OldKnowledge,New Vessel.
Essays on on-device AI, data sovereignty, and building systems that keep knowledge where it belongs.
Enter the archive →— 00 / ThesisEvery business runs on knowledge older than its software.
Latest Entries
Small Language Models in Healthcare 2026: Private, On-Prem Clinical AI That Fits Inside the Hospital
PixiJS in Production 2026: High-Performance 2D Web Graphics, WebGPU, and When 2D Beats 3D
Three.js in Production 2026: WebGPU, Realtime, and the New Web-Graphics Standard
AI SRE Agents: Architecture for Autonomous Incident Response
Why AI Proofs-of-Concept Die Before Production (and the Architecture That Ships)
अथ प्रथमोऽध्यायः ॥
विद्या ददाति विनयं विनयाद् याति पात्रताम् ।
पात्रत्वाद्धनमाप्नोति धनाद्धर्मं ततः सुखम् ॥
We digitize centuries-old manuscripts. Then we build with the same discipline.
AppScale's roots are in a quiet project: structuring classical Sanskrit texts into faithful digital form. Extraction, structure, provenance, sovereignty — the same principles now power our client work.
Built for Every Business
Intelligence that never leaves the device
Offline-first, data-sovereign systems — LLMs, vector search, inference at the edge.
02Document IntelligenceAir-gapped document vaults
Sealed, searchable, provable.
03AI ProductsFull-cycle AI product builds
Mobile-first, production-grade.
04MCP & AI SecuritySecuring the agentic stack
Tool poisoning, prompt injection, confused deputies — studied and hardened.
05Indic Language AIOCR for scripts the world forgot
Sovereign OCR for Indic scripts.
06Consulting & Architecture19+ years shipping at scale
OTT, e-commerce & mobility platforms serving millions.

Small Language Models in Healthcare 2026: Private, On-Prem Clinical AI That Fits Inside the Hospital
A specialized 7B model on the hospital’s own GPU never ships PHI off-site — and now matches larger models on medical exams. The private, on-prem clinical-AI architecture for 2026.

PixiJS in Production 2026: High-Performance 2D Web Graphics, WebGPU, and When 2D Beats 3D
The 2D particle scene flew on the laptop and melts on a real phone. PixiJS in production 2026: sprite batching, texture atlases, memory disposal, and WebGPU with WebGL fallback.

Three.js in Production 2026: WebGPU, Realtime, and the New Web-Graphics Standard
The 3D prototype hit 120fps on the laptop and crawls on a real phone. Three.js in production 2026: draw calls, memory disposal, asset pipelines, and WebGPU with fallback.

AI SRE Agents: Architecture for Autonomous Incident Response
The same agent that mitigates an incident in 3 minutes can cause one in 3 seconds. AI SRE agents: the diagnose-and-remediate loop, the five guardrails, and where to start.

Why AI Proofs-of-Concept Die Before Production (and the Architecture That Ships)
The demo dazzled the board; six months later it still is not live. Why 62-70% of AI pilots die in the PoC-to-production gap, the five gaps, and the architecture that ships.

Embedding Models Comparison 2026: OpenAI vs Cohere vs Voyage vs BGE
Head-to-head comparison of the top embedding models in 2026: OpenAI text-embedding-3, Cohere Embed v3, Voyage AI, and BGE. Benchmarks, cost per 1M tokens, context windows, and a decision framework for RAG, code search, multilingual, and self-hosted deployments.

Embedding Models Comparison 2026: OpenAI vs Cohere vs Voyage vs BGE
Head-to-head comparison of the top embedding models in 2026: OpenAI text-embedding-3, Cohere Embed v3, Voyage AI, and BGE. Benchmarks, cost per 1M tokens, context windows, and a decision framework for RAG, code search, multilingual, and self-hosted deployments.

Embedding Models Comparison 2026: OpenAI vs Cohere vs Voyage vs BGE
Head-to-head comparison of the top embedding models in 2026: OpenAI text-embedding-3, Cohere Embed v3, Voyage AI, and BGE. Benchmarks, cost per 1M tokens, context windows, and a decision framework for RAG, code search, multilingual, and self-hosted deployments.

Embedding Models Comparison 2026: OpenAI vs Cohere vs Voyage vs BGE
Head-to-head comparison of the top embedding models in 2026: OpenAI text-embedding-3, Cohere Embed v3, Voyage AI, and BGE. Benchmarks, cost per 1M tokens, context windows, and a decision framework for RAG, code search, multilingual, and self-hosted deployments.

KV-Cache Offloading: Serving 10x More Users by Not Recomputing
Your GPU re-prefills the same 15,000-token prompt ten thousand times a day. KV-cache offloading to DRAM and NVMe turns that recompute into a cheap fetch — 10x users.

Feature Flag Architecture: Rollouts, Kill Switches, and Flag Debt
Ship code to production without releasing it to everyone. Feature flags done right: release vs kill-switch types, sticky rollouts, failure defaults, and killing flag debt.

Distributed Locks: Redlock, Fencing Tokens, and Why Your Lock Doesn't Lock
A billing system double-charged 14 customers through a lock that worked exactly as documented. Redlock, fencing tokens, advisory locks, and when to delete the lock.

The CAP Theorem, PACELC, and What Real Databases Actually Choose
The pick-two triangle is a misunderstanding: partitions happen regardless. PACELC, quorum arithmetic, and what DynamoDB, Spanner, and Postgres actually choose.

Connection Pooling: PgBouncer, RDS Proxy, and the Serverless Postgres Problem
A team raised max_connections from 100 to 5,000 and throughput fell. Connection pooling done right: PgBouncer modes, RDS Proxy, sizing math, and the serverless trap.

Multi-Cloud AI: Cost-Aware LLM Routing Without the Lock-In
An enterprise spent nine months duplicating its AI stack across two clouds; every real incident was a quota ceiling. Multi-provider routing: the architecture that pays.

LLM-as-a-Judge vs Deterministic Heuristics: Who Grades the Model?
A judge scored 4.6/5 for months while users churned over wrong answers. When to use LLM-as-a-judge, when deterministic checks win, and the calibration that separates them.

Rust vs C++ for On-Device Inference Engines
Memory safety versus bare-metal control for mobile AI engines. Where the three crash classes live, why ecosystem gravity beats safety, and where to put the unsafe seam.

The CACTUS Framework: Automated Data Quality Gates for RAG
Your RAG hallucinates because the corpus is wrong, not the model. The six CACTUS data-quality gates, the heuristic-vs-judge cost split, and how to measure it.

Batch LLM Inference: Processing Millions of Documents Without Going Broke
A classification backfill through the chat endpoint turned a $400 job into $3,000 and still failed. Batch LLM inference: half the cost, resumable, built for finite jobs.

PromptOps: Managing Prompts as Code Before They Break Production
One prompt edit, shipped with no version history and no eval, silently regresses production for a week. PromptOps: prompts as versioned, tested, rollback-able code.
Bring this thinkingto your business.
conversation
One essay, most weeks. No noise.