/ Projects

Things I've built

Open-source tools and full products — from an LLM quantisation optimiser to a multi-tenant hiring platform. Everything here has its code on GitHub.

Open source · LLM inference

OpenDynamicGGUF

Dynamic per-tensor mixed-precision GGUF quantisation — for any model.

Measures how sensitive every tensor is to quantisation, then picks the bit-width for each one under your size or VRAM budget — and ships a reproducible recipe, not just a smaller file.

  • Every bit decision traced to a measured ΔKL-divergence probe.
  • Recipes that rebuild the exact GGUF file, bit for bit.
  • One pipeline for dense, MoE and SSM models, built on llama.cpp.
View on GitHub Read the deep dive

odg — recipe.yaml

$ odg quantize --model functiongemma:latest \
    --target-size 3.2GB

token_embd         → Q8_0   # pinned
attn_v    (all)    → Q6_K   # sensitive, small
attn_q/k  (early)  → Q5_K
ffn_gate  (mid)    → Q3_K   # large, robust
ffn_down  (all)    → Q4_K
output             → Q8_0   # pinned
Language
Python
License
Apache-2.0
Status
Alpha
  • Python
  • PyTorch
  • llama.cpp
  • GGUF
  • Hugging Face
  • KL-divergence

Open source · Full-stack SaaS

Niyam

An open-source applicant tracking system for modern hiring teams.

A multi-tenant ATS in the spirit of Greenhouse and Lever — pipelines, structured interviews, e-signatures and an audit trail, with AI screening that cut manual work by 90%.

  • Workspace isolation with account-level data boundaries.
  • Stage-triggered e-sign automation and a full audit trail.
  • FastAPI services with Celery workers, Redis and S3-compatible storage.
View on GitHub

Hiring flow

  1. 01Kickoff
  2. 02Job published
  3. 03Applied
  4. 04Screening
  5. 05Interview loop
  6. 06Debrief
  7. 07Offer
  8. 08E-sign
  9. 09Hired
Language
Python · TypeScript
License
MIT
Status
Active
  • FastAPI
  • PostgreSQL
  • Redis
  • Celery
  • MinIO
  • React 19
  • TypeScript
  • pgvector
AI video

Kalki

An AI editorial engine that turns a talking-head clip into a finished, hook-first vertical reel.

Whisper transcribes, a DeepSeek agent pipeline writes the captions and on-screen design, and FFmpeg renders a finished reel — in English, Hindi or Hinglish.

  • faster-whisper
  • DeepSeek
  • FastAPI
  • Pillow
  • FFmpeg
Open source · Recommender systems

Kairo

Build recommendation systems, not recommendation pipelines.

Candidate generation, features, XGBoost ranking and evaluation in one guided platform — every dataset, model and deployment versioned, served through one clean API.

  • FastAPI
  • React
  • XGBoost
  • SQLAlchemy
  • Pandas

+ More on GitHub

Smaller experiments

GitHub profile
SecureLLM Security assessment and defence platform for self-hosted LLMs and AI agents. TypeScript ForgeAPI A Rails-inspired, convention-over-configuration layer for FastAPI — generators, migrations, background jobs. Python rails_ai_kit A Rails gem for vector classification, embeddings and similarity search on pgvector. Ruby TinyGPT A ~1M-parameter autoregressive language model built from scratch in JAX. Python Mimicra Personalised talking-head training from a short video. Python AutoPM An AI-native autonomous project manager for context-aware software execution workflows. Python