tinyvoice — Fine-Tune a Model in Your Own Voice in an Afternoon

Active

MLXQwen2.5-0.5B (4-bit LoRA)PythonFastAPIServer-Sent Events

▶15-second product demo

Problem

Training your own language model sounds like a PhD job that needs a GPU cluster, so most developers never try. The tooling looks intimidating from the outside.

Business Impact

Fine-tuning is treated as inaccessible, so teams default to prompt-stuffing forever. tinyvoice is a working proof that small-model LoRA on-device is minutes and a laptop — a replicable recipe for anyone who wants a model in their own style.

System Approach

Dataset script pairs each published post's one-line brief with its final copy (26 train / 2 valid)
MLX LoRA on Qwen2.5-0.5B-Instruct-4bit, 400 steps, ~3 minutes on Apple silicon
FastAPI streams stock and tuned models side by side over Server-Sent Events
Single-HTML compare UI — same topic in, two voices out
Fixed an MLX streaming bug: inference state is thread-local, so all MLX work runs on one thread

Key Decisions & Trade-offs

0.5B 4-bit over a large model — the whole point is that it runs on a laptop
LoRA over full fine-tune — minutes, not hours, and a tiny adapter to ship
On-device MLX over cloud training — no keys, no bill, no data leaving the Mac
Side-by-side UI over a metrics table — the difference should be visible, not described

Current Status

Public and open source (MIT) at github.com/rohitguta2432/tinyvoice. The trained adapter is committed so the repo runs as-is; the stock-vs-tuned duel is verified live in the browser in the demo.

Roadmap

Swap in a larger base model to test whether knowledge (not just style) transfers
Feed the full post corpus as it grows
One-command train-and-serve script

What I'd Improve Next

Dataset de-duplication and augmentation for a cleaner signal
A quantitative voice-match score to measure the tuning objectively

Explore More

MyFinancial — Personal Financial Advisor — Financial planning in India is fragmented across banks, insurance, and tax documents.
PropCheck — AI Property Trust Score for India — Indian property buyers lose lakhs to fraudulent listings on Magicbricks, 99acres, Housing.
StellarMIND — Chat-to-SQL with RAG — Business users need to query databases without knowing SQL.
ClinicAI — WhatsApp AI Clinic Assistant — India has 12 lakh+ small clinics running on phone calls and paper diaries.
MicroItinerary — AI Travel Planner — Travel apps optimize for proximity and ratings.
SanatanApp — Hindu Devotional App — Devotional users in India juggle 5+ separate apps for Chalisa, Gita, Aarti, Ramayan, and Mahabharat.
SynFlow — Enterprise Intelligence Platform — Private deal networks rely on manual introductions and spreadsheets.
FinBaby (Jama) — Personal Finance Tracker — Indian middle-class families track expenses across UPI apps, bank statements, and paper notebooks.
RetailOS — Multi-Tenant Retail SaaS — Indian kirana stores and small retailers use paper registers or basic billing software with no inventory tracking, no GST compliance, and no offline support.
TripHive — Offline-First Collaborative Trip Planner — Group trip planning is fragmented across WhatsApp, Google Docs, Maps, Splitwise, and email.
ScamRakshak — On-Device AI Scam Detector — Indians lose thousands of crores annually to digital scams via WhatsApp, SMS, and social media.
PaisaGuard — Family Budget Survival App — Middle-class families worldwide track expenses inconsistently — UPI apps show transactions but don't enforce budgets.
rohitraj.tech — Engineering work is often invisible.
Agent Autopsy — Forensic Debugger for Failed AI Agent Runs — Most AI agent failures are silent — the run completes, the status code is green, and the result is still wrong.
snap3d — One Photo In, an Editable 3D Model Out — A photo shows you one side of an object; the other five sides are a guess.
Reliability & Production Readiness — Load testing, observability, and API contracts.
Open Source Repos — Browse the source code behind these projects.

System Approach

Dataset script pairs each published post's one-line brief with its final copy (26 train / 2 valid)

MLX LoRA on Qwen2.5-0.5B-Instruct-4bit, 400 steps, ~3 minutes on Apple silicon

FastAPI streams stock and tuned models side by side over Server-Sent Events

Single-HTML compare UI — same topic in, two voices out

Fixed an MLX streaming bug: inference state is thread-local, so all MLX work runs on one thread

Key Decisions & Trade-offs

0.5B 4-bit over a large model — the whole point is that it runs on a laptop

LoRA over full fine-tune — minutes, not hours, and a tiny adapter to ship

On-device MLX over cloud training — no keys, no bill, no data leaving the Mac

Side-by-side UI over a metrics table — the difference should be visible, not described

Explore More

MyFinancial — Personal Financial Advisor — Financial planning in India is fragmented across banks, insurance, and tax documents.

PropCheck — AI Property Trust Score for India — Indian property buyers lose lakhs to fraudulent listings on Magicbricks, 99acres, Housing.

StellarMIND — Chat-to-SQL with RAG — Business users need to query databases without knowing SQL.

ClinicAI — WhatsApp AI Clinic Assistant — India has 12 lakh+ small clinics running on phone calls and paper diaries.

MicroItinerary — AI Travel Planner — Travel apps optimize for proximity and ratings.

SanatanApp — Hindu Devotional App — Devotional users in India juggle 5+ separate apps for Chalisa, Gita, Aarti, Ramayan, and Mahabharat.

SynFlow — Enterprise Intelligence Platform — Private deal networks rely on manual introductions and spreadsheets.

FinBaby (Jama) — Personal Finance Tracker — Indian middle-class families track expenses across UPI apps, bank statements, and paper notebooks.

RetailOS — Multi-Tenant Retail SaaS — Indian kirana stores and small retailers use paper registers or basic billing software with no inventory tracking, no GST compliance, and no offline support.

TripHive — Offline-First Collaborative Trip Planner — Group trip planning is fragmented across WhatsApp, Google Docs, Maps, Splitwise, and email.

ScamRakshak — On-Device AI Scam Detector — Indians lose thousands of crores annually to digital scams via WhatsApp, SMS, and social media.

PaisaGuard — Family Budget Survival App — Middle-class families worldwide track expenses inconsistently — UPI apps show transactions but don't enforce budgets.

rohitraj.tech — Engineering work is often invisible.

Agent Autopsy — Forensic Debugger for Failed AI Agent Runs — Most AI agent failures are silent — the run completes, the status code is green, and the result is still wrong.

snap3d — One Photo In, an Editable 3D Model Out — A photo shows you one side of an object; the other five sides are a guess.

Reliability & Production Readiness — Load testing, observability, and API contracts.

Open Source Repos — Browse the source code behind these projects.