernels
CUDA / Numba-CUDA
TensorRT-LLM
Nsight profiling

(3)

Compilers & Hardware Enablement

Making PyTorch run correctly on silicon it was not born on. I work as a PyTorch RISC-V Community Partner on riscv64 enablement: vectorised CPU backends, runtime feature detection, cache topology, and the graph-compiler paths that lower models onto new hardware.

What it buys you: Models that compile and run on new architectures without silent numerical divergence.

RISC-V RVV
ATen vectorisation
TorchInductor
torch.compile / Dynamo
MLIR
cpuinfo

(4)

RL Post-Training & Alignment

Turning a pretrained model into one that reasons and follows preference. Policy-gradient and preference-optimisation pipelines with rollout generation, reward modelling, KL control and evaluation harnesses that catch reward hacking before it ships.

What it buys you: A post-training loop where reward goes up because the model got better, and you can prove which one it was.

GRPO / PPO
DPO / KTO
Reward modelling
KL-regularised RL
LoRA / QLoRA
Eval harnesses

(5)

Document AI & Vision

Research and production vision systems: formula and handwriting OCR with encoder-decoder transformers ranked on character error rate, optical mark recognition for exam sheets, and face-embedding retrieval with zero-downtime re-indexing on model swap.

What it buys you: Scanned pages become structured data you can trust, measured on held-out sets.

Vision-encoder decoders
CER / WER evaluation
Optical mark recognition
ArcFace embeddings
HNSW vector search
DPDP-compliant pipelines

(6)

Building Companies

Founder of ScoreOne, The Karma Compass and Srijan Sarees, co-founder of JobCoachX and LakkeSide. Product, architecture and engineering under one roof: multi-tenant platforms, server-authoritative scoring and proctoring, payments, and security reviews before a breach finds the gap.

What it buys you: Products real people pay for and depend on, from exam halls to saree orders across India.

Multi-tenant architecture
Server-side proctoring
Threat modelling
Payment integrations
TypeScript / React
Edge deployment

(7)

Framework Correctness & Open Source

88 merged pull requests outside my own repos, 16 of them in PyTorch, Megatron-LM, TensorFlow, TensorRT-LLM, NeMo and Numba. Silent dtype promotion, broadcasting edge cases, meta-function shape checks, and CI that was green because it could not fail.

What it buys you: Bugs fixed upstream once, for everyone, with a regression test that keeps them fixed.

Minimal repros
Meta-function validation
Dtype and shape semantics
Regression testing
CI sharding and flake triage
Upstream review etiquette

(8)

Agents & Retrieval Systems

LLM systems that act, not just answer: tool-calling agents grounded in a curated corpus, real-time assistants that listen and respond inside a live call, and retrieval pipelines with chunking, embedding search and citation so every answer traces to a source.

What it buys you: Assistants that answer from your data, cite it, and say so when they do not know.

Tool calling
Corpus grounding
Embedding search
Grounded generation
Streaming inference
Agent evals

(9)

Speech & Audio

Speech pipelines for Indian languages and long recordings: timestamped ASR that turns long recordings into clean transcripts, and neural text-to-speech for Marathi, a low-resource language most models handle badly.

What it buys you: Hours of audio become searchable text, and Indian languages get a voice that sounds native.

Whisper ASR
Timestamped transcripts
Neural TTS
Low-resource fine-tuning
Forced alignment
Streaming audio

(10)

Data Engineering at Scale

Corpora and datasets built to be trained on: a 700,000-record scrape across 42 Indian cities, an 18,781-artifact research corpus, a past-paper question bank for exam ML, and daily market-data collectors. Deduplication, schema contracts and provenance on every row.

What it buys you: Clean, deduplicated, provenance-tracked data, so the model learns the signal and not the scraping noise.

Distributed crawling
Deduplication
Schema contracts
Dataset versioning
Parquet / Arrow
Backfill pipelines

(How I work)

01
Reproduce
Every fix begins with a failing reproduction. If I cannot make it break, I do not claim to have fixed it.
02
Root cause
I trace each failure to its cause, reading the framework source beneath the stack trace.
03
Smallest fix
The fewest lines that hold, guarded by one test that fails without them.
04
Ship and verify
Upstream, green, and verified in main before I call it done.

Guinness World Records

Technical leader on a Guinness World Record.

Led the technical effort for SPPU's Department of Technology at the Pune Book Festival in December 2024, an effort recognised with two Guinness World Records certificates.

2Guinness certificates

Dec 2024Pune Book Festival

(Open source) merged upstream

My code runs inside the frameworks you train on.

riscv64 / rv64gc

PyTorch RISC-V
Community Partner

Named contributor on PyTorch's RISC-V enablement roadmap (#180975) alongside the XuanTie and Ruyi teams and PyTorch core maintainers. Landed RISC-V fast paths in c10 and ATen, fixed the cross-compile toolchain, and run a native riscv64 CI lab on RISE hardware.

3RISC-V PRs landed

5phases on the roadmap

rv64native CI lab

Open source, NVIDIA ecosystem

Distributed systems engineer across NVIDIA's AI stack

An independent contributor working in the open across NVIDIA's training and inference repositories: distributed init in Megatron-LM, tensor-parallel weight splits in TensorRT-LLM, precision in MoE routing, and the tooling around them.

6PRs landed

5NVIDIA repositories

TP/MoEparallelism + precision

Megatron-LM #2571distributed init deadlock Megatron-LM #2765fp32 MoE routing TensorRT-LLM #15351tensor-parallel QKV split TransformerEngine #3195datatype + hardware matrix NeMo-Run #562PEP 604 union dispatch numba-cuda-mlir #228sub-record regression test

16 PRs landed in the big frameworks, 65 more in the ThinkRL RL library.

PyTorchpytorch #189685 Fix int32 overflow in CUDA bincount/histogram bin index Sep 2026 PyTorchpytorch #190003 Forward cross-compilation toolchain variables before project() Jul 2026 PyTorchpytorch #188999 RISC-V CPU pause hint in the cpu_atomic_add_float spin loop Jul 2026 PyTorchpytorch #189552 Fix multilabel_margin_loss decomposition for -1 padded targets Jul 2026 PyTorchpytorch #189000 RISC-V rdtime fast path for the profiler's getApproximateTime Jul 2026 TransformerEngineNVIDIA #3195 Datatype and hardware support matrix docs Sep 2026 NeMo-RunNVIDIA-NeMo #562 Dispatch PEP 604 unions (X | Y) in TypeParser Aug 2026 numba-cuda-mlirNVIDIA #228 Regression test for nested scalar sub-record field reads Aug 2026 TensorRT-LLMNVIDIA #15351 AutoDeploy: rescale fused QKV split sizes under tensor parallelism Jul 2026 Megatron-LMNVIDIA #2765 Sigmoid in fp32 for MoE aux loss stability with 32+ experts Feb 2026 Megatron-LMNVIDIA #2571 Fix world-size mismatch causing a distributed init deadlock Dec 2025 TensorFlowtensorflow #122563 Fix tf.experimental.numpy.compress XLA compilation failure Aug 2026 TensorFlowtensorflow #122567 Support tf.bitcast on complex tensors under XLA Jul 2026 TensorFlowtensorflow #121252 Clarify stateless RNG consistency guarantees under XLA Jul 2026 TensorFlowtensorflow #102094 16KB page alignment for the core JNI library (Android 15+) Oct 2025 Transformershuggingface #43047 Stop accelerate splitting the timm vision encoder across devices Apr 2026

Code review, TensorFlow core

Maintainer-requested reviewer on #118744 (tf.bitcast / tf.cast eager vs XLA divergence). Reviewed an OOB write in the MaxPoolGradWithArgmax GPU kernel (CWE-787, #115326), unbounded recursion in Variant deserialization (#118269), a Grappler ArgMin/ArgMax rewrite (#116709) and rank invariants in StringNGrams (#118868).

(Works) things I built

Selected works.

01 / 12

01 Founder, Apr 2026 to now

ScoreOne

AI-backed JEE and NEET prep. NTA-parity scoring, server-side anti-cheat the client cannot fake, OMR scanning, institute dashboards. 180+ merged PRs.

TypeScript / React / Back4AppView ↗

02 Open-source SDK on PyPI

Midwicket

Local-first cricket intelligence over 20,000+ matches and 9.1M+ deliveries. Win-probability models with a versioned registry and rollback.

DuckDB / PyArrow / FastAPIView ↗

03 Founder, live

The Karma Compass

Founded with my brother. Vedic, Western and Rashi charts, palm readings and kundli matching for 5000+ seekers worldwide.

HTML / JS / PHP / RazorpayLive ↗

04 Founder & CEO

Srijan Sarees

One of the largest Banarasi saree manufacturers. 10 lakh+ pieces sold across India and abroad, crores in revenue.

Next.js / Vercel / AI imageryLive ↗

05 Co-founder, live

JobCoachX

Co-founded career coaching: resume, LinkedIn and hiring-manager referrals to get people interviews fast. 1500+ consulted.

HTML / JS / VercelLive ↗

06 SPPU, Nov 2025 to Mar 2026

FotoOwlAI

Face recognition attendance across four cloud services: T4 inference, Qdrant embeddings, a Telegram bot, Google Drive. DPDP consent built in.

InsightFace / Qdrant / RailwayAcademic

07 SPPU, Department of Technology

Event Image Retrieval

Find every photo you appear in. Face search over 2,000+ event images and 8,000+ embeddings with sub-second results, cutting about 90% of manual searching.

InsightFace / FAISS / FastAPIAcademic

08 Open source

Engram

An open-source memory layer for LLM agents.

PythonView ↗

09 Full stack

TranscriptAI

Paste a URL, get a transcript. Whisper backend with a zero-dependency local start.

FastAPI / Whisper / Next.jsView ↗

10 Research

SLM LoRA

PEFT/LoRA fine-tuning of Pythia-160m on 10 datasets, tracked in W&B, benchmarked on Kaggle T4.

PyTorch / PEFT / W&BView ↗

11 Data tooling

Data Sanitizer

Schema detection, dedupe, outlier detection and LLM enrichment for messy datasets.

Python / LLMsView ↗

12 Data engineering

NSE Collector

Five-year backfill plus a daily job collecting Indian market data into a versioned dataset.

Python / ActionsView ↗

(Journey) so far

The path so far.

2022

Co-founded LakkeSide

A GPU-backed AI agency with four other founders. Now 20 H100s and 3+ client deployments running at once.

Oct 2024

Started B.Tech Data Science

Savitribai Phule Pune University. ML, distributed systems, computer vision. GPA 8.5/10.

Dec 2024

Guinness World Record

Technical leader for SPPU's Department of Technology at the Pune Book Festival.

Mar 2025

Co-founded JobCoachX

Interviews, placed fast. 1500+ people consulted since, from students to career switchers.

Oct 2025

First merge: TensorFlow

16KB page alignment for the JNI library, unblocking Android 15+ apps on Google Play.

Nov 2025

Founded The Karma Compass

With my brother as co-founder. I read palms, he reads charts. 5000+ seekers guided since.

Dec 2025

Into NVIDIA Megatron-LM

Fixed a distributed init deadlock. Followed in Feb with an fp32 MoE stability fix.

Apr 2026

Founded ScoreOne

Started building ScoreOne, AI-backed JEE and NEET mock exams that score like the real paper.

Jul 2026

PyTorch RISC-V Community Partner

Landed RISC-V fast paths in PyTorch core and joined the riscv64 enablement roadmap. TensorRT-LLM merge the same month.

Sep 2026

65 PRs into ThinkRL

Rebuilt an RL training library's GRPO, PPO and checkpoint paths. TransformerEngine and a PyTorch CUDA fix land too.

Now

Shipping, and graduating 2028

Leading five ventures, advancing PyTorch on RISC-V, and seeking the hardest problems in ML systems.

ML systems

Distributed training, FSDP, MoE, NCCL, async scheduling, vLLM, TensorRT

Frameworks

PyTorch, TensorFlow, Transformers, Megatron-LM, TransformerEngine, NeMo

Infra / CUDA

CUDA, Numba-CUDA, Triton, ONNX Runtime, RISC-V, H100 cluster management

Languages

Python, TypeScript, JavaScript, Java, R

N g1

(Off the clock) the 64 squares

Grandmaster^*

*Not a grandmaster. A deeply committed enthusiast.

Bullet

1004

Thinks fast. Blunders faster still.

Rapid

477

More time on the clock, more ways to lose.

Blitz

448

Somewhere between bullet confidence and rapid regret.

Puzzles (best)

1117

Unbeatable when the other side cannot move.

Challenge me on chess.com ↗ Ratings from chess.com