About

Offline AI That Ships

PlanCodeShip builds offline AI applications for underserved populations in India. Three apps — Yojana-Sathi (government schemes), Krishi-Edge (crop advisory), and Shiksha-Edge (NCERT tutor) — run entirely on $100 Android phones with zero internet dependency.

The technical stack: fine-tuned Qwen 2.5 models with LoRA on MLX, quantized F32→F16→Q4_K_M, compiled via NDK for ARM64, served through a JNI bridge to llama.cpp. Shruti STT (29MB) for Hindi voice. Bundled SQLite with FTS5 for BM25-ranked search. The commercial focus is ChatGo — 6 tools to cut WhatsApp Business API costs by up to 75% ahead of Meta's October 2026 pricing change.

Founder

Amit Saharana

IIT Delhi CS · Codeforces Grandmaster (2416, 3rd in India) · 61st in Google Distributed Code Jam (1st in India) · AIR 79 / 1.15M in AIEEE.

10 years building low-latency systems: crash reporting at Facebook (Menlo Park), multi-region video routing and AI transcription at 100ms (E5), LLM orchestration and RAG at Wokay (CTO, 50K+ daily interactions).

Full resume and experience →

Principles

How We Build

On-device, not cloud-dependent

llama.cpp compiled for ARM64 via NDK. SQLite with FTS5 bundled natively. Shruti STT at 29MB. Models quantized to Q4_K_M. Everything runs offline on budget phones.

Measured, not assumed

30/30 voice-to-search E2E accuracy. 10/10 scope gate accuracy. 22.5 tok/s on ARM64 emulator. 348ms model load. Every claim has a benchmark.

Infrastructure depth over wrapper depth

Custom JNI bridge with UTF-8 Devanagari handling. F32→F16→Q4_K_M quantization pipeline. Post-STT normalization with transliteration. Not thin API wrappers.

Research with a revenue path

EdgeHub is the technical moat. ChatGo is the revenue driver — 6 tools for WhatsApp Business cost optimization ahead of Meta's October 2026 pricing change.

Timeline

Milestones

  1. 2025

    PlanCodeShip founded. Built unified Go backend, multi-agent orchestration, and RAG pipeline infrastructure.

  2. 2026 Q1

    Shipped Jev (agent safety guardrails) and LocalReader (on-device neural TTS for iOS).

  3. 2026 Q2

    Built Shiksha-Edge — offline AI tutor running on ₹8,000 phones with llama.cpp and LoRA fine-tuning.

  4. 2026 Q3

    Launched ChatGo — 6 tools to cut WhatsApp Business costs ahead of Meta's Oct 2026 pricing change.

  5. 2026 Q4

    Built EdgeHub Android — compiled llama.cpp for ARM64, fine-tuned 3 models (Yojana/Krishi/Shiksha), integrated Shruti STT, bundled FTS5 SQLite. All verified on API 35 emulator. 22.5 tok/s, 100% E2E accuracy.