[08]hackathon
★ First outstation hackathonEclipse 6.0
ParticipantMy first outstation offline hackathon: 24 hours at Thapar, Patiala. The brief was a real-time voice agent on a fine-tuned small language model. We built ASHA-VANI, an offline Hinglish voice assistant for ASHA health workers in rural India, with no internet needed. Every part worked on its own. Getting them all working together was the real boss fight.
- When
- April 2026
- Where
- TIET, Patiala, Punjab
- Hosted by
- ACM Thapar · Thapar Institute of Engineering & Technology
- Format
- in person · 24 hours
- Track
- SLM Voice Agents · Oumi sponsor track (EC602)

[01]The squad
Evolve AI
- Govind Jindal← that's me
- Manan Kochhar
- Piyush Sharma
- Bhavishya Grover
- Asmit Chitkara
[02]The rounds
Round 1
The shortlist
An online first round picked the teams for Thapar. Ours made the cut, which meant packing bags for Patiala.
Selected ✓
Round 2
The 24 hours
Round 2, offline at TIET on 4–5 April: frontend, backend, voice and dataset, split four ways and built side by side.
Built ✓
Round 3
The final submission
The last step was a Git repo of the whole thing. Ours wasn't put together yet, so we didn't submit it.
📦 Repo left behind
[03]What we built
ASHA-VANI

A fine-tuned 3B model
Qwen2.5-3B-Instruct fine-tuned with Oumi LoRA (rank 8) on about 1,800 Hinglish ASHA protocol pairs. Training loss fell from 2.76 to 0.76 in 312 steps.
4-bit, fully offline
The adapter is fused into the base model and quantised to 4-bit on Apple MLX, so it runs on a laptop with no network at all.
Hinglish speech-to-text
faster-whisper (small, int8) primed with a Hinglish medical vocabulary, plus rule-based fixes for the words it mishears, like "prasab" → "prasav".
A Hindi voice
Piper TTS with the Priyamvada Hindi voice reads every answer aloud, all on-device.
Calm triage
Words like "jhatke", "khoon" or "behosh" flag a case as urgent and shift the UI into a critical state, without alarming anyone.
Knows when to say "I don't know"
Out-of-scope questions (cancer, surgery, an ECG) get one answer: go to the PHC or call 108 right away. No guessing.
A latency budget
Each stage has a target: STT 150 ms, model 100 ms, speech 200 ms, under 500 ms end to end. Anything slower is flagged red.
[04]Side quests
The night before
The prep started the night before: ideas, roles and setup, all before we'd even reached Patiala. 🌙
Prep mode"Saheli"… or Jarvis?
The pitch had a custom wake word, "Saheli". The custom model never got trained, so the code fell back to a stand-in. Our ASHA assistant answered to "Hey Jarvis". 🦾
Hey JarvisSponsored by Quenzy
The desk slowly filled with sponsor chips and cans. Count them in the photos below. 🥔
Fuelled ✓The repo we left behind
The final step was a Git submission. Ours wasn't ready, so we walked out without submitting. A day later, back home, we pushed it anyway. 😅
Unsubmitted
[05]Plot twist
Four parts that worked. One app that wouldn't. 🧩

Then 😱
The split was perfect: frontend, backend, voice and dataset, one owner each. Every piece ran on its own laptop. We were sure that plugging them together would take ten minutes. 😌
Now 🧠
Integration is the project. The hard part was never one piece. It was the contracts between them: audio formats, ports, model paths, and four people with four different "works on my machine". 🔌
[06]The repo, after the fact
$ git log --reverse --format="%h %ad %s" --date=format:"%-d %b"
- e6b4c836 AprInitial commit
- 32aea9d6 AprFirst Commit-> backend + frontend
- 77734ee6 AprMerge branch 'main' of https://github.com/GovindJindal/ASHA-VANI
- 1b3132f7 AprFixed Voice Input and Output
- 11ef71e8 AprASHA-VANI MLX Migration and Cleanup
- 48d1dfd25 JunUpdated Lora and training updates
$
- Commits
- 6
- Avg. message
- 34.5 chars
- Shortest
- “Initial commit”
The hackathon ended on 5 April with nothing pushed. The first commit landed a day later, the voice was fixed on 7 April, and the 4-bit MLX migration came on the 8th. We didn't need a deadline to finish it. 🧩✅
[07]From the floor








[08]The certificate
Delivered ✓Certificate found. 📬
Round 2, offline hackathon, team Evolve AI, Thapar. 📜 Earned in 24 hours on roughly zero hours of sleep. 😴✅
“Splitting the work was the easy part. Joining it back together was the real project. Thapar taught me to agree how the pieces fit before writing any of them, and that an unsubmitted repo can still turn into a finished one.”
Next hackathon
NeuraCode 2.0: Where Logic Meets Glory