Pl. VI Singapore 新加坡
About·a working manuscript, revised 27 September 2026
Divyanshu Kumar
Research engineer. I work on making AI systems efficient, secure and private, and on what breaks when they are none of those.
I am an AI Research Engineer at Anaconda, which acquired Enkrypt AI this August. Most of my work sits on one stack: making models smaller and faster to run, finding out how they fail, building the guardrails that catch it, and keeping the data they learned from private. The stack keeps getting taller, which I don’t mind.
From Berai, Sarai, Bihar · B.Tech, Kolkata · IISc Bangalore · Enkrypt AI, ~300 models red-teamed · Anaconda, 2026
Ask this site something, or try to make it misbehave →§ 1the stack
What I work on
Quantization is how you get a large model onto hardware you can afford. In 2024 we found it also makes the model easier to talk out of its rules. I think about that result a lot, because it is the pattern. Efficiency, security and privacy are usually treated as separate subjects, and they are not. Change one and the other two move.
At Enkrypt I was the founding ML research engineer. I built the adversarial testing pipelines we ran against about 300 foundation models, and the production guardrails that came after. Guardrails are not free. Every refusal you add costs something in usefulness, and measuring that trade is most of the job.
Before that I was a Research Associate at IISc Bangalore with Prof. Prathosh A.P., on machine translation for Indic languages, machine unlearning in generative models, and privacy-preserving ML. That is where differential privacy got hold of me. I later wrote a four-part series on it, from the definition to DP-SGD.
Two questions are on my desk now. With Fabrizio Frasca at Technion: does an LLM’s answer about a graph change when you only change how the graph is written down? And in diffusion language models, which write by denoising rather than left to right: is there a step at which the output turns harmful, and can the model take it back?
My first research was text summarization at Helppr AI, in 2020. Outside work I read more code than I write, tinker with Go and ssh at weekends, and this June I finished BlueDot Impact’s Technical AI Safety course. If any of this is your problem too, write to me.
§ 2places
Where this happened
I am from Berai, a village in Sarai, Bihar. Since then I have lived in Patna, Kolkata, Delhi and Bangalore, in that order. Kolkata is where I studied electronics and communication engineering, at Narula Institute of Technology, until 2022. Bangalore is where IISc and Enkrypt happened. Singapore is where I presented our AAAI-26 workshop papers, in person. Malaysia, the Philippines and Bali are where I wander around.
Pl. VII Malaysia ماليزيا
Pl. IX Bali ᬩᬮᬶ
§ 3the record
Where I’ve worked
Newest first. Where a place has a plate in §2, it links there.
Work and research
Anaconda
Joined with Enkrypt AI’s acquisition.
Enkrypt AIacquired by Anaconda
Maintained Enkrypt’s public LLM Safety Leaderboard of about 300 models, and built an automated red-teaming system that finds issues in generative AI applications built by Fortune 500 companies.
Built proofs of concept and automated red-teaming for enterprise customers in finance, consulting and cybersecurity, and agentic document tools for fund administration: an LPA extractor with hybrid vector search, and an Excel formula-tracing agent with cell-level provenance.
Shipped low-latency guardrails on ONNX and Triton, a differentially private fine-tuning framework (PEFT and SFT), and SecureLLM, which uses homomorphic encryption.
Indian Institute of Science
Unlearning targeted features in GANs by perturbing the latent space, with differential privacy; tested on MNIST, CIFAR-10 and CelebA.
Knowledge-distilled Hindi–Kannada translation: a custom mBART, and an IndicBART encoder with part-of-speech embeddings.
FactMata
Worked on the Information Nutrition Label model, and on back-translation to augment its training data.
AIcrowd
Wrote baselines for NLP and vision challenges (LSTM, BERT, ALBERT, YOLO, Detectron).
Made ML puzzles for learners, with GANs and transformers.
HighRadius
Predicted when clients would pay with XGBoost (MAE 0.19).
Wrote Java web agents that parse 10 remittances a minute, in production.
Helppr.ai
Fine-tuned ALBERT, BERT, BART and T5 for summarisation.
Deployed one at 78% accuracy, with 30% lower inference time.
Open source
NimbleEdge
Built federated matrix reconstruction for recommendation, and the Sphinx docs for EnvisEdge’s Python APIs.
As a Season of Docs mentor I edited docs and tutorials, and ran weekly community meetups.
OpenMined
Mentored the Secure and Private AI course, on PySyft, with weekly check-ins on each mentee.
Education
Narula Institute of Technology
CGPA 8.78 out of 10.
Service
Reviewing
ICLR 2024, SET workshop.
NeurIPS 2024 and 2025, Datasets and Benchmarks track.
COLM 2025 and 2026.