Bengaluru, India
I help models become working software.
Dhriman Deka — an ML engineer in Bengaluru learning how to take models out of notebooks and run them as dependable services.
astroknot-sheep/meridian · feat: deterministic safety gates
Evals passing$ python -m meridian.deploy
▲ weights loaded — MentalRoBERTa ONNX in 0.8s
✓ RAG index warm · ✓ safety gates armed · ✓ 0 notebooks in prod
warming cache — blr edition
Notebooks are where models go to die
From data pipelines
to deployed services.
Selected Work
A few things I've helped build — a mental health agent with safety gates, a semantic search for an internal tool, and a multilingual hate-speech classifier. Each one taught me something about how projects actually run end to end.
Focus: LLM evaluation · retrieval · MLOps
Projects
Meridian 2025
Clinical-grade mental health agent — LangGraph orchestrator, safety gates, MentalRoBERTa screening
BERT semantic search 2025
Semantic search for an internal tool — private company project
Hate-speech detector 2024
Multilingual classifier fine-tuned on ~1.5M tweets — private company project
Pension portfolio optimiser 2023
Monte Carlo + classical optimisation on Indian pension data
Synthetic PAN/GST generator 2025
A small generative model for compliance test data — private company project
The Journey
Studied economics, spent some time in a research group, and now work on ML systems at Probe42. Still learning how all the pieces fit together.
Data Science & ML Engineer
Probe42 · Bengaluru
- + Moving models from Jupyter notebooks onto EC2 as reliable services with monitoring, retraining, and clean interfaces
- + Built a BERT-based semantic search and a generative model for synthetic compliance data, now used across internal tooling
Research Intern, LaRGo LLM Group
Prof. Kripa Bandhu Ghosh · Kolkata
- + Fine-tuned a multilingual hate-speech classifier across several Indian languages on ~1.5M tweets
- + Learned eval design, dataset curation, and reproducible pipelines
Economics
Econometrics, statistics, causal inference, portfolio optimisation
(Institute of National Importance)
Connected and ready
The technical arsenal.
Core ML
The daily drivers
Infra & Data
Where models meet production
Java ecosystem
For services that need the JVM
Learning & Ops
Sharpening the edge
Focus areas
What I care about
Papers, posts, and postmortems
Writing & publications.
Invited reviewer — SD4H Workshop @ ICML 2026
OpenReview
A comparative look at NPS and UPS from an employee's perspective
SSRN
Teaching AI to play Ludo: a 10,000-episode deep reinforcement learning journey
Medium
I built a transformer to predict stocks — it achieved 53.7% accuracy and lost all its money
Medium
I built a Python library to strip sensitive data from my training sets — here's what I learned
Medium
I built a virtual economy with 126 AI agents to see if they could survive
Medium
No pricing page here — just people
Talk to Dhriman.
Drop a message
$0 / foreverWrite what's on your mind — it lands straight in Dhriman's inbox.
Elsewhere
OpenCode, demos, and writing — the receipts live here.
- ↗GitHub — @astroknot-sheep
- ↗LinkedIn — /in/dhrimandeka
- ↗Hugging Face — astroknotsheep
- ↗Medium — @not_mordecai
See what's new on GitHub →
