Ashutosh
Kumar.

AI API cost tracker — unified spend dashboard for OpenAI, Anthropic, Google, OpenRouter.

Full-stack SaaS on Cloudflare's edge: SvelteKit on Pages, D1 (SQLite) for storage, a standalone cron worker for hourly usage syncs, budget alerts with email + Slack, and AES-256-GCM encryption for every API key at rest.

• → Hourly KV-locked cron worker syncs usage incrementally across 4 providers — idempotent upserts prevent double-counting on concurrent deploys
• → daily_rollups pre-aggregation: overview renders from 30 rows, not thousands of usage_records
• → Budget alerts with timezone-aware period math, email + Slack, deduped via last_triggered_at — no double-fire within the same window
• → AES-256-GCM encryption for all API keys and Slack webhooks before insert — ciphertext columns excluded from every list query
• → Magic-link auth (SHA-256 hashed token, 15-min TTL) + KV sessions — zero passwords in the database
• → Constant-time HMAC-SHA256 Lemon Squeezy webhook verification + test-mode guard to prevent silent $0 Pro upgrades

Kafka-driven async video transcoding pipeline — S3 → FFmpeg → ECS → multi-res.

A TypeScript pipeline that ingests video upload events via Kafka, fans them out to ECS workers running FFmpeg, and stores 360p / 720p / 1080p outputs back in S3 — a YouTube-style async processing architecture without the SaaS tax.

• → Kafka topic partitioning for parallel multi-resolution transcoding — 360p, 720p, 1080p in one pass
• → ECS task auto-scaling tied to consumer lag — spins up workers on demand, idles at zero cost
• → S3 pre-signed URL workflow for secure upload ingestion and output delivery (no public buckets)
• → Fully async end-to-end: producer acknowledges immediately, consumers commit offsets — zero polling loops
• → ~2.3x faster than sequential transcoding on a 5-min 1080p source (parallel fan-out vs single-worker)
• → Consumer-lag-based autoscaling: 0 → N workers in <90s, scale-down to zero when queue drains

Offline-first bulk email CLI — CSV, Sheets, SMTP, 10k+ recipients.

A Go CLI for sending personalised HTML email campaigns without SaaS. CSV and Google Sheets input, Go templates, any SMTP server, worker-pool dispatch. Ships as a single static binary for Windows, macOS, and Linux.

• → Worker-pool SMTP dispatch with connection reuse — 40% throughput gain over per-send dialing
• → Token-bucket rate limiting (configurable msg/sec + burst) to stay under provider limits
• → Exponential backoff retry that distinguishes transient 4xx from permanent 5xx failures
• → AND/OR/NOT rule engine over CSV columns — cut bounce rate by 10% on real campaigns
• → ~450 emails/sec on commodity hardware · 18 MB resident for 100k-row CSV (streaming)

Token-efficient code intelligence for AI coding agents

MCP server that indexes codebases and gives AI agents symbol-aware answers instead of raw file reads — 15–30× token reduction. Multi-tier search (exact → BM25 → fuzzy → semantic), call graph analysis with PageRank repo maps, incremental indexing via Merkle snapshots and parallel tree-sitter parsing, plus a Cloudflare Worker for auth, billing, and hosted embeddings.

• → Built a code intelligence MCP server in Rust — tree-sitter (9 langs), SQLite/FTS5, Tantivy BM25. 20 tools, 50K+file repos, 15–30× token reduction for AI agents
• → Multi-tier search pipeline — exact lookup (<1ms) → BM25 → fuzzy (nucleo/FTS5 trigram) → semantic embeddings (768-dim). Reciprocal Rank Fusion for hybrid merging across 4 tiers.
• → Graph analysis engine with custom CSR — bidirectional BFS (O(b^(d/2))), layered call traversal (50K visit cap),Union-Find components, push-based Personalized PageRank touching ~200–500 nodes vs full 50K.
• → Incremental indexing — Merkle snapshots (blake3), parallel parsing via lock-free pools (crossbeam), 250ms debounced watcher, WAL-mode SQLite with 256MB mmap zero-copy reads.
• → Cloudflare Worker API — GitHub OAuth + HMAC-signed licenses, Razorpay billing with webhook verification, 6-bucket rate limiting (fail-closed), KV-cached embedding proxy → Modal T4 GPU (~250ms warm).
• → Lock-free concurrency — ArcSwap caches, DashMap stores, single-flight graph builds, mimalloc. Sub-millisecond responses. Fat LTO + codegen-units=1 release. 8-crate workspace.