EDGE AI RESEARCH LAB · INDIA
ZeroShotLab
On-device AI agents for phones — local computer, efficient micro-models, privacy by default.
We build the stack where the model, tools, and memory stay on the device. First product ship is L2 Hinglish micro-models; the parallel track is a phone-native agent harness (file/shell sandbox + E2E messaging later). We do not start with a giant model zoo.
Website: zeroshotlab.in · GitHub: zeroshotlab-ai
Mission
Make capable AI agents practical on real ₹8–15k Androids: low latency, offline-friendly, and under the user’s control — not locked behind cloud APIs.
Product stack
| Track | What | Target ship | Status |
|---|---|---|---|
| L2 — Micro-models | Hinglish LID · transliteration · normaliser · PII (20–100M, ONNX int8) — also harness pre/post tools | Now → Jan 2027 | START HERE |
| Agent runtime | On-device harness + local file/shell (“own computer”) + OTS GGUF/LiteRT bases | Phase 0 Nov 2026 → Kotlin MVP Jan–Feb 2027 | Parallel with L2 |
| L3 — Instruct | Hinglish instruct 1–4B (LoRA/QLoRA ± DPO), quantised — only if L2 eval proves the gap | Feb 2027 (or when needed) | After L2 |
| L4 — TTS | Hindi/Hinglish on-device TTS | Mar 2027 (conditional) | Cut first if late |
| L5 — SDK | Edge SDK + Android sample (agent + L2 tools ± L3) | Apr 2027 | Alpha |
| L1 — Zoo | Indic edge model zoo + device benchmarks | Optional later | Deferred — not a gate |
Public roadmap (Nov 2026 → Apr 2027)
Now → mid-Jan 2027 — L2 alpha
- Freeze scope: 4 micros × ONNX int8
- Datasets + licences + rule baselines to beat
- ONNX int8 export + phone RTF measurements on budget Androids
- Train/distill micros; kill any run that does not beat rule baseline
- Ship gate: ≥2 of 4 models on this Hub + short write-up + Android/ONNX sample
Parallel — Agent Phase 0 → MVP
- Study harness patterns (AndroidHarness / AndroPI / Mochu — study only, no clone)
- Termux + off-the-shelf GGUF smoke on a real phone
- Native Kotlin harness MVP (Jan–Feb 2027) feeding L5
Feb 2027 — L3 (if needed)
- Tool-aware Hinglish instruct on OTS Gemma/Qwen-class bases
- Must beat base on held-out eval and run on a budget phone
Mar 2027 — L4 (conditional)
- Only with consented data + native-listener MOS
- If late: drop / slip — do not block L2/L3/L5
Apr 2027 — L5 SDK + agent sample
- Python + Android sample: L2 tools + OTS runtime ± L3
- Public HinglishEdge-Eval-style table
L2 detail — what ships first
| Micro-model | Job |
|---|---|
| LID | Language / script / Hinglish-mix detect |
| Transliteration | Roman ↔ Devanagari (scoped) |
| Normaliser | Clean noisy Hinglish for downstream |
| PII | Redact/flag phone, email, etc. before tools/logs |
Hard rules: no train-from-scratch · no L1 zoo as a blocker · every release needs a public eval · prefer Apache/MIT bases · Modal free / Kaggle first (budget alert at $30).
Week-by-week (L2 critical path)
| When | Milestone |
|---|---|
| 6–12 Nov 2026 | Scope freeze · orgs · LICENSES · Modal alert |
| 6–14 Nov | Dataset + licence shortlist · rule baselines |
| 9–20 Nov | ONNX int8 pipeline · phone RTF script |
| 16–29 Nov | First 2 micros trained vs baseline |
| 16 Nov–20 Dec | Eval set v0 (reviewed Hinglish) |
| 30 Nov–20 Dec | Remaining micros · HF packages · harness hooks |
| by mid-Jan 2027 | L2 alpha publish (≥2/4) |
Why investors / partners should care
- Clear wedge: phone-native agents + India-first Hinglish tooling — not another hosted chatbot.
- Lean path: L2 GPU budget on the order of tens of dollars; phones ₹20–40k test fleet; revenue from paid edge pilots, not a race-to-bottom API.
- Honest gates: ship only what beats baselines; L4 and L1 zoo are explicitly cuttable.
- Apr 2027 north star: public models + agent sample + eval table + pilot/LOI conversations from results.
Links
- Website: zeroshotlab.in
- GitHub org: github.com/zeroshotlab-ai
- Hub org: huggingface.co/zeroshotlab
India · Edge AI · Building in public · Plan as of 6 Nov 2026 IST · dates are targets, not promises