OverviewLive
Eka is an AI companion with voice in English, Hindi and Kannada. Its four personas (founder / chanakya / gita / reflection) are written as system prompts and served through a FastAPI backend.
Published an eval comparing prompt-only and RAG setups. The two rows ran different model sizes, so it is not a like-for-like result, and the write-up says so next to the numbers.
Four supervised models trained on synthetic datasets and published on Hugging Face: a sentiment classifier, a summarizer, a complexity router, and a LightGBM ranker.
The fine-tuning chapter closed honestly rather than being padded out: personas ship as system prompts against Qwen2.5-7B-Instruct via Groq, since prompting already matched the fine-tune's measured ceiling.
Stack used
PythonFastAPIReactQdrantLightGBMGroqSarvam AIHuggingFace
What I learned
- •Training and publishing real models to Hugging Face, not just calling someone else's API
- •Knowing when to stop: shipping the honest, measured result instead of chasing a fine-tune that wouldn't have moved the needle
- •Building an eval harness before making any claims about model quality