Ishaan Chaturvedi

• Self-hosted, live, paid AI product on an 11-service TypeScript / NestJS architecture (~289K LOC, ~40% tests); a two-stage LLM recognition layer classifies each input into a typed plan, files it autonomously, and auto-compiles a per-user knowledge base with hybrid search.
• Agentic decision layer: migrated hand-written routing rules to model-owned decisions inside a deterministic spine (queues, budgets, gates, receipts unchanged); autonomy bounded by reversibility: autonomous filing, propose-only structural change.
• Agentic workflows: multi-step plans with automatic re-grounding, branching, and end-of-thread synthesis.
• Evals: golden eval suites as CLI runners with per-run cost budgets; prompts as tested, versioned components.
• Multi-provider + structured output: Anthropic / OpenAI / Google / Perplexity abstraction with complexity-based cost routing, prompt caching, and strict tool-use + Zod schemas with retry-repair.
• RAG: pgvector-backed hybrid (semantic + keyword) search over an event-driven reindex pipeline.
• Production-grade: encryption-by-default, CI-enforced privacy invariants, per-user cost caps and entitlement gates, 6-hourly encrypted backups with a rehearsed restore, staging parity, GDPR export/erasure, a full specification / ADR process (31 ADRs).

• Designed and operate the multi-agent organisation that builds Alter: seven roles plus a scheduler over a git-backed ledger, delivering 178 merged PRs across 14 phases in nine months, solo.
• Concurrency-safe state model: single-writer files, one-file-per-fact sharding, and compiled (never hand-edited) indexes, eliminating a documentation-conflict livelock that was serialising parallel agents through full CI re-runs.
• Independent verification: every change fans out to a verifier and an adversarial breaker spawned on a deliberately thin brief (work-item + diff range only) so review cannot inherit the implementer's reasoning; black-box UAT structurally blind to the repo.
• Mechanical guardrails: capacity leases (file locks, disk-headroom refusal) with pre-execution hooks denying raw equivalents; limits enforced by the system, not by prompt instructions.
• Release engineering: typed release cuts with compiled user-facing changelogs; rebuilt the pipeline to cut hotfix-to-deployable-image from 60–90 min to 2–5 min via native per-architecture builds and digest-identical carry-forward of unchanged services.

• 0-to-1 AI filmmaking tool built around revision, not prompt-to-video: v1 a single agent over a typed dependency graph (cascade modes, consistency locks, a “converge” step); v2 a council of seven agent departments (producer, director, story, look & cast, camera, sound, unit + edit) through four production gates, all run from data by one five-role harness (assemble → generate → critique → resolve → record).
• Event-sourced core: an append-only production diary with actor, cause and cost on every event; every screen a rebuildable projection: causation chains (“why?”), time-travel replay, a per-department cost ledger, export/import of a whole project as one log.
• Coordination enforced in code, not prompts: single writer per artifact via a tool gateway; flags ruled on by a Director, never peer-to-peer; a convergence rule (no open revise/block flag, every department has evaluated the current version of every lock binding it); a bounded round cap that ends loops in a question to the human; paid generation kept outside the system as a human click.
• Working browser prototype through all four gates with replayable diaries; the mechanism write-up, the design document and the prototype are on this site.

• Directed and produced 2 AI short films end-to-end (story → moodboard → prompt engineering → consistent keyframes → ElevenLabs original scores); built an automated content pipeline.

Dissertation: long-term planning agent for sparse-reward environments using Monte Carlo Graph Search & Quality-Diversity search, directly relevant to modern agentic systems.