Skip to content
View Deepnar's full-sized avatar

Block or report Deepnar

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Deepnar/README.md

Deepesh Sonar

I build systems, then try to break my own claims about them.

LinkedIn Email X


Third-year Computer Engineering student in Mumbai. I work on information retrieval, long-term memory for language models, and evaluation — how you measure whether a system actually does what you claim, which turns out to be the harder half.

Most of what I build starts as a question I couldn't find a satisfying answer to. Lately those have been: what does a model actually need to remember, and how would you know if it remembered the right thing? Before that: can a constraint solver produce a timetable nobody has to fix by hand?

🔬 How I work

I write the criticism of my own work before someone else does. My largest project ships with a fidelity audit listing which of its components were genuinely defective, which the benchmark never exercised, and which actually carried the result — because ablation studies routinely conflate those three, and a component that never ran isn't a component that failed.

I'd rather state a result precisely than round it up. When my system matched a baseline instead of beating it, that's what I reported — along with the 32% context reduction that made the match interesting.

I document more than most people think is reasonable. Architecture notes, threat models, cost models. It's how I find out whether I actually understand something.

🛠️ Work

Project What it is
ice A local-first memory layer between any OpenAI-compatible client and a locally served model. Six retrieval legs fused with weighted Reciprocal Rank Fusion, a knowledge graph where superseded facts stay queryable as history, and a classifier gating whether retrieval fires at all. Evaluated with LSREP, a longitudinal protocol I designed for it. Paper submitted to ACM TIST · Apache-2.0
timetable-generator Constraint-driven scheduling. One engine covering seven timetable types, OR-Tools CP-SAT alongside a greedy solver, behind a hard-constraint registry that fails closed.
prompt-routing-classifier Multi-label topic and intent classification for routing prompts to specialised models. Built the pipeline end to end, from dataset construction to CPU inference.
DS-Practice Algorithms, data structures, OS and networking from coursework — written from the algorithm rather than adapted from a library.
micrograd-from-scratch Scalar autodiff and manual backpropagation, no ML libraries.

Off GitHub: I've taken a civic-reporting platform through two national and state competitions with a six-person team, where I wrote the security threat model and the infrastructure cost model, and led the pitch.

🐧 Elsewhere

Python · Java · PyTorch · FastAPI · PostgreSQL and pgvector · OR-Tools · Docker. I run Arch on a machine I partitioned the hard way.


Open to research and startup internships in retrieval, memory, or applied ML systems. If you're working on any of that, I'd like to hear from you.

Pinned Loading

  1. ice ice Public

    A local-first memory layer for conversational AI: persistent, structured memory for any OpenAI-compatible model, with a longitudinal protocol (LSREP) for evaluating it.

    Python 2

  2. timetable-generator timetable-generator Public

    Constraint-driven timetable scheduling service for colleges: FastAPI + PostgreSQL + OR-Tools CP-SAT, with cross-timetable safety, soft-constraint scoring, and PDF/CSV/iCal exports.

    Python

  3. prompt-routing-classifier prompt-routing-classifier Public

    Multi-label topic and intent classifier for routing prompts to specialised models: sentence embeddings + one-vs-rest logistic regression, CPU-only inference.

    Jupyter Notebook

  4. DS-Practice DS-Practice Public

    Algorithms, data structures, OS and networking programs from Computer Engineering coursework — implemented from the algorithm, not adapted from a library.

    Python

  5. micrograd-from-scratch micrograd-from-scratch Public

    A minimal autodiff engine and neural network from first principles: scalar computation graph, manual backpropagation, no ML libraries.

    Jupyter Notebook 1

  6. presentation-forge presentation-forge Public

    Self-hosted AI presentation generator: topic or sources in, researched themed editable PPTX and report out. Docker, BYOK or Ollama; 74 slide types, 34 themes.

    JavaScript 6